Patent Application
20160224465
Davis T., et al., “The University of Florida sparse matrix collection,” ACM Transactions on Mathematical Software (TOMS), 38, no. 1 (2011): 1.
Kvatinsky S., et al. “Models of Memristors for SPICE Simulations,” Proceedings of the IEEE Convention of Electrical and Electronics Engineers in Israel, pp. 1-5, November 2012.
Machine learning algorithms executed on High Performance Computers (HPC) address complex challenges such as mapping the human genome, investigating medical therapies, and executing high frequency derivative trading.
These problems are sequential in the sense that each parallelizable step depends on the outcome of the preceding step, and typically, large amount of data is exchanged (synchronized) within the chip between sequential and parallel processing cores in each step—as illustrated by timing diagram 10 of
Single Instruction Massively Parallel (SIMD) architectures are a class of parallel computers with multiple processing units performing the same operation on multiple data points simultaneously. Such machines exploit data level parallelism, and are thus well suited for machine learning over Big Data. High utilization of SIMD processor requires very high computation-to-bandwidth ratio and large data sets. Excess of coarse SIMD computing elements operating at high rates results in irregular thermal density and hotspots, further limiting SIMD scalability. Power dissipation and on-chip communication are the primary factors limiting the scalability of on-chip parallel architectures.
It has been shown that the maximal acceleration of a fine grain workload containing sequential and concurrent parts, occurs when the sequential section is assigned to a large, high ILP sequential processor (containing, for example, accelerators such as double precision floating point unit, function generator and branch prediction unit), denoted 23 in
An immediate limiting factor of such architectures is synchronization requiring data exchange (illustrated by double sided arrow SYNCHRONIZATION) between the sequential processor and the processing array, depicted as a red bus in
There may be provided a hybrid computer that may include a sequential processor, a single instruction massively parallel (SIMD) processor, and shared memory module that may be shared between the sequential processor and the SIMD processor.
The SIMD processor may be integrated with the shared memory module.
The SIMD processor may be directly coupled to the shared memory module.
The sequential processor may be coupled to the memory model only via a cache memory.
The SIMD processor may be arranged to perform a row access of the shared memory module while the sequential processor may be arranged to perform a column access of the shared memory module.
The SIMD processor and the sequential processor may be arranged to concurrently access the shared memory module.
The SIMD processor and the sequential processor may be arranged to access the shared memory module in an interleaved manner.
The shared memory module may include at least one SIMD memory region that may be writable only by the SIMD processor and at least one sequential processor memory region that may be writable only by the sequential processor.
The shared memory module may include at least one SIMD memory region that may be writable only by the SIMD processor.
The shared memory module may include at least one SIMD memory region that may be writable only by the SIMD processor; wherein the sequential processor may be configured to read content stored in the SIMD memory region.
The shared memory module may include at least one SIMD memory region that may be writable only by the SIMD processor; wherein the sequential processor may be prevented from reading content stored in the SIMD memory region.
The shared memory module may include shared memory cells that may be shared by the SIMD processor and by the sequential processor, and non-shared memory cells that may be accessed only by one of the SIMD processor and the sequential processor; wherein the shared memory cells may include more electrical components than the non-shared memory cells.
The shared memory module may include shared memory cells that may be shared by the SIMD processor and by the sequential processor, and non-shared memory cells that may be accessed only by one of the SIMD processor and the sequential processor; wherein at least some shared memory cells may be coupled to at least four control lines.
The shared memory module may include at least one sequential processor memory region that may be writable only by the sequential processor; wherein the SIMD processor may be configured to read content stored in the sequential processor memory region.
The sequential processor may be configured to control an access of the SIMD processor to the shared memory module.
The SIMD processor may include multiple processing units that may be configured to operate in parallel to each other.
The SIMD processor may include multiple one bit processing units.
The SIMD processor may include multiple multi-bit processing units.
The SIMD processor may include multiple processing units, a sequencer that issues commands to at least some of the multiple processing units, an interconnect and a network on chip (NOC) unit.
The SIMD processor may include multiple processing units, a sequencer that issues commands to at least some of the multiple processing units, a reduction tree and a network on chip (NOC) unit.
The SIMD processor may include multiple processing units and a reduction tree, wherein the reduction tree may include a group of adders that may be coupled to each other and to outputs of the multiple processing units.
The SIMD processor may include multiple processing units, each processing unit may include an adder, multiple registers, and a function generator.
The SIMD processor may include multiple one bit processing units, each one bit processing unit may include a one bit adder, multiple one bit registers, and a one bit function generator.
The SIMD processor may include a first number (N) of processing units, wherein the SIMD processor may be arranged to perform arithmetic operations on vectors that have a second number (N) of elements by performing multiple iterations of N length arithmetic operations; wherein N exceeds N.
The SIMD processor may include a first number (N) of processing units, wherein the SIMD processor may be arranged to perform a search for a key within a vector of a second number (N) of elements by performing multiple compare operations with different segments of the vector; wherein a length of the key does not exceed N and wherein N exceeds N.
The hybrid computer may include a group of SIMD processors, wherein the shared memory module may be virtually segmented to a group of shared memory segments; wherein each SIMD processor of the group of SIMD processors may be arranged to access a single shared memory segment of the group of shared memory segments.
The hybrid computer may include a group of SIMD processors, wherein the shared memory module may be virtually segmented to a group of shared memory segments; wherein the sequential processor may be arranged to access to each shared memory segment of the group of shared memory segments; wherein each SIMD processor of the group of SIMD processors may be arranged to access a single shared memory segment of the group of shared memory segments.
The shared memory module may include memory cells that may include memristive devices.
The shared memory module may include memory cells that may include memristive devices and diodes.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices and diodes.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices; wherein one end of each shared memory cell may be coupled to SIMD processor access circuits and wherein another end of each shared memory cell may be coupled to sequential processor access circuits.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices; wherein one end of each shared memory cell may be coupled to column sensing circuits and row drivers and wherein another end of each shared memory cell may be coupled to row sensing circuits and column drivers.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices; wherein one end of each shared memory cell may be coupled to column sensing circuits and row drivers and wherein another end of each shared memory cell may be coupled to row sensing circuits and column drivers; wherein the SIMD processor and the sequential processor may be configured to access the shared memory cells in a non-overlapping manner.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices; wherein one end of each shared memory cell may be coupled to column sensing circuits and row drivers that may be coupled to the sequential processor and wherein another end of each shared memory cell may be coupled to row sensing circuits and column drivers that may be coupled to the SIMD processor.
The SIMD processor may include multiple processing units that may include memristive devices.
The shared memory module and the SIMD processor may be integrated and may include memristive devices that may be configured to operate as memory cells and to perform arithmetic operations.
The shared memory module may include memory cells that may include multiple bit memristive devices.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include a memristive device; wherein the hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include a memristive device; wherein the hybrid computer may be configured to perform a subtraction operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases a resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include a memristive device;
The hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device;
The hybrid computer may be configured to perform a subtraction operation of the first and second operands, using the certain memristive device, by storing in the certain memristive device the value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases the resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include a memristive device.
The hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device;
The hybrid computer may be configured to perform a subtraction operation of the first and second operands, using the certain memristive device, by storing in the certain memristive device the value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases the resistivity of the certain memristive device;
The hybrid computer may be configured to perform long multiplications and long divisions by repeating subtraction and addition operations.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include a memristive device; wherein the hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include only a single memristive device; wherein the hybrid computer may be configured to perform a subtraction operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases a resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include only a single memristive device.
The hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device;
The hybrid computer may be configured to perform a subtraction operation of the first and second operands, using the certain memristive device, by storing in the certain memristive device the value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases the resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include only a single memristive device;
The hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device;
The hybrid computer may be configured to perform a subtraction operation of the first and second operands, using the certain memristive device, by storing in the certain memristive device the value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases the resistivity of the certain memristive device;
The hybrid computer may be configured to perform long multiplications and long divisions by repeating subtraction and addition operations.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include only a single memristive device; wherein the hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device.
The SIMD processor, the sequential processor and the shared memory may be implemented on a same die.
The SIMD processor, the sequential processor and the shared memory share at least one physical layer.
There may be provided a device that may include an array of memristive devices and a programming module; wherein the programming module may be arranged to initially store in a certain memristive device of the memristive device array a value that corresponds to a second operand and to feed the certain memristive device with one or more programming pulses that correspond to a first operand in a direction that decreases a resistivity of the certain memristive device thereby causing the certain memristive device to have a resistance that may be indicative of a sum of the first and second operands.
The device may include a read module that may be arranged to evaluate the resistance of the certain memristive device.
There may be provided a device that may include an array of memristive devices and a programming module; wherein the programming module may be arranged to initially store in each memristive device of the memristive device array a value that corresponds to a second operand and to feed the memristive device with one or more programming pulses that correspond to a first operand in a direction that decreases a resistivity of the memristive device thereby causing the memristive device to have a resistance that may be indicative of a sum of the first and second operands.
There may be provided a device that may include an array of memristive devices and a programming module; wherein the programming module may be arranged to initially store in a certain memristive device of the memristive device array a value that corresponds to a second operand and to feed the certain memristive device with one or more programming pulses that correspond to a first operand in a direction that increases a resistivity of the certain memristive device thereby causing the certain memristive device to have a resistance that may be indicative of a difference between the first and second operands.
There may be provided a method for subtracting a first operand from a second operand, the method may include: initially storing in a memristive device a value that corresponds to the second operand; feeding the memristive device with one or more programming pulses that correspond to the first operand in a direction that increases a resistivity of the certain memristive device thereby causing the memristive device to have a resistance that may be indicative of a difference between the first and second operands.
There may be provided a method for executing first and second instructions, the method may include executing the first instructions by a sequential processor; and executing the second instructions by a single instruction massively parallel (SIMD) processor; wherein the executing of the first and second instructions may include accessing a shared memory module that may be shared by the sequential processor and the SIMD processor.
There may be provided a device that may include an array of memristive devices, wherein each memristive device has a first end and a second end, wherein first ends of the memristive devices of the array may be coupled to a row driver a row sensing circuit; wherein second ends of the memristive devices of the array may be coupled to a column driver and to a column sensing circuit; wherein the row driver and the column sensing circuit may be arranged to operate during first time periods and wherein the column driver and the row sensing circuit may be arranged to operate during second time periods that do not overlap with the first time periods.
Each memristive device of the array may be coupled to an isolating element that may be arranged to reduce a coupling between a row coupled to the memristive device and a column coupled to the memristive device.
The isolating element may be a diode.
There may be provided a device that may include an array of memristive devices, wherein each memristive device has a first end and a second end, wherein first ends of the memristive devices of the array may be coupled to a row driver a row sensing circuit; wherein second ends of the memristive devices of the array may be coupled to a column driver and to a column sensing circuit; wherein the row driver and the column sensing circuit may be arranged to provide row accessibility and wherein the column driver and the row sensing circuit provide column accessibility.
Each memristive device of the array may be coupled to an isolating element that may be arranged to reduce a coupling between a row coupled to the memristive device and a column coupled to the memristive device.
The isolating element may be a diode.
There may be provided a hybrid computer that may include a sequential processor, a single instruction massively parallel (SIMD) processor, and shared memory module that may be shared between the sequential processor and the SIMD processor.
The SIMD processor may be integrated with the shared memory module.
The SIMD processor may be directly coupled to the shared memory module.
The sequential processor may be coupled to the memory model only via a cache memory.
The SIMD processor may be arranged to perform a row access of the shared memory module while the sequential processor may be arranged to perform a column access of the shared memory module.
The SIMD processor and the sequential processor may be arranged to concurrently access the shared memory module.
The SIMD processor and the sequential processor may be arranged to access the shared memory module in an interleaved manner.
The shared memory module may include at least one SIMD memory region that may be writable only by the SIMD processor and at least one sequential processor memory region that may be writable only by the sequential processor.
The shared memory module may include at least one SIMD memory region that may be writable only by the SIMD processor.
The shared memory module may include at least one SIMD memory region that may be writable only by the SIMD processor; wherein the sequential processor may be configured to read content stored in the SIMD memory region.
The shared memory module may include at least one SIMD memory region that may be writable only by the SIMD processor; wherein the sequential processor may be prevented from reading content stored in the SIMD memory region.
The shared memory module may include shared memory cells that may be shared by the SIMD processor and by the sequential processor, and non-shared memory cells that may be accessed only by one of the SIMD processor and the sequential processor; wherein the shared memory cells may include more electrical components than the non-shared memory cells.
The shared memory module may include shared memory cells that may be shared by the SIMD processor and by the sequential processor, and non-shared memory cells that may be accessed only by one of the SIMD processor and the sequential processor; wherein at least some shared memory cells may be coupled to at least four control lines.
The shared memory module may include at least one sequential processor memory region that may be writable only by the sequential processor; wherein the SIMD processor may be configured to read content stored in the sequential processor memory region.
The sequential processor may be configured to control an access of the SIMD processor to the shared memory module.
The SIMD processor may include multiple processing units that may be configured to operate in parallel to each other.
The SIMD processor may include multiple one bit processing units.
The SIMD processor may include multiple multi-bit processing units.
The SIMD processor may include multiple processing units, a sequencer that issues commands to at least some of the multiple processing units, an interconnect and a network on chip (NOC) unit.
The SIMD processor may include multiple processing units, a sequencer that issues commands to at least some of the multiple processing units, a reduction tree and a network on chip (NOC) unit.
The SIMD processor may include multiple processing units and a reduction tree, wherein the reduction tree may include a group of adders that may be coupled to each other and to outputs of the multiple processing units.
The SIMD processor may include multiple processing units, each processing unit may include an adder, multiple registers, and a function generator.
The SIMD processor may include multiple one bit processing units, each one bit processing unit may include a one bit adder, multiple one bit registers, and a one bit function generator.
The SIMD processor may include a first number (N) of processing units, wherein the SIMD processor may be arranged to perform arithmetic operations on vectors that have a second number (N) of elements by performing multiple iterations of N length arithmetic operations; wherein N exceeds N.
The SIMD processor may include a first number (N) of processing units, wherein the SIMD processor may be arranged to perform a search for a key within a vector of a second number (N) of elements by performing multiple compare operations with different segments of the vector; wherein a length of the key does not exceed N and wherein N exceeds N.
The hybrid computer may include a group of SIMD processors, wherein the shared memory module may be virtually segmented to a group of shared memory segments; wherein each SIMD processor of the group of SIMD processors may be arranged to access a single shared memory segment of the group of shared memory segments.
The hybrid computer may include a group of SIMD processors, wherein the shared memory module may be virtually segmented to a group of shared memory segments; wherein the sequential processor may be arranged to access to each shared memory segment of the group of shared memory segments; wherein each SIMD processor of the group of SIMD processors may be arranged to access a single shared memory segment of the group of shared memory segments.
The shared memory module may include memory cells that may include memristive devices.
The shared memory module may include memory cells that may include memristive devices and diodes.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices and diodes.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices; wherein one end of each shared memory cell may be coupled to SIMD processor access circuits and wherein another end of each shared memory cell may be coupled to sequential processor access circuits.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices; wherein one end of each shared memory cell may be coupled to column sensing circuits and row drivers and wherein another end of each shared memory cell may be coupled to row sensing circuits and column drivers.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices; wherein one end of each shared memory cell may be coupled to column sensing circuits and row drivers and wherein another end of each shared memory cell may be coupled to row sensing circuits and column drivers; wherein the SIMD processor and the sequential processor may be configured to access the shared memory cells in a non-overlapping manner.
The shared memory module may include shared memory cells that may be accessible to both the SIMD processor and to the sequential processor; wherein the shared memory cells may include memristive devices; wherein one end of each shared memory cell may be coupled to column sensing circuits and row drivers that may be coupled to the sequential processor and wherein another end of each shared memory cell may be coupled to row sensing circuits and column drivers that may be coupled to the SIMD processor.
The SIMD processor may include multiple processing units that may include memristive devices.
The shared memory module and the SIMD processor may be integrated and may include memristive devices that may be configured to operate as memory cells and to perform arithmetic operations.
The shared memory module may include memory cells that may include multiple bit memristive devices.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include a memristive device; wherein the hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include a memristive device; wherein the hybrid computer may be configured to perform a subtraction operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases a resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include a memristive device;
The hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device;
The hybrid computer may be configured to perform a subtraction operation of the first and second operands, using the certain memristive device, by storing in the certain memristive device the value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases the resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include a memristive device;
The hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device;
The hybrid computer may be configured to perform a subtraction operation of the first and second operands, using the certain memristive device, by storing in the certain memristive device the value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases the resistivity of the certain memristive device;
The hybrid computer may be configured to perform long multiplications and long divisions by repeating subtraction and addition operations.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include a memristive device; wherein the hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include only a single memristive device; wherein the hybrid computer may be configured to perform a subtraction operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases a resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include only a single memristive device;
The hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device;
The hybrid computer may be configured to perform a subtraction operation of the first and second operands, using the certain memristive device, by storing in the certain memristive device the value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases the resistivity of the certain memristive device.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include only a single memristive device;
The hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device.
The hybrid computer may be configured to perform a subtraction operation of the first and second operands, using the certain memristive device, by storing in the certain memristive device the value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that increases the resistivity of the certain memristive device;
The hybrid computer may be configured to perform long multiplications and long divisions by repeating subtraction and addition operations.
The SIMD processor and the shared memory module may be integrated and may include combined memory and computation cells; wherein each memory and computation cell may include only a single memristive device; wherein the hybrid computer may be configured to perform an addition operation of first and second operands, using a certain memristive device of a certain combined memory and computation cell, by storing in the certain memristive device a value that corresponds to the second operand and feeding the certain memristive devices with one or more programming pulses that correspond to the first operand in a direction that decreases a resistivity of the certain memristive device.
The SIMD processor, the sequential processor and the shared memory may be implemented on a same die.
The SIMD processor, the sequential processor and the shared memory share at least one physical layer.
There may be provided a device that may include an array of memristive devices and a programming module; wherein the programming module may be arranged to initially store in a certain memristive device of the memristive device array a value that corresponds to a second operand and to feed the certain memristive device with one or more programming pulses that correspond to a first operand in a direction that decreases a resistivity of the certain memristive device thereby causing the certain memristive device to have a resistance that may be indicative of a sum of the first and second operands.
There may be provided a device that may include a read module that may be arranged to evaluate the resistance of the certain memristive device.
There may be provided a device may include an array of memristive devices and a programming module; wherein the programming module may be arranged to initially store in each memristive device of the memristive device array a value that corresponds to a second operand and to feed the memristive device with one or more programming pulses that correspond to a first operand in a direction that decreases a resistivity of the memristive device thereby causing the memristive device to have a resistance that may be indicative of a sum of the first and second operands.
There may be provided a device that may include an array of memristive devices and a programming module; wherein the programming module may be arranged to initially store in a certain memristive device of the memristive device array a value that corresponds to a second operand and to feed the certain memristive device with one or more programming pulses that correspond to a first operand in a direction that increases a resistivity of the certain memristive device thereby causing the certain memristive device to have a resistance that may be indicative of a difference between the first and second operands.
There may be provided a method for subtracting a first operand from a second operand, the method may include: initially storing in a memristive device a value that corresponds to the second operand; feeding the memristive device with one or more programming pulses that correspond to the first operand in a direction that increases a resistivity of the certain memristive device thereby causing the memristive device to have a resistance that may be indicative of a difference between the first and second operands.
There may be provided a method for executing first and second instructions, the method may include executing the first instructions by a sequential processor; and executing the second instructions by a single instruction massively parallel (SIMD) processor; wherein the executing of the first and second instructions may include accessing a shared memory module that may be shared by the sequential processor and the SIMD processor.
There may be provided a device that may include an array of memristive devices, wherein each memristive device has a first end and a second end, wherein first ends of the memristive devices of the array may be coupled to a row driver a row sensing circuit; wherein second ends of the memristive devices of the array may be coupled to a column driver and to a column sensing circuit; wherein the row driver and the column sensing circuit may be arranged to operate during first time periods and wherein the column driver and the row sensing circuit may be arranged to operate during second time periods that do not overlap with the first time periods.
Each memristive device of the array may be coupled to an isolating element that may be arranged to reduce a coupling between a row coupled to the memristive device and a column coupled to the memristive device.
The isolating element may be a diode.
There may be provided a device that may include an array of memristive devices, wherein each memristive device has a first end and a second end, wherein first ends of the memristive devices of the array may be coupled to a row driver a row sensing circuit; wherein second ends of the memristive devices of the array may be coupled to a column driver and to a column sensing circuit; wherein the row driver and the column sensing circuit may be arranged to provide row accessibility and wherein the column driver and the row sensing circuit provide column accessibility.
Each memristive device of the array may be coupled to an isolating element that may be arranged to reduce a coupling between a row coupled to the memristive device and a column coupled to the memristive device.
The isolating element may be a diode.
The subject matter regarded as the invention is particularly pointed out and distinctly claimed in the concluding portion of the specification. The invention, however, both as to organization and method of operation, together with objects, features, and advantages thereof, may best be understood by reference to the following detailed description when read with the accompanying drawings in which:
In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be understood by those skilled in the art that the present invention may be practiced without these specific details. In other instances, well-known methods, procedures, and components have not been described in detail so as not to obscure the present invention.
The subject matter regarded as the invention is particularly pointed out and distinctly claimed in the concluding portion of the specification. The invention, however, both as to organization and method of operation, together with objects, features, and advantages thereof, may best be understood by reference to the following detailed description when read with the accompanying drawings.
It will be appreciated that for simplicity and clarity of illustration, elements shown in the figures have not necessarily been drawn to scale. For example, the dimensions of some of the elements may be exaggerated relative to other elements for clarity. Further, where considered appropriate, reference numerals may be repeated among the figures to indicate corresponding or analogous elements.
Because the illustrated embodiments of the present invention may for the most part, be implemented using electronic components and circuits known to those skilled in the art, details will not be explained in any greater extent than that considered necessary as illustrated above, for the understanding and appreciation of the underlying concepts of the present invention and in order not to obfuscate or distract from the teachings of the present invention.
The term “memristor” means any memristive device. It may include any two-port resistive device that is used to store data, including MIM (Metal-Insulator-Metal) resistive switches, MTJ (Magnetic Tunnel Junction), PCM (Phase Change Memory) and others.
Any reference to resistive memory (for example ReRAM) should be interpreted as a reference to any memristive device memory.
There is provided a hybrid general purpose SIMD computer architecture (GP-SIMD) that resolves the issue of synchronization by in-memory computing, through combining data storage and massively parallel processing.
GP-SIMD employs a two-dimensional access memory with modified SRAM storage cells and a bit-serial processing unit per each memory row. An analytic performance model of the GP-SIMD architecture is presented, comparing it to Associative Processor and to conventional SIMD architectures. Cycle-accurate simulation of four workloads supports the analytical comparison. Assuming a moderate die area, GP-SIMD architecture outperforms both the Associative Processor and conventional SIMD co-processor architectures by almost an order of magnitude while consuming less power.
The GP-SIMD delivers a number of advantages over Conventional SIMD (CSIMD) architecture:
Data processing and data storage are unified. There is no need for data transfer between sequential memory and the SIMD memory;
GP-SIMD allows concurrent operation of the sequential processor and SIMD co-processors on the shared memory, allowing the sequential processor to offload a task to the SIMD while continuing to process some other sequential functions.
The number of GP-SIMD fine grain processing units matches the number of memory rows, striving to match the entire dataset and obtain maximum parallelism.
The GP-SIMD architecture enables the sequential processor to associatively address the memory array. It may thus allow reduction of software complexity for certain sequential algorithms.
GP-SIMD power dissipation is distributed uniformly over the entire processing array rather than being concentrated around a smaller number of large, power hungry processing cores. Thus, there are fewer hotspots leading to further reduction of temperature dependent leakage power [005].
In this application there is provided a description of an integration of a fine grain massively parallel SIMD co-processor with a standard general purpose sequential processor and a shared memory with two-dimensional access, leading to improvement in performance as well as reduction in power dissipation. In this application there is provided a comparative performance and power analysis of GP-SIMD, a Conventional SIMD processor (CSIMD) and an Associative Processor (AP) [0044], supported by analytical modeling. GP-SIMD performance and power are validated with cycle-accurate simulations. It has been found that an inflection point where GP-SIMD outperforms both Conventional SIMD (CSIMD) and Associative Processor (AP) in both performance and power.
The GP-SIMD Processor
Top Level Architecture
The GP-SIMD is a hybrid general purpose and SIMD computer architecture 50 that resolves the issue of data synchronization by in-memory computing, through combining data storage and massively parallel processing.
As illustrated in
The sequential processor 33 accesses either one word at a time, or multiple words. Typically, such a transaction accesses one physical row at a time.
The SIMD processor 55 reads/writes a bit-slice (having r bits) comprising the same bit-number from all words in some partition of the memory. Physically, it may access multiple bits in a physical row and all rows per access, namely accessing multiple columns of the physical array (one after another).
The PUs are interconnected using an interconnection network (discussed in the context of
Note the key architectural differences between CPUs and GPU/APUs and GP-SIMD: While CPUs and GPU/APUs only share a memory, the GP-SIMD also enables in-memory processing (PIM). Further, while CPUs and GPU/APUs can only access their shared memory in a word-sequential manner, the SIMD of GP-SIMD accesses memory in a word-parallel manner, reaching all memory words at once. In addition, while current CPUs and GPU/APUs access their shared memory through a multi-level cache hierarchy, the SIMD in GP-SIMD accesses the shared memory directly and no transfers to cache are needed. As a result of these differences, [0030] details an advantage of power efficiency (GFLOPS/W) relative to such conventional architectures.
When the SIMD reads data from the shared memory, the contents of a bit slice of the memory are transferred into the register slice (RAs, RBs or RCs). Upon writing to the shared memory by the SIMD, the contents of one of the register slices are transferred into the GP-SIMD memory array. The PU is controlled by micro-instructions that are broadcast by the sequencer to all PUs. Some operations are conditioned by the single bit register RD, as determined by a two-bit field of the micro-instruction, RD_INST. The conditional operations are listed in TABLE 1. These conditional micro-instructions are handy for implementation of a variety of algorithms. While the first four operations are self-explanatory, the last two operations allow the sequential processor to perform associative commands on the memory array.
ST Value
y access (read/write) by the sequential processor or the
o-processor.
onal read command, according to a RD:
? RB = 0: RB = Memory output)
o-processor memory-write of RA
onal write command, according to RD:
? Memory input = RB: Memory input = RA).
onal (masked) read/write: Disable row for memory access
ential processor. Used during associative access
onal (masked) read/write: Enable row for memory access
ential processor. Used during associative access
indicates data missing or illegible when filed
Operating Modes
The GP-SIMD may be operated in three modes:
a. Sequential Processing mode, in which the general purpose sequential processor accesses data in memory during the execution of the sequential segments of a workload; data is typically transferred in lines to and from L1 cache (when the L1 cache line is longer than the row in shared memory, a single line from L1 spans multiple rows).
b. SIMD Processing mode, in which the parallelizable segments of a workload are executed. The SIMD addresses a column of the memory, and reads or writes to all rows of memory. Note that each PU corresponds to a single memory row. Thus in a single SIMD memory access, all PUs simultaneously read from or write into a single column of the memory array.
c. Concurrent Mode, in which both the sequential processor and the SIMD co-processor can access the shared memory.
In contrast with typical operation of Off-Memory SIMD accelerators, the sequential and SIMD processors may operate concurrently. In such a case, the sequential processor issues a SIMD instruction to the SIMD sequencer in a non-blocking manner, and proceeds to execute its own sequential code. The WAIT( ) synchronization instruction waits for the sequencer to complete a pending sequence of SIMD instructions before the sequential processor can proceed. There are no coherency issues (as there are no race conditions). The hardware does not provide special means, rather, the software is designed to enable non-interfering operation of both CPU and SIMD, as demonstrated in [0030]. The memory block is divided into a shared memory segment, which is shared by the sequential processor and the SIMD, and a SIMD-Only segment, which is accessible only by the SIMD. Data does not need to be transferred by SIMD to the CPU. Rather, shared data is created in the shared segment, either by the SIMD or the CPU. Thus, no data synchronization between sequential and parallel segments is required since both the sequential processor and SIMD co-processor access the very same memory array. Consequently, time and power are no longer incurred for data synchronization.
Arithmetic/Logic Operations
GP-SIMD can implement a wide range of arithmetic and logic processing tasks. Consider a workload using two datasets, A and B, each containing D elements, where each element is m bits wide. These vectors are mapped into the GP-SIMD memory array 74 such that two m bit adjacent column-groups hold vectors A and B (denoted 74 and 75 in
The addition is performed in m single-bit addition steps:
c[*]|s[*]
i
=a[*]
i
+b[*]
i
+c[*]∀i=0, . . . ,m−1 (1)
where i is the bit index and ‘*’ indexes a PU and a memory row. A bitwise vector addition of datasets having four elements (D=4) of four bits word-length (m=4) is shown in
Fixed point multiplication and division in GP-SIMD are also implemented bit-serially but word-parallel, consisting of a series of add-shift and subtract-shift vector operations. Shift is implemented by appropriate column addressing and therefore requires no extra cycles.
Thus, fixed point m×m bit vector multiplication requires 3m*m∈O(m2) cycles, regardless of the vector size, D. Floating point arithmetic for GP-SIMD is somewhat more complex to implement. IEEE single precision floating point vector multiplication takes close to 2500 cycles, regardless of the length of the dataset, D (assuming the entire dataset fits into GP-SIMD memory).
In This application, we consider single bit processing unit, and single bit SIMD memory access. However, the processing unit and memory access may employ larger word length. In such a case the following considerations may apply:
If the PU is made x-bit parallel with y-bit parallel memory access, the two dimensional memory array should increase, as typical memory rows are single bit high, allowing very few horizontal wires. Note further that y-bit SIMD access to memory complicate sequential processor memory access as a single word is mapped into y separate rows.
If the PU is made x-bit parallel with serial (1-bit) memory access, its sheer size may not allow placing it in the form-factor of memory rows. Further, the PU hardware may become heavily underutilized due to waiting for data.
Matrix Multiplication on GP-SIMD
It is assumed that two input matrices, the N×M multiplier matrix A and M×L multiplicand matrix B, are stored in the GP-SIMD memory in the Coordinate List (COO) format, where nonzero elements are stored along with their row and column indices. The output C=A×B is of dimension N×L.
In this example, the value field (Val) is allocated 32 bit-slices (memory columns 0:31) to accommodate a single precision floating-point number. The Row and Col fields are allocated 1 bit-slice each (memory columns 32:33).
Dense Matrix Multiplication
Dense matrix multiplication is explained by means of the example of
Broadcast: match the row-elements of B with the appropriate column-elements of A (namely, match A*,i with Bi,*).
Multiply: multiply pairs of matched elements
Reduce: add the singleton products together
The broadcast procedure is performed as a sequence of associative operations on the sequential processor, while Multiply and Reduce are performed as parallel operations on the SIMD processor, as follows. In GP-SIMD, the sequential processor can associatively match rows of memory array with a Key of k bits (in O(k) cycles, cf. Sect. [00266]). Once executed, all rows matching the Key are tagged (RD of the tagged rows contain ‘1’). The sequential processor can then execute associative read/write from/to the tagged rows (in O(1) cycles), achieving Broadcast.
Element A1,1 is read by the sequential processor (O(1) cycles),
The Row index fields of B are compared with the column index of A1,1 (‘1’) and matching memory rows are tagged (O(1) cycles),
A1,1 is written into field T1 of the tagged rows (O(1) cycles).
Thus, broadcasting a single element of A into matching pairs of B takes three cycles. In the general case, for N×M multiplier matrix, Broadcast entails O(M log M) cycles per row. Once all elements of the first row of matrix A are broadcast into the appropriate places (as depicted in
Since the multiplicand matrix B is transposed, the sum of the singletons residing in rows 5 and 6 yields C1,1, the first element of the output matrix, and the sum of singletons residing in rows 7, 8 yields C1,2. The Reduction process is effectively implemented by the hardware reduction tree. For single precision floating-point vector summation, the hardware reduction tree takes about 32 cycles (bit-serially feeding the 32 bits of the numbers into the tree, which subsequently reduces these number off-line in a pipeline fashion) and following the transfer of these singletons into the tree, processing of the next matrix row (Broadcast and Multiply) is initiated. The outputs of reduction tree are fed directly to the sequential processor that stores the computed elements of C into the designated memory addresses.
Although we detailed a single precision floating-point Broadcast, Multiply, and Reduce procedures, any operand word length may be considered. Further, although the example shows matrices A and B stored in separate PUs and memory rows, typically they would be stored side by side to enable handling larger matrices.
The hardware reduction tree of floating-point operands entails considerable area penalty. For effective implementation, note that Reduce follows a very long Multiply procedure of 2500 cycles. Pipelined Multiply-Reduce is thus preferable, in which a single floating-point accumulator can serially add up to 2500 singleton products, before piping its results to the next level of the tree. With that, the area and power of the reduction tree is kept at bay.
The Broadcast, Multiply, and Reduce procedures are now repeated for the next row of the multiplier matrix A, until all rows have been processed. The pseudo code of the algorithm is depicted below. It includes two nested loops. The external loop goes over the rows of matrix A. The internal loop performs Broadcast of all row elements of A.
~32 cycles
~1 cycle
O(log(M))
~1 cycle
The complexity of the algorithm is as follows:
M(1+log2 M+1)+CMult+CReduction] (2)
where CMuit=2500 cycles and CReduction=32 cycles.
Note that for large N, M, the complexity of GP-SIMD DMM approaches O(NM log2 M) and O(N2 log2 N) for square matrices. For example, the estimated complexity of multiplying a 10,000×10,000 matrix A exceeds 1B cycles.
The GP-SIMD DMM algorithm may also be used for multiplying dense matrix by dense vector, but it takes the same number of cycles regardless of the number of columns (L) of B (as illustrated in 24). Thus, the efficiency of the GP-SIMD DMM algorithm grows with the number of columns of the multiplicand matrix B (efficiency is defined as the number of actually performed arithmetic operations divided by the number of PUs times total cycle count, namely the maximum number of operations possible during the execution time).
d. Sparse Matrix Multiplication
The following description refers to an example of an implementation of multiplication of sparse matrix A by dense matrix B. The algorithm is similar to dense matrix multiplication but instead of processing all NxM elements of A, we only process the nonzero elements:
Broadcast is executed only for nonzero elements of A,
Multiply and Reduce are performed only for nonzero rows of A.
The pseudo code of the GP-SIMD SpMM algorithm is depicted below:
~32 cycles
~1 cycle
~32 cycles
~1 cycle
The complexity of GP-SIMD SpMM is approximately
N
NNZ
[M
NNZ(1+log2 M+1)+CMult+CReduction]
Where NNNZ is the number of nonzero rows of matrix A and MNNZ is the average number of nonzero elements per row (equal to number of nonzero elements of the multiplier matrix A, denoted by ANNZ, divided by NNNZ). Thus, for large NNNZ, MNNZ, the complexity of the algorithm approaches O(ANNz log2 M). For example, given sparse 10,000×10,000 matrix A with only 1,000 nonzero elements (10−5 density) and large MNNZ, the complexity of GP-SIMD SpMM is about 16,000 cycles, 105 times shorter than GP-SIMD DMM. Similar to the GP-SIMD DMM algorithm, GP-SIMD SpMM efficiency grows with the number of columns in the multiplicand matrix B.
Note that GP-SIMD SpMM is similar to GP-SIMD DMM, except for several cycles spent by the sequential processor handling matrix indices. It is thus preferable to employ the GP-SIMD SpMM algorithm even for dense matrices (with a minor modification for skipping zero elements). Hence, in the following section, only the results of GP-SIMD SpMM cycle accurate simulations are presented.
Associative Operations
GP-SIMD, besides being a massively parallel SIMD accelerator, can implement classical CAM operations such as associative search, sorting and ordering. The CAM allows comparing all data words to a key, tagging the matching words, and possibly reading some or all tagged words one by one. Consider a large vector, where each element is m bits wide, illustrated by column A in
Content-addressable access is achieved as follows. Assume that the memory array contains a vector of unique indices (A), adjacent to a vector of data (B). Comparing vector A with a key, followed by setting the RD_INST Bus to ‘10’ while issuing read to the memory array, allows the sequential processor to fetch a single value of vector B, corresponding to the row in which vector A matched the Key (that is, Output=B[i|A[i]=Key]). When multiple rows match the key, the values must be read one by one. The associative commands have proved valuable for efficient implementation of, e.g., sparse matrix multiplication. In addition, GP-SIMD can also implement a TCAM.
Interconnection Network
Arithmetic, logic and associative operations presented above require the relevant operands to be available at the memory row of the PU. However, common workloads require inter-PU data communications. Depending on the workload, communication requirements may vary from no communication (for “embarrassingly parallel” tasks such as Black-Scholes option pricing) to relatively intense communications (e.g., for FFT). In some cases, support for special pre-defined communication patterns or permutations can be useful (e.g., the permutations required for FFT). A dedicated interconnect is employed to allow all PUs to communicate in parallel.
Since GP-SIMD processing operation is mainly bitwise, the interconnection can be a relatively simple circuit-switched network. An example of an efficient network is a ±8 nearest neighbor, as in
Reduction Tree
A common reduction operation sums up a large array of values. Other common reduction operations are minimum or maximum. Reduction Tree (earlier introduced as a ‘response counter’ in) is an adder tree, enabling bit-serial parallel summation of PU values. Consider a vector A of D fixed point m-bit elements, as illustrated in
The reduction tree can also be implemented in software using the PUs and the interconnection network. Consider for example a vector A of eight fixed point 7-bit elements, as in
Similarly to the hardware implementation, the parallel binary tree is implemented as follows:
In the first (initialization) step 111, vector A is copied onto S.
In the second step 112, vector S is shifted one row up and stored in vector T, followed by adding vectors T and S. The oddly addressed elements of vector S contain the results of the first step of the binary tree addition.
In the third step 113, vector S is shifted two rows up and stored in vector T, followed by adding vectors T and S. Vector S now contains the results of the second level of the addition tree.
In the fourth step 114, vector S is shifted four rows up and stored in vector T, followed by adding vectors T and S. Vector S now contains the results of the third level of the tree, and the sum of all elements of A appears in the first element of S.
In general, given a ±log2N interconnection network, fixed point m-bit reduction entails O(2m+(2m+3m) log2(N))∈O(m log2(N)) cycles. In case of large datasets and limited interconnection networks, the last few shifting and summation steps can be executed serially by the sequential processor, so as to save processing time and power.
Circuit Implementation
The GP-SIMD PU utilizes its memory row as its register file for storing variables and temporaries. The width of the memory array is much higher than the word width of the sequential processor (e.g. 256 distinct columns vs. 32-bit word). A typical GP-SIMD memory array 124 is depicted in
A shared-memory cell consisting of a 7-transistor SRAM bit (transistors 136, 138, 139 and transistors in inverters 137 and 137′ of a latch) is used in the shared columns (cell 130 of
In the shared-memory cell, the standard 6T SRAM bit cell (including transistors 128, 138, latches 137 and 137′ as well as bit_line 131, bit_line’ 132 and word_line 134) is amended by a pass-gate (transistor 136) to enable column read/write access (column_line 135 and row_Data 133).
In the SIMD-only static bit cell (
e. By sequential processor: to read data from memory, the bit_line and bit_line-not lines are pre-charged, the word_line is asserted, and the bit lines are sensed. To write data to the memory, the bit_line, bit_line-not and the word_line are asserted.
f. By SIMD co-processor: to read data from memory, the column_line is asserted, and the row_data is pre-charged. To write data to the memory, the column_line and the row_data are asserted.
When the width of the memory is sufficiently small (e.g., 256 columns), a single pass-gate transistor is tied to the row data. In wider arrays, a differential pair, namely, row_data and row_data-not and an additional pass-gate transistor may be required. GP-SIMD's 6T and 5T SRAM cell outweigh the AP 12T cell of [0044] both area and power-wise. For the same die area, GP-SIMD array thus packs nearly twice the number of memory rows, and offers twice the AP performance for the same die area. These cells have been validated with Spice and demonstrate no degradation in performance vs. standard SRAM cell. There are several alternatives to placing the PUs alongside the memory rows: (1) Customized layout positioning the PU within the pitch of the memory rows, specifically since the size of the PU is relatively small; (2) k PUs placed side by side next to k memory rows, as in
The memory architecture enables the sequential processor and the SIMD co-processor to concurrently read from and write to the shared memory. Further, the architecture inherently allows segmentation of the memory array along both dimensions, so as to avoid conflict and allow processing on parts of the data, as follows:
g. Segmentation in the horizontal dimension: since the memory width is larger than sequential processor data-bus (for example, 64 bit data-bus and 256 bit wide array), the SIMD co-processor may be programmed to read from the entire memory array, but write only to the SIMD-Only section, while the sequential processor continues to process the shared memory space.
h. Segmentation along the vertical dimension: RD can be used to mask the word_line of certain rows.
i. Full segmentation: Access by the sequential processor and the SIMD co-processor may be interleaved, to avoid all interference.
The sequential processor may not have access to the entire SIMD memory. Thus, the software must assure minimal transfers between the Shared and SIMD-Only segments. Since the number of memory rows in a typical GP-SIMD implementation could be quite high (above one million), the GP-SIMD array can be partitioned into n smaller blocks each having D/n processors (and D/n complete memory rows). Each block could have, for example, its own sequencer. In such a case, all blocks share a piped version of the bus to the sequential processor.
A potential layout of such an implementation is depicted in
Note that since the sequential processor accesses no more than a single memory row at a time, such access involves only a single GP-SIMD block and the path to it; only that path is activated to conserve power. Reading data from a GP-SIMD block could entail a few cycles due the long path the signal may have to traverse. To mitigate such access delays, the sequential processor includes a (write through) data cache. If cache hierarchy exists, the shared memory array having a designated and distinct address space, shall connect to the sequential processor's L1 cache (see
GP-SIMD Programming Model and Performance
The SIMD co-processor is controlled by its sequencer. The sequencer converts each SIMD instruction into a sequence of micro-instructions for the SIMD PUs. The SIMD instruction set is listed in TABLE 2, and the special sequential processor instruction set is listed in TABLE 3. In contrast with typical operation of Off-Memory SIMD accelerators, the sequential and SIMD processors of GP-SIMD may operate concurrently. The sequential processor issues a SIMD instruction to the SIMD sequencer in a non-blocking manner, and proceeds to execute its own sequential code. Naturally, the sequencer does not process an instruction prior to completing the previous one. The WAIT( ) instruction waits for the sequencer to complete its pending sequence of SIMD instructions before the sequential processor can proceed. TABLE 2 also summarizes the arithmetic/logic performance of the SIMD co-processor. The SIMD instruction sequences are either programmed manually or generated by compilers. Typically, SIMD operations should be made available as library functions.
erformance (cycles)
RC, DST)
OR/XOR (SRC1,
ORI/XORI (SRC1,
ogic (Imm.)
SUB_FIXED (SRC1,
th., Fixed Point
SUB_FP (SRC1,
DIV_FIXED (SRC1,
th., Fixed Point
DIV_FP (SRC1,
OWN_FIXED (SRC,
C., Fixed Point
WN_FP (SRC, NUM,
XED (SRC, m, DST)
th., Fixed Point
P (SRC, DST)
W_FIXED (SRC, m,
ction, SW, Fixed
3m + 1 + 2m) log2 N
W_FP (SRC, DST)
uction, SW, FP
0 + 1 + 64) log2 N
W_FIXED (SRC, m,
ction, HW, Fixed
m + log2 N) ∈ O(m +
W_FP (SRC, DST)
uction, HW, FP
2 + log2 N) ∈ O(log N)
RD (COL)
d a col into RD
MASK (COL)
asked Read
_MASK (COL,
asked Write
COND (COL, DATA,
ssociative Read
_COND (COL,
ssociative Write
ciative Compare
RE (DATA, COL, m)
indicates data missing or illegible when filed
ormance (cycles)
D (ROW, COL, m)
ITE (ROW, COL, DATA, m)
)
indicates data missing or illegible when filed
Mnemonics:
Due to the novel bitcell of the GP-SIMD shared memory array (
Analytic Model and Comparative Analysis
In this section, we provide an analytical performance and power consumption model of the CSIMD, AP and GP-SIMD and compare their relative performance, area and power consumption under constrained area and power resources. For the purpose of modeling, it is assumed that the sequential processor is identical in all architectures, and focus on concurrent workloads. We further assume that the CSIMD executes a single fused multiply-add operation in 1.5 GHz, while the GP-SIMD and AP employ faster clocks of 2.5 GHz, as they execute bitwise rather than word-parallel instructions and thus have shorter critical paths.
Consider a workload having WL single cycle instructions (i.e., arithmetic, control, register file access and alike). The workload consists of N fine grain concurrent execution segments, requiring TRef=WL*1 cycles to execute on a baseline reference floating point engine capable of performing 1 FLOP/Cycle (single precision floating point operation per cycle). Such floating point engine consumes an area of 0.01 mm2 in 45 nm and dissipates 10 mW. With 22 nm process technology, the same floating point engine would consume an area of 0.003 mm2 and dissipate roughly 5 mW. These figures pertain to a baseline floating point engine, without the accompanying connectivity logic, register file and the like.
A portion of the workload instructions, CSync, is dedicated to data synchronization before and after each segment starts and ends respectively, while a further portion of the workload, CInter, is dedicated to data exchange among the execution units. The remaining portion of the workload, CProc, is actual processing.
have: CProc+CSync+CInter=1
In a large design comprising thousands of PUs, it is impractical to consider a complex network topology that would consume super-linear area. We thus assume a linear complexity ring topology interconnection network in all three architectures, where in CSIMD the network is 32 bit wide, and in GP-SIMD and AP, the network is 1 bit wide. In a ring topology, within O(N) serial steps, N PUs can exchange data according to any permutation. To simplify the analysis, CSIMD, AP and GP-SIMD analytical equations were consolidated into TABLE 7 of
Conventional SIMD Processor
A Conventional SIMD (CSIMD) co-processor is depicted on the left side of
The area of the CSIMD processor can be presented as in section (12) of table 4, where AALU is the ALU area and ARF is the register file area. The inter-PU connection network is omitted for simplicity. For efficient comparison between PU and memory areas, we represent all area values (ALU, registers, memory) in terms of baseline SRAM cell area. Let the baseline SRAM cell area be 1. In 22 nm CMOS technology, the actual figure is in the range of 0.1 μm2.
Then we can write the expression of section (15) of table 4, where ACSIMDALUo is the area of a single bit of a high speed parallel ALU and ACSIMDRFo is the area of a register bit (a flip-flop), both measured in baseline SRAM cell area units; m is data wordlength and k is the size of the register file. This model assumes that a floating point ALU is similar in area to a large integer multiplier. The model is quite simplistic and does not take into account numerous aspects of CSIMD design (instruction cache, communication and control, etc.). Its purpose is providing the best case reference figures for the comparative analysis of the CSIMD processor's speedup, area and power.
The average power of the CSIMD processor can be written as in section (17) of table 4, where PProc-CSIMD is the average power consumed during a single processing operation; PInter is the average power consumed during inter-PU communication; PSync-CSIMD is the average power consumed during synchronization, and PLeak-CSIMD is the leakage power. Just as in the case of area comparison, we represent all power values (ALU, registers, memory) through the write power consumption of a baseline SRAM memory cell. Let the power consumption of the baseline SRAM cell during write from ‘0’ to ‘1’ or from ‘1’ to ‘0’ be 1. In 22 nm CMOS technology, the actual figure is in the range of 1 μW at 4 GHz. Then we can further write the CSIMD power consumption as in section (20) of table 4, where PCCSIMDALUo and PCSIMDRFo are the average per-bit power consumptions of the ALU and RF respectively during computation. PINTERo is the per-bit power consumption during the inter-PU communication. PSYNCo is the per-bit power consumed during synchronization. PProc-CSIMD, PInter-CSIMD and PSync-CSIMD are measured in SRAM cell write power consumption units. Leakage power can be expressed as in section (26) of table 4, where A is the area, V is the supply voltage, α and β are constants, and γ is the leakage area coefficient that depends on silicon process and operating conditions.
Associative Processor
In this section we detail the analytical model for the speedup, area and power consumption of the Associative Processor (AP). AP is based on Content Addressable Memory (CAM), which allows comparing the stored data words to a key, tagging the matching words, and writing new data in tagged rows. AP is a non-von Neumann in-memory computing accelerator. The results of basic computing operations are pre-calculated and compiled into a sequence of AP instructions. The operands are stored in the CAM array. In response to an input combination, the result is written into the CAM, typically to multiple locations.
The architecture of AP 170 is presented in
SA is a small sense amplifier. A detailed AP design is presented in [0044]. In a conventional CAM, a compare operation is typically followed by a read of the matched data word. In AP, a compare is usually followed by a parallel write into the unmasked bits of all tagged rows. In a similar manner to GP-SIMD, a fixed-point m bit addition takes 16m∈O(m) cycles. Subtraction is performed similarly and also requires O(m) cycles.
Fixed point multiplication and division in AP require O(m2) cycles.
In a similar manner to GP-SIMD, the AP [0044] does not entail data synchronization unless the entire data set does not fit in the AP's memory array, thus, see section (4) of table 4. The execution time of the concurrent portion of the workload can be written as in section (7) of table 4, where nAP is the number of PUs in the AP, SAPe is the speedup of associative PU relative to the CSIMD's PU. Lacking a-priori knowledge of the workloads to be executed on the AP, it is assumed that the worst case scenario comprising a continuous series of single precision floating point multiplications, which in one direct implementation takes 8800 cycles vs. a single cycle on the baseline CSIMD's PU. In this case SAPE=1/8800. The speedup of the AP can then be written as in (10).
The area of the AP can be written as in section (13) of table 4, where k is the width of the associative processor memory array (in m-bit data words) reserved for temporary storage, AAPTAGo is the AP TAG cell area, AAPo is the AP memory cell area, both measured in SRAM cell area units, and 2AALUo is the per-PU hardware reduction tree size, assuming the adders of each tree level are of growing word length [0044]. Similarly to the CSIMD coprocessor, we ignore the area of the interconnection network for simplicity.
The average power of the AP can be written as in (18), where PProc-AP is the average power consumed during a single processing operation; PInter-AP is the average power consumed during inter-PU communication; PSync-AP is the average power consumed during synchronization when the entire data set does not fit in the AP, and PLeak-AP is the leakage power. In a similar manner to the previous section, we can further write the AP power consumption as in (21), where PAPo is the average per-bit power consumptions of the AP during computation. PINTER-o is the per-bit power consumption during the inter-PU communication. PSYNCo is the per-bit power consumed during synchronization. PProc-AP, PInter-AP and PSync-AP are measured in SRAM cell write power consumption units. Leakage power can be expressed as in section (27) of table 4. Note that for comparison purposes we use the same leakage power formulation (represented as a function of area only as in section (26) of table 4) for both the AP and the CSIMD processor. This may overestimate AP power somewhat: First, the leakage power per area could be lower for memory than for logic. Second, APs have fewer hotspots. Since the leakage power is highly temperature dependent, hotspots may lead to higher leakage in the CSIMD processor.
GP-SIMD Processor
In this section we detail the analytical model for the speedup, area and power consumption of the GP-SIMD. As discussed above, the GP-SIMD does not entail data synchronization unless the entire data set does not fit in the GP-SIMD's memory array. Thus section (5) of table 4. The execution time of the concurrent portion of the workload can be written as in section (8) of table 4, where nAP is the number of PUs in the GP-SIMD, SIMD, SGP-SIMDe is the speedup of associative PU relative to the CSIMD's PU. Similar to the AP analysis, it is assumed that the worst case scenario comprising a continuous series of single precision floating point multiplications, which in one direct implementation takes 2500 cycles vs. 1 cycle on the baseline CSIMD's PU. In this case SGP-SIMDe=1/2500. The speedup of the GP-SIMD can then be written as in section (11) of table 4.
It is assumed that the GP-SIMD's sequential processor accesses only the first L bits of the memory array. Thus, the average size of the memory bit cell is in section (14) of table 4, where k is the width of the GP-SIMD memory array (in m-bit data words) reserved for temporary storage, ASHAREDo is the shared memory cell area and ASIMDONLYo is the SIMD-only memory cell area, both measured in SRAM cell area units. The total area of the GP-SIMD can be written as in section (16) of table 4, where AGPSIMDPUo is the GP-SIMD processing unit cell area measured in SRAM cell area units, and 2AALUo is the per-PU hardware reduction tree size. Similarly to the CSIMD coprocessor and the AP, we ignore the area of the interconnection network for simplicity.
The average power of the GP-SIMD can be written as in section (19) of table 4, where PProc-GPSIMD is the average power consumed during a single processing operation; PInter-GPSIMD is the average power consumed during inter-PU communication; PSync-GPSIMD is the average power consumed during synchronization when the entire data set does not fit in the AP, and PLeak-GPSIMD is the leakage power.
In a similar manner to the previous section, we can further write the GP-SIMD power consumption as in section (22) of table 4, where PGPSIMDo is the average per-bit power consumptions of the AP during computation. PINTERo is the per-bit power consumption during the inter-PU communication. PSYNCo is the per-bit power consumed during synchronization. PProc-GPSIMD, Pinter-GPSIMD and PSync-GPSIMD are measured in SRAM cell write power consumption units. Leakage power can be expressed as in section (28) of table 4.
Note that for comparison purposes we use the same leakage power formulation (represented as a function of area only as in section (26) of table 4) for the CSIMD, AP and GP-SIMD architectures. This might be somewhat unfair to the GP-SIMD: First, the leakage power per area could be lower for memory than for logic. Second, GP-SIMD, in a similar manner to AP, may have fewer hotspotsError!Reference source not found. Since the leakage power is highly temperature dependent, hotspots may lead to higher leakage in the CSIMD processor. Note that the AP's arithmetic/logic performance is 2.5-5× lower than GP-SIMD's arithmetic performance, regardless of the number of processing elements, N.
Consider a given die total area, ATotal. The number of the CSIMD, AP and the GP-SIMD processing units can be found as in sections (26)-(28) of table 4. Substituting nSIMD, nAP and nGPSIMD into the equations listed in the previous sub-sections, we determine the speedup and the power equations of each architecture. The area and power parameters we use for modeling purposes are listed in TABLE 4.
Speedup vs. area for the CSIMD coprocessor, the AP and the GP-SIMD are shown in curves 181(1)-181(3) of graph 181 of SIMD bandwidth remains fixed. After sufficient area becomes available the speedup saturates.
ttributed
bit cell area
bit cell power
cell area
cell power
bit (FF) area
bit (FF) power
dup relative to sequential
ll area
ll power
cell area
D speedup relative to
D Shard bit cell area
D SIMD-Only cell area
D PU area
D PU power
D Intercommunication power
D Synchronization power
rd length
file size (in 32-bit words)
al processor word length
wer coefficient
mW/mm2
d portion dedicated for
d portion dedicated for
ication
th (number of 32b words
indicates data missing or illegible when filed
Area and power parameters are relative to the area and power of SRAM bit cell respectively; the values are based on typical standard cell libraries.
For smaller total area, the speedups of the AP and GP-SIMD are lower than the speedup of the CSIMD coprocessor. The breakeven point lies around 14 mm2 for the GP-SIMD. Diminishing returns affect the GP-SIMD and AP speedups to a lesser extent than in the CSIMD, since they only occur when the data set does not fit into the internal memory array. Following [0044], to emphasize this effect, and for illustrative purpose only, it is assumed that that the data set size grows with the GP-SIMD and AP size respectively, and up until ATotal=25 mm2 (at which point, nGPSIMD=1.0×106 and nAP=5.3×105) the GP-SIMD and AP internal memory array have just enough space to hold all necessary program variables. However, from that point on, both processors will have to exchange additional data with, for example, an external DRAM. It is assumed that for each additional processing unit beyond ATotal=25 mm2, a single 32b data-word per 300 additional processing units will incur synchronization delay. These assumptions cause the GP-SIMD and AP speedup to saturate as well. Note that without this assumption, both GP-SIMD and AP Speedup would have continued to climb linearly with area. Thus, Off-chip bandwidth affects GP-SIMD in a similar manner to conventional GPUs. Note however, that at ATotal=25 mm2 (just prior to saturation), the GP-SIMD provides speedup of 5.5× over the CSIMD, and 7× over the AP. The speedup factor over CSIMD continues to grow till the GP-SIMD saturates at ATotal=40 mm2.
Energy vs. area is shown in curves 182(1)-182(3) of graph 182 of
The speedup/energy ratio vs. area is shown in curves 183(1)-183(3) of graph 183 of
In general, GP-SIMD achieves higher performance than CSIMD due to CSIMD stalling on synchronization. GP-SIMD achieves higher performance-per-area than AP due to its lower memory cell area (˜2×), allowing the integration of twice the amount of PUs and memory rows per a given area, and its faster floating point implementation (˜3.5×) due to GP-SIMD arithmetic PU vs. associative-only PU of AP.
GP-SIMD Cycle-Accurate Simulation
The GP-SIMD, CSIMD and AP high level analytical models cannot be used for realistic simulations of known workloads, but are useful for early architectural design space exploration and trend analysis. In this section, we detail a home-grown cycle-accurate simulator of GP-SIMD, useful for the design and performance evaluation of real workload. Following [0046] and [0044], four workloads have been selected for performance and energy consumption simulations:
N-point Fast Fourier Transform (FFT)
N-option pairs Black-Scholes option pricing (BSC)
N-point Vector Reduction (VR)
Dense Matrix Multiplication of two √{square root over (N)}×√{square root over (N)} matrices (DMM)
Where N is the data set size, scaled for simplicity to the processor size (following the methodology suggested in and [0044]), i.e. N=nGPSIMD. Note that simulations do not cover the cases where the data size exceeds the size of the processor (requiring data synchronization).
The workloads are hand-coded for the simulations. For FFT, we use optimized parallel implementation outlined in. For N point FFT, the sequential computing time is O(N log N). For N PU parallel implementation, computing time is reduced to O(log N). Following each computing step, N intermediate results need to be exchanged among the PUs. We thus assign each multiply-accumulate operation to a single PU. For Black-Scholes (BSC), we used a direct implementation. With sequential computing time of O(N), Black and Scholes option pricing requires no interaction between separate option calculations (Black and Scholes option pricing is an example of an ‘embarrassingly parallel’ task). We thus assign a single PU to handle a single call option of a single security at a single strike price and a single expiration time. Vector reduction (VR) is implemented using the software reduction tree where a single PU retains a single vector element. For √{square root over (N)}×√{square root over (N)} dense matrix multiplication (DMM), the sequential execution time is O(N3/2). The GP-SIMD DMM algorithm utilizes N PUs and yields parallel execution time of O(N), with √{square root over (N)} data elements being broadcast every step [0030]. The DMM algorithm employs the hardware reduction tree to accelerate multiple vector summations in parallel: at half the depth of the reduction tree, all √{square root over (N)} vector sums are available and are transferred directly to the sequential processor.
The simulator is cycle based, keeping records of the state of each register of each PU, and of the memory row assigned to it. Each command (for example, floating point multiply) is broken down to a series of fine-grain single bit PU operations. In a similar manner to SimpleScalar [006], the simulator also keeps track of the registers, buses and memory cells that switch during execution. With the switching activity and area power models of each baseline operation detailed in TABLE 4, the simulator tracks the total energy consumed during workload execution. We follow the area assumptions detailed in TABLE 4, and simulate speedup and power for growing datasets corresponding to growing number of PUs and hence growing total area.
Cycle-based simulated speedup is presented in curves 191(1)-191(4) of graph 191 of
Speedup is defined in section (11) of table 4. Energy consumption is shown in curves 192(1)-192(4) of graph 192 of
Curves 193(1)-193(4) of graph 193 of
Sparse Matrix multiplication (SpMM) on GP-SIMD [0030] on a large set of matrices, demonstrated potentially improved power efficiency of 20× relative to a number of GPU designs. Specifically, GP-SIMD SpMM power efficiency ranges between 0.1-100 GFLOPS/W. SpMM power efficiency of GPUs such as NVidia's K20 and GTX660 is in the 0.1-0.5 GFLOPS/W range (as reported by references cited in [0030]). A wide variety of multi-core processors such as quad-core AMD Opteron 2214, quad-core Intel Xeon E5345, eight-core Sun UltraSparc T2+T5140 and eight-SPE IBM QS20 Cell reportedly reach SpMM power efficiency of up to 0.03 GFLOPS/W.
Resistive GP-SIMD
The resistive GP-SIMD employs a resistive shared memory array instead of the CMOS memory of the original GP-SIMD design [0031]. In this section, this application may illustrate the ReRAM memory crossbar array, symmetric two-dimensional access to memory in the resistive GP-SIMD, and the potential for multi-valued computations within the resistive memory array.
Resistive Memory Crossbar
One of the limitations of commercially available flash non-volatile memory is its low endurance. Flash cells may withstand around 100,000 write cycles before the wear begins to deteriorate the integrity of the storage [0040]. Further, endurance drops with scaling to smaller geometries, and modern multilevel 16 nm NAND flash are limited to ˜2000 write cycles. The size and power consumption of circuits such as charge pump further limits flash scalability. As VLSI geometries get smaller, the yield and reliability of flash cells decrease dramatically [0034]. These limitations lead to the development of alternative nonvolatile memory technologies, such as Ferroelectric RAM (FeRAM), Magnetoresistive random access memory (MRAM), Spin-transfer torque random access memory (STT-RAM) potentially replacing MRAM, Phase-change memory (PRAM), Conductive-bridging RAM (CBRAM), and ReRAM, resistive random access memory. ReRAM consists of a normally insulating dielectric material that can be made to conduct through a filament or conduction path formed after the application of a sufficiently high voltage.
Once the filament is formed, it may be RESET (broken, resulting in high resistance state) or SET (re-formed, resulting in lower resistance state) by an appropriately applied voltage. Further promising features include low operating voltage, small cell area, and fast write time. To further increase ReRAM density, a crossbar structure is employed. In crossbar arrays, ReRAM cells are sandwiched between wordlines and bitlines, having area as small as 4F2 (where F is the feature size). Moreover, since a crossbar array permits a multilayered structure, its effective cell area may be further reduced [001][0013][0020].
The resistance of ReRAM is changed by applying electrical current or voltage. The resistance is bounded by a minimum resistance RLRS (low resistive state, logic ‘1’) and a maximum resistance RHRS (high resistive state, logic ‘0’). Write reliability is a serious concern in ReRAM crossbar arrays: Voltage applied across a crossbar varies based on the location of the cell as well as the data pattern stored in other ReRAM cells in the array, due to wire resistance and sneak currents. To that end, several schemes for write operation have been proposed. When multiple cells are selected for write in a single row, it is impossible to complete the write operations (both SET and RESET) in one step. Thus, SET and RESET operations are separated into two steps: first SET is performed for all designated columns, followed by RESET of the other columns [0013]. Read operation is also affected by sneak current, especially when the selected cell of a row is in high resistance state and the others are in low resistance state [007]. In order to alleviate read disturbance, parallel read can be employed to read all the cells in the same row. A per-cell diode [001][0036] may be used is to terminate read and write sneak paths.
A conventional resistive crossbar (
Symmetric 2-D Access to Resistive Shared Memory Array
We introduce a shared resistive memory crossbar array that facilitates row and column addressability. The conventional array (FIG. 20Error!Reference source not found.A) is designed for access to columns of one row at a time.
There is provided a two-dimensional symmetric access array (
There is provided a two-dimensional symmetric access array includes an array of memristors, wherein one end of the memristors is connected to word lines that are coupled to row driver 201 and row sensing circuit 206. Another end of the memristors are connected to bit lines that are coupled to column driver 207 and column sensing circuit 202.
It is designed to enable area efficient access to either the columns of one row (by the sequential processor) or to all rows of a single column (by the SIMD processor). The top level architecture of the resistive GP-SIMD (ReGP-SIMD) closely follows the CMOS based GP-SIMD architecture of
Reading and writing the memory crossbar array by the sequential processor and the SIMD are interleaved. Depending on the current sequential processor instruction and SIMD microinstruction (namely, if there is access to the array), up to four operations out of the following six possibilities may take place during a single access, requiring either one read or two writes for each processor:
Sequential processor access (
1. Read: SIMD's row and column circuits are disconnected from the array. Sequential processor's column sensing circuits and row drivers are connected to the array. The bit_lines are precharged and the appropriate row driver corresponding to the sequential processors' addressed row grounds its word_line. Current flows from the bit_lines through all SET ReRAM devices connected to the grounded word_line. The columns sensing circuits sense changes on the bit_line, and output the corresponding bits to the sequential processor.
2. Write ones, SET: SIMD's row and column circuits are disconnected from the array. Sequential processor's row drivers are connected to the array. The bit_lines designated to be SET are grounded, and the appropriate row driver corresponding to the sequential processors' addressed row asserts its word_line. Current flows from the word_line through all designated ReRAM devices connected to the asserted bit_line, and sets them to low resistance.
3. Write zeros, RESET: SIMD's row and column circuits are disconnected from the array. Sequential processor's row drivers are connected to the array. The bit_lines designated to be RESET are asserted, and the appropriate row driver corresponding to the sequential processors' addressed row grounds its word_line. Current flows from the bit_lines through all designated ReRAM devices connected to the grounded word_line, and resets them to high resistance.
SIMD co-processor access (
1. Read: Sequential processor's row and column circuits are disconnected from the array. SIMD's Row sensing circuits and column drivers are connected to the array. The word_lines are precharged and the appropriate column driver corresponding to the SIMD co-processors' addressed column grounds its bit_line. Current flows from the word_lines through all SET ReRAM devices connected to that column. The rows sensing circuits sense changes on the word_lines, and outputs the corresponding bits to the SIMD PUs.
2. Write ones, SET: Sequential processor's row and column circuits are disconnected from the array. The SIMD's column drivers are connected to the array. The word-lines designated to be SET are grounded, and the appropriate column driver corresponding to the SIMD co-processors' addressed column asserts its bit_line. Current flows from the bit_line through all designated ReRAM devices connected to the asserted bit_line, and sets them to low resistance.
3. Write zeros, RESET: Sequential processor's row and column circuits are disconnected from the array. The SIMD's column drivers are connected to the array. The word_lines designated to be RESET are asserted, and the appropriate column driver corresponding to the SIMD co-processors' addressed column grounds its bit_line. Current flows from the word_lines through all designated ReRAM devices connected to the grounded bit_line, and resets them to high resistance.
Reading and writing by the SIMD or by the sequential processor are not concurrent. Thus, the SIMD sequencer may issue: (a) no access operation if the SIMD neither reads nor writes to the array; (b) a single access operation if the SIMD reads from the array; (c) two access operations if the SIMD writes to the array, and similarly for the sequential processor. Hence, zero to four shared memory operations are executed per a single clock cycle of the sequential processor (combining one SIMD memory operation and one memory operation of the sequential processor). The shared memory array is thus clocked with a frequency of four times the clock of the sequential and SIMD processor. That is, four distinct shared memory operations may be executed in a single sequential- or SIMD-processor cycle, and interleaving happens within that single clock cycle. Note that the interleaved addressing enables the sequential processor and the SIMD co-processor to read from and write to the shared memory without conflicts, as accesses are separated into distinct operation.
A more advanced implementation may allow the sequential processor to access the shared memory array on all four timeslots if the SIMD processor is not accessing it. However, our simulation and comparative analysis does not assume such implementation.
The GP-SIMD sequencer, depending on the combination of sequential and SIMD instructions being executed, selects from one to four operations. Note that the interleaved addressing enables the sequential processor and the SIMD co-processor to concurrently read from and write to the shared memory without conflicts, as accesses are separated into distinct operation.
In ReGP-SIMD, sneak currents affect both the read and write operations as in a standard ReRAM crossbar. The purpose of per-cell diode is to terminate such sneak paths. Note that the diode is placed above ReRAM and hence takes no additional area. Utilizing a ReRAM crossbar reduces ReGP-SIMD storage cell to 4F2/k footprint (where k is the number of vertically integrated ReRAM layers [001]). Such level of integration allows placing 200M (k=1) 256-bit PU ReGP-SIMD on a single 200 mm2 silicon die. To compare, the CMOS GP-SIMD cell area is approximately 190F2 [008], limiting the GP-SIMD to 8.2M PU in the same silicon area. Only the shared memory array scales with ReRAM. The PU however is implemented in CMOS, and thus limits the scalability of ReGP-SIMD.
The switching time of ReRAM may reach the range of a hundred picoseconds [0041], allowing memory access rates at frequencies of more than 1 GHz. The energy consumption of the resistive cell during read is less than 1fJ per bit. Unfortunately, the ReRAM write energy is in the range of a few tens of fJ per bit [008][0036], and in addition to the energy required by the PU CMOS logic [0031], prohibit scaling the ReGP-SIMD to 200M rows. However, write energy depends on the ReRAM material, and fortunately research of efficient ReRAM materials is likely to result in lower write energy consumption in the future. Further, resistive implementation of logic [0023] may result in a power- and area-efficient ReRAM PU. Another factor potentially limiting the use of ReGP-SIMD is the resistive memory endurance, which is in the range of 1012 [0032]. Given that during arithmetic operations with write operation typically occurring once every third cycle (taking vector add as an example, and considering only SIMD accesses to the array), the probability of a single bit to be written is (1/256)·(1/3)≈1·10−3 per cycle. At 1 GHz, it limits the endurance-driven MTBF of a ReGP-SIMD to ˜1·106 sec, less than two weeks. However, recent studies predict that the endurance of resistive memories is likely to grow to the 1014-1015 range, which may extend the endurance-driven MTBF of a ReGP-SIMD to a number of years.
A notable advantage of ReRAM is that it may be placed over logic and over other circuits [0026]. Moreover, the area over a single PU may accommodated about 500 resistive memory cells; if the per-PU shared memory word is 500 or fewer bits, there is no need for any area dedicated for memory, and the majority of the ReGP-SIMD area consists of PUs with shared memory overlaid on top. In addition to saving area, access time by the PU is significantly reduced because no long wires are needed.
This area reduction benefit is much more pronounced in processing-in-memory architectures than in other many-cores. For instance, if CMOS memory occupies 50% of the chip area in a many-core, replacing memory by resistive cells laid on top the logic might cut chip area by two; CMOS GP-SIMD memory occupies more than 95% of chip area and converting to ReGP-SIMD may result in area reduction by a factor of 20.
Single RERAM, Multi-Valued Adder/Subtracter
ReRAM cells may be able to store multi-valued data [0047]. The stored multi-valued data corresponds to the voltage and time interval the resistive cell was subjected to. Assume we can subject the resistive cell to a sequence of pulses, each timed to set the multi-valued cell to a number corresponding to powers of 2 (that is, 1, 2, 4, 8, 16 . . . ). If the dielectric material of the resistive cell can change from conductance state to insulating (and vice versa), a single resistive element can not only store a multi-valued data but also perform addition and subtraction.
Addition is performed by passing only the pulses corresponding to the bits that are set in the first operand, to the resistive cell storing the second operand. The voltage is applied in the direction that decreases the resistivity of the cell. Subtraction is performed by passing only the pulses corresponding to the bits that are set in the first operand, to the resistive cell storing the second operand. The voltage is applied in the direction that increases the resistivity of the cell. To exemplify this process, assume a resistive cell capable of storing four bits or 16 levels. We may subject the cell to four distinct pulses, each corresponding to the time interval it takes to set the cell to values 1, 2, 4, and 8, and to voltage applied per the operation direction (increase or decrease resistivity). Assume that the cell currently stores the value ‘1001’, and we wish to subtract from it the value ‘0110’. Out of the four available voltage pulses, we pass the second and third pulses in the direction that increases the resistivity of the cell. At the end of the operation, the cell contains the value ‘0011’.
Subtraction and addition can be used to perform long multiplication and division. RESET is accomplished by reading the current multi-valued value and subtracting from it the negated value. This unique capability potentially transforms the resistive crossbar to massively parallel multi-cell processor grid, capable of executing many calculations in parallel, in a small die.
Simulations and Comparative Analysis
In this section, we evaluate power and performance of ReGP-SIMD and compare it to CMOS GP-SIMD. In our simulations, we use the previously reported GP-SIMD cycle-accurate simulator [0031]. The simulator employs the ReGP-SIMD performance and power figures obtained by SPICE simulations. We use CMOS 22 nm process with operating frequency of 2.5 GHz, and single precision floating point arithmetic. To evaluate ReRAM characteristics, we use the TEAM model [0022][0024][0045]. Access time to the shared memory by the SIMD PU is significantly shorter than in CMOS GP-SIMD, due to the elimination of long wires (all word bits are laid above the corresponding PU). Access time for the sequential processor is the same in ReGP-SIMD as in the CMOS version; the memory is partitioned for this purpose[0031].
Dense Matrix Multiplication of a two √{square root over (N)}×√{square root over (N)} matrices (DMM) workload has been selected for our simulations, where N is the data set size, scaled for simplicity to the processor size, i.e., N=nGPSIMD. Note that simulations do not cover the cases where the data size exceeds the size of the processor (requiring off-chip data synchronization). For √{square root over (N)}×√{square root over (N)} dense matrix multiplication (DMM), the sequential execution time is O(N3/2). The GP-SIMD DMM algorithm utilizes N PUs and yields parallel execution time of O(N), with √{square root over (N)} data elements being broadcast every step [0030]. The GP-SIMD simulator is cycle based, recording the state of each register of each PU, and of the memory row assigned to it. Each command (for example, floating point multiply) is broken down to a series of fine-grain single bit PU operations. In a similar manner to SimpleScalar [006], the simulator also keeps track of the registers, buses and memory cells that switch during execution. With the switching activity and area power models of each baseline operation detailed in [0031], the simulator tracks the total energy consumed during workload execution.
We follow the assumptions detailed in [0031] and in section [00371], and simulate speedup and power for growing datasets corresponding to the growing number of PUs and hence growing total area. The number of PUs as a function of the GP-SIMD area for the CMOS and resistive designs are depicted in graph 211 of
The power consumption and the power efficiency results are depicted in graphs 215 and 216 of
a (mm )
of PUs (×106)
Dimension ({square root over (N)})
S
W)
ensity (W/mm2)
fficiency (GFLOPs/W)
indicates data missing or illegible when filed
To address the high power requirement of ReGP-SIMD, we consider a version constrained to dissipating 150 W in Table 2. It requires only 62 mm2. A same area CMOS version achieves only ˜1/25 of its density and ˜1/5 of its performance. Power efficiency remains ˜1:3 in favor of the CMOS GP-SIMD, but both versions exhibit better DMM power efficiency (50.8 and 15.2 GFLOPS/W) than NVidia GTX480 GPU (5 GFLOPS/W according to [0011] scaled to 22 nm CMOS).
When both area and power constraints are applied, CMOS and resistive GP-SIMD exhibit the relative ratios shown in TABLE 9 (the highlighted cells shows the limiting factor of each architecture).
Subtraction and addition can be used to perform long multiplication and division. RESET is accomplished by reading the current multi-valued value and subtracting from it the negated value. This unique capability potentially transforms the resistive crossbar to massively parallel multi-cell processor grid, capable of executing many calculations in parallel, in a small die.
Cycle Accurate Simulation, SpMM
In the previous section, ReGP-SIMD and CMOS GP-SIMD were simulated using the compute intensive Dense Matrix Multiplication (DMM) benchmark. In this section, ReGP-SIMD and CMOS GP-SIMD are compared using a Sparse Matrix Multiplication (SpMM) benchmark, leading to irregular read and write patterns. Our cycle accurate simulation results are compared with those of nVidea K20 [0037], Intel XEON PHI [0037] and the Associative Processor (AP) [0044]. We followed the methodology outlined in [0030] and simulated 1,000 floating-point square matrices with the number of nonzero elements spanning from one hundred thousand to eight million, randomly selected from the collection of sparse matrices from the University of Florida [0014].
As detailed in ]28[, the sparse N×M matrix multiplication algorithm has a computational complexity of O(NNNZMNNNZ log M), where NNNZ is the number of nonzero rows of the multiplier matrix A and MNNZ is the average number of nonzero elements per row (equal to number of nonzero elements of the multiplier matrix A, denoted by ANNZ, divided by NNNZ), and the efficiency grows with the number of multiplier matrix's average nonzero elements per row, and with the number of multiplicand matrix's columns. Further, GP-SIMD performance depends on data word-length rather than on data set size. If matrix elements are presented in a floating-point format, the word-length is 32 bit (IEEE754 single precision). Data set size in SpMM typically equals the number of nonzero elements in the sparse matrix.
Although the efficiency of the GP-SIMD SpMM algorithm [0030] grows with the number of columns of the multiplicand matrix, for fair comparison we limit our analysis to multiplicand matrices of 16 columns, as used in [0037]. Further, we assume that both ReGP-SIMD and CMOS GP-SIMD integrate 8 million PUs. We consider test-case sparse multiplier matrices having 1M columns or less and up to 8M nonzero elements, and dense multiplicand matrices having 16 columns. Figure Error!Reference source not found.(c) presents the sparse by dense matrix multiplication execution times for these matrices.
The spread in execution times depicted in Figure (c) stems from the sensitivity of the GP-SIMD SpMM algorithm to the average number of non-zero elements per row. This sensitivity is shared, although possibly to a lesser extent, by conventional SpMV and SpMM implementations (on GPU or multi-core) [0021][0027]. Since the average of nonzero elements per row in our test-set is somewhat capped (see Figure (b)), as the number of nonzero elements grows, so does the average number of rows and columns of matrices. The execution time of Broadcast command (see [0030]) depends on both the number of columns of matrix A and the number of rows. As the matrix gets larger (large number of rows each having a small number of nonzero elements), the execution time hinders the power efficiency of GP-SIMD. For two matrices with a similar number of nonzero elements, the difference of two orders of magnitude in the average number of nonzero elements per row results in a similar difference in the execution time.
The performance of the GP-SIMD SpMM algorithm as a function of the average number of nonzero elements per row is presented in Figure (d). The figure demonstrates a close to logarithmic dependency of the GP-SIMD SpMM algorithm performance on the average number of nonzero elements per row. Hence, if the average number of nonzero elements per row is small (which is consistently the case in the University of Florida collection matrices), the effectiveness of the ReGP-SIMD and CMOS GP-SIMD SpMM algorithm is limited. Also shown in Figure (d) are performance data for AP and two commercial processors (Intel XEON-PHI and K20 [0037]). The simulated power consumption of the GP-SIMD SpMM and AP (as well as reported power of NVidia K20 [0015]) is presented in Figure (e).
The ReGP-SIMD and GP-SIMD SpMM power efficiency is in the range of 0.1 to 100 GFLOPS/W (Figure (f)) with ReGP-SIMD taking the lead thanks to power efficient resistive implementation. The power efficiency declines with the number of nonzero elements (requiring higher power consumption). The SpMM/SpMV power efficiency of advanced contemporary GPUs such as NVidia's K20 and GTX660 is in the 0.1-0.5 GFLOPS/W range [0015]. A wide variety of multicore processors such as quad-core AMD Opteron 2214, quad-core Intel Xeon E5345, eight-core Sun UltraSparc T2+T5140 and eight-SPE IBM QS20 Cell reportedly reach the SpMM power efficiency of up to 0.03 GFLOPS/W [0042]. Several FPGA SpMV and SpMM implementation were proposed (for example [0030][0012]), however these studies focused on optimization of performance or energy-delay, and power dissipation figures were not reported.
The ReGP-SIMD, GP-SIMD and AP power efficiency advantage stem from in-memory computing (there are no data transfers between processing units and memory hierarchies) and from low-power design made possible by the very small size of each processing unit.
GP-SIMD Vs. ReGP-SIMD Having Identical PU Count
At area of 196 mm2, the CMOS GP-SIMD integrates 8.1M PUs and achieves maximum DMM performance of 337 GFLOPS while consuming only 35.9 W. The Resistive GP-SIMD having the same number of PUs occupies only 8 mm2 and achieves the same DMM performance of 337 GFLOPS while consuming only 29.4 W (TABLE 6).
esistive
CMOS
P-SIMD
P-SIMD
sistive/CMOS)
a (mm2)
of PUs (×106)
Dimension ({square root over (N)})
S
W)
ensity (W/mm2)
fficiency (GFLOPs/W)
indicates data missing or illegible when filed
Given the same number of PUs, the CMOS GP-SIMD occupies area 24.5 times larger than the resistive version and achieves only about four-fifths of the power efficiency.
The significant area savings of ReGP-SIMD may outweigh the additional cost of the resistive layer making ReGP-SIMD the preferred solution for cost effective implementation.
Area Constrained ReGP-SIMD
At area of 200 mm2, the CMOS GP-SIMD contains 8.2M PUs and achieves maximum DMM performance of 340 GFLOPS while consuming only 36.6 W (TABLE 7). Given the same area, the ReGP-SIMD would consume unrealistic power of ˜830 W. At these figures, ReGP-SIMD achieves about 5 times the performance of the CMOS version, but only about one-fifth of its power efficiency.
esistive
CMOS
P-SIMD
P-SIMD
sistive/CMOS)
a (mm2)
of PUs (×106)
Dimension ({square root over (N)})
S
W)
ensity (W/mm2)
fficiency (GFLOPs/W)
indicates data missing or illegible when filed
Power Ratio Analysis
CMOS GP-SIMD and the ReGP-SIMD software models are identical, and for the purpose of simulation, the very same code is executed on both, making the number of read/write operations to the shared memory array identical. TABLE 7 states that while the number of ReGP-SIMD PUs is 24.5× larger than the number of CMOS GP-SIMD for the same silicon area, power dissipation of ReGP-SIMD is only 22.7× higher than that of CMOS GP-SIMD. This effect is explained by observing DMM implementation on both architectures.
For large matrices, broadcast consumes the largest part of compute time.
Broadcast consists of many reads and a single write [0030]. Since the energy for reading ReRAM cells is significantly lower than reading CMOS memory, power increases less than the increase in the number of PUs.
Power Constrained ReGP-SIMD
To address the high power requirement of ReGP-SIMD, we consider a version constrained to dissipating 150 W in Error!Reference source not found. It requires only 39 mm2. A same area CMOS version achieves only ˜1/25 of its density and ˜1/5 of its performance. Power efficiency remains ˜1:5 in favor of the CMOS GP-SIMD, but both versions exhibit similar or better DMM power efficiency (23.6 and 4.8 GFLOPS/W) than NVidia GTX480 GPU (5 GFLOPS/W according to [0011] scaled to 22 nm CMOS).
esistive
CMOS
P-SIMD
sistive/CMOS)
a (mm2)
of PUs (×106)
Dimension ({square root over (N)})
S
W)
ensity (W/mm2)
fficiency (GFLOPs/W)
0 Performance (GFLOPs/W)
indicates data missing or illegible when filed
Area and Power Constrained ReGP-SIMD
When both area and power constraints are applied, CMOS and resistive GP-SIMD exhibit the relative ratios shown in TABLE 9 (the highlighted cells shows the limiting factor of each architecture).
We note that the GP-SIMD retains the entire data structure (two large dense matrices, temporaries and program variables) on chip, and thus neither time nor power are spent for data exchange with off-chip memory. Note further that a 200M rows GP-SIMD array having 256 bits per row translate to 6.4 GByte on-chip memory. However, the architect may increase the number of bits per row to 475 (matching PU width) without changing the characteristics of ReGP-SIMD. Additional layers may be used to increase the overall ReGP-SIMD on-chip storage, further reducing off-chip communication bandwidth (and hence power) in cases of very large work-load. For example, a 200 mm2 die may have 11.8 GByte (475 bits per row) on-chip storage, 23.6 GBytes with double ReRAM layer, and so on.
ReRAM Multi-valued Arithmetic
Multi-valued resistive memory may become available in the future [0029][0039][0010][0043][0047][0035]. Certain such devices may portray a linear range of operation: the resistivity may be set to a value k by applying a certain fixed voltage for time kt0(a ‘pulse’), where to is a constant time unit. Moreover, the device is additive: If a first pulse lasting k1 time units is followed by a second pulse lasting k2 time units, the resulting stored value is k1+k2 (as long as the result lies within the range of allowed values for the multi-valued device). Reading a multi-valued cell requires converting the voltage to a digital number, e.g. by means of an analog-to-digital converter at the periphery of the ReRAM array. Writing, on the other hand, calls for converting a digital number into pulse duration.
Addition may be performed by passing a timed voltage pulse corresponding to the first operand, to the resistive cell storing the second operand. The voltage is applied in the direction that decreases the resistivity of the cell. Subtraction is performed similarly, by applying the voltage in the opposite direction. To exemplify the accumulation process, assume a resistive cell capable of storing four bits or 16 levels. If the cell stores 9, and we wish to subtract 6, we subject the cell to a 6t0 voltage pulse in the direction that increases the resistivity of the cell. At the end of the operation, the cell contains 3. Since no carry operation is implemented, this is saturation arithmetic (note that we assume power-of-2 multi-valued cells).
The single bit ReRAM cell dissipates about 1fJ when driven from RON to ROFF and vice versa. Multi-valued arithmetic requires driving the cells from and to levels within the ReRAM's dynamic range limited by RON and ROFF. Thus, a single cell arithmetic operation may dissipate up to 1fJ. However, the peripheral circuitry incorporates A/D converter for reading out the contents of the multi-valued cell, taking larger area and consuming more power. These parameters have been simulated and validated using the TEAM model [0022][0024][0045] and the SPICE simulator.
Assume N bit multivalued ReRAM Cell storing a 2's complement binary number in which the last (most significant) bit represents the overflow (“O”), and the second to last bit represents sign (“S”). The addition and subtraction flow diagram is depicted in. Initially, the first operand (A) is stored in the ReRAM cell. The cell is then subjected to a timed pulse corresponding to the second operand (B). The cell accumulates both operands and retains the result in-place. The result is read back to the PU, and subsequently overflow condition is tested as follows:
The PU retains the sign bit of the operands (OPA_SIGN, OPB_SIGN) and result sign bit (RES_SIGN). Following arithmetic operation,
If (OPA_SIGN==OPB_SIGN==0) && (RES_SIGN==1)
//both operands are positive and the result is negative
Overflow detected
elseif (OPA_SIGN==OPB_SIGN==1) && (RES_SIGN=0)
//both operands are negative and the result is positive
Overflow detected
Reset “O” bit
The overflow detection protects from overflowing the cell's dynamic range. If overflow is found, saturation is applied followed by write-back. When the overflow (“O” bit) becomes “1”, the ReRAM is driven with sufficient interval to reset it back to zero, preserving the result that has not over-flown. Thus, fixed point m bit addition/subtraction consumes, worst case, four cycles, regardless of the ReGP-SIMD vector size. Note that we assume that it takes a single cycle to switch the resistive cell from low to high resistivity (i.e., from the value 0 to N−1 in N bit multivalued cell). A long word containing K bits (including the sign bit), where K>N−1, may be partitioned into L=[K/(N−1)] ReRAM cells. The long addition and subtraction flow diagram is depicted in
Fixed point single-cell multiplication and division in multi-valued ReGP-SIMD are computed serially, performing a series of add-shift and subtract-shift vector operations (see flow diagram in 29). When multiplying A×B, and assuming the product may be stored in a single ReRAM cell (that is, log2 A×B<N−1), both operands are fetched to the PU. If the ith bit of B is one, the operand A (left-shifted by i places) is added to the accumulator which resides in the ReRAM. For each i, the loop takes five cycles. Thus, fixed point m×m bit vector multiplication requires 1+5m∈O(m) cycles. The above may be extended to long word operands following the flow-diagram of
The potential energy and area savings of multi-valued ReRAM computation is a promise worth investigating. In a standard 45 nm CMOS process, a register file access consume 6pJ (12pJ for read and write) while an 8 bit integer addition consume only 0.03pJ [0017]. The total 8 bit load/add/store operation thus entails 12.03pJ. With an 8 bit multi-bit ReRAM cell, the energy for the same operation is expected to be less than 1fJ, four orders of magnitude lower energy than typical CMOS. More so, the addition is performed on a tiny 4F2 cell, with processing time similar to SRAM cell access time.
This ReRAM multi-valued arithmetic capability can potentially transform the resistive crossbar to a massively parallel multi-cell processor grid, capable of executing many calculations in parallel, in a small footprint. Further, multi-valued arithmetic may significantly improve ReGP-SIMD power efficiency as the operations are performed right on the cell, concurrently for the entire cell-column, rather than moving information in bit-wise manner back and forth between the cell-column and the PUs.
GP-SIMD architecture combines a general purpose sequential processor and a massive parallel SIMD co-processor with a large shared memory array. Since the shared memory is accessible by both processors, data is not transferred (synchronized) back and forth between separate memories of the two processors. Unlike conventional processors or SIMD accelerators, GP-SIMD's core component (and by far, the largest area-wise) is the shared memory array, thus it stands to benefit the most from a reduced bit-cell architecture offered by a resistive crossbar.
This paper explores a resistive design of the GP-SIMD (ReGP-SIMD), which has the potential to scale the GP-SIMD from a few millions of PUs to a few hundred millions of PUs on a single silicon die. In resistive GP-SIMD, a novel resistive row and column addressable 4F2 crossbar is utilized, replacing the modified CMOS 190F2 SRAM storage previously proposed for GP-SIMD architecture. We find that PiM architectures and specifically, GP-SIMD, benefit more than other many-core architectures from using resistive memory.
Our simulations show that although high power density and finite endurance of the ReRAM and the CMOS PU limit the potential of ReGP-SIMD, it allows significantly better scalability and performance as compared to a CMOS GP-SIMD as well as other conventional SIMD accelerators. Future progress in ReRAM technologies is likely to offer continuous scalability, improved power efficiency, and higher endurance for ReRAM and thus further enhance ReGP-SIMD. A hybrid CMOS-memristor PUs are possible to further scale the ReGP-SIMD. Additional notable advantage is the placement of the PUs under the resistive shared memory crossbar leading to a significant area savings over the CMOS GP-SIMD. Finally, we noted that the large ReGP-SIMD on-chip memory helps to reduce the latency and power consumption of addressing an off-chip memory.
A conceptual multi-valued arithmetic in a single resistive storage device is explored. Such a device can potentially transform the resistive crossbar to massively parallel multi-cell processor grid, capable of executing a large amount of calculations in parallel, in a small die. Future research is needed to optimize such devices and to devise PiM architectures that realize and extend these density and performance potentials.
In resistive GP-SIMD, a novel resistive row and column addressable 4F2 crossbar is utilized, replacing the modified CMOS 190F2 SRAM storage previously proposed for GP-SIMD architecture. We find that PiM architectures and specifically, GP-SIMD, benefit more than other many-core architectures from using resistive memory.
GP-SIMD is a novel processor architecture that combines a general purpose sequential processor, a massively parallel SIMD array, and a large memory shared by both processors. The shared memory may be accessed simultaneously by the sequential processor (providing a row address and accessing the data columns) and by the SIMD processor (providing a column address while each processing unit, PU, accesses a memory row). Thus, data do not need to be transferred (synchronized) back and forth between separate memories of the two processors.
CMOS GP-SIMD is compared with two other massively parallel, processing-in-memory (PIM) architectures: Conventional SIMD (CSIMD) employing separate memories for the SIMD and sequential processors and requiring data synchronization (the SIMD array comprising a moderate number of FPUs, a-la GPU), and an Associative Processor (AP) combining storage and processing in each cell. Comparison is made by means of analytical study, and a cycle-accurate simulator is employed to validate performance and energy of four benchmarks on GP-SIMD.
GP-SIMD speedup grows faster with area than that of CSIMD due to finer granularity of its processing unit, resulting in more processing units per equivalent area. GP-SIMD speedup also grows faster with area than that of the Associative Processor due to the finer granularity of its storage elements, resulting in larger memory per equivalent area. Unlike CSIMD, the performance of GP-SIMD depends on the data word-length rather than data set size. As area budget grows beyond the speedup breakeven point, the GP-SIMD's energy remains below that of the CSIMD coprocessor, and it thus outperforms CSIMD in terms of speedup/energy ratio over a broad spectrum of area and power budget for workloads with high data-level parallelism. In general, GP-SIMD achieves higher performance than CSIMD due to CSIMD stalling on synchronization. GP-SIMD achieves higher performance-per-area than AP due to its lower memory cell area, allowing the integration of twice the amount of PUs and memory rows per given area, and its more than twice faster floating point implementation due to GP-SIMD arithmetic PU vs. associative-only PU of AP. GP-SIMD however is not universally efficient. While yielding high speed up when implementing fine-grain massively data-parallel workloads (such as sparse linear algebra and machine learning algorithms), its efficiency is much lower under workloads with low data-level parallelism, or where little or no data synchronization is required. To conclude, GP-SIMD architecture may achieve significantly higher performance or power efficiency over other conventional architectures.
GP-SIMD architecture combines a general purpose sequential processor and a massive parallel SIMD co-processor with a large shared memory array. Since the shared memory is accessible by both processors, data is not transferred (synchronized) back and forth between separate memories of the two processors. Unlike conventional processors or SIMD accelerators, GP-SIMD's core component (and by far, the largest area-wise) is the shared memory array, thus it stands to benefit the most from a reduced bit-cell architecture offered by a resistive crossbar.
This application explores a resistive design of the GP-SIMD (ReGP-SIMD), which has the potential to scale the GP-SIMD from a few millions of PUs to a few hundred millions of PUs on a single silicon die. Our simulations show that although high power density and finite endurance of the ReRAM and the CMOS PU limit the potential of ReGP-SIMD, it allows significantly better scalability and performance as compared to a CMOS GP-SIMD as well as other conventional SIMD accelerators. Future progress in ReRAM technologies is likely to offer continuous scalability, improved power efficiency, and higher endurance for ReRAM and thus further enhance ReGP-SIMD. A hybrid CMOS-memristor PUs are possible to further scale the ReGP-SIMD. Additional notable advantage is the placement of the PUs under the resistive shared memory crossbar leading to a significant area savings over the CMOS GP-SIMD. Finally, we noted that the large ReGP-SIMD on-chip memory helps to reduce the latency and power consumption of addressing an off-chip memory.
A conceptual multi-valued arithmetic in a single resistive storage device is explored. Such a device can potentially transform the resistive crossbar to massively parallel multi-cell processor grid, capable of executing a large amount of calculations in parallel, in a small die. Future research is needed to optimize such devices and to devise PiM architectures that realize and extend these density and performance potentials.
In the foregoing specification, the invention has been described with reference to specific examples of embodiments of the invention. It will, however, be evident that various modifications and changes may be made therein without departing from the broader spirit and scope of the invention as set forth in the appended claims.
Moreover, the terms “front,” “back,” “top,” “bottom,” “over,” “under” and the like in the description and in the claims, if any, are used for descriptive purposes and not necessarily for describing permanent relative positions. It is understood that the terms so used are interchangeable under appropriate circumstances such that the embodiments of the invention described herein are, for example, capable of operation in other orientations than those illustrated or otherwise described herein.
The connections as discussed herein may be any type of connection suitable to transfer signals from or to the respective nodes, units or devices, for example via intermediate devices. Accordingly, unless implied or stated otherwise, the connections may for example be direct connections or indirect connections. The connections may be illustrated or described in reference to being a single connection, a plurality of connections, unidirectional connections, or bidirectional connections. However, different embodiments may vary the implementation of the connections. For example, separate unidirectional connections may be used rather than bidirectional connections and vice versa. Also, plurality of connections may be replaced with a single connection that transfers multiple signals serially or in a time multiplexed manner. Likewise, single connections carrying multiple signals may be separated out into various different connections carrying subsets of these signals. Therefore, many options exist for transferring signals.
Although specific conductivity types or polarity of potentials have been described in the examples, it will be appreciated that conductivity types and polarities of potentials may be reversed.
Each signal described herein may be designed as positive or negative logic. In the case of a negative logic signal, the signal is active low where the logically true state corresponds to a logic level zero. In the case of a positive logic signal, the signal is active high where the logically true state corresponds to a logic level one. Note that any of the signals described herein may be designed as either negative or positive logic signals. Therefore, in alternate embodiments, those signals described as positive logic signals may be implemented as negative logic signals, and those signals described as negative logic signals may be implemented as positive logic signals.
Furthermore, the terms “assert” or “set” and “negate” (or “deassert” or “clear”) are used herein when referring to the rendering of a signal, status bit, or similar apparatus into its logically true or logically false state, respectively. If the logically true state is a logic level one, the logically false state is a logic level zero. And if the logically true state is a logic level zero, the logically false state is a logic level one.
Those skilled in the art will recognize that the boundaries between logic blocks are merely illustrative and that alternative embodiments may merge logic blocks or circuit elements or impose an alternate decomposition of functionality upon various logic blocks or circuit elements. Thus, it is to be understood that the architectures depicted herein are merely exemplary, and that in fact many other architectures may be implemented which achieve the same functionality.
Any arrangement of components to achieve the same functionality is effectively “associated” such that the desired functionality is achieved. Hence, any two components herein combined to achieve a particular functionality may be seen as “associated with” each other such that the desired functionality is achieved, irrespective of architectures or intermedial components. Likewise, any two components so associated can also be viewed as being “operably connected,” or “operably coupled,” to each other to achieve the desired functionality.
Furthermore, those skilled in the art will recognize that boundaries between the above described operations merely illustrative. The multiple operations may be combined into a single operation, a single operation may be distributed in additional operations and operations may be executed at least partially overlapping in time. Moreover, alternative embodiments may include multiple instances of a particular operation, and the order of operations may be altered in various other embodiments.
Also for example, in one embodiment, the illustrated examples may be implemented as circuitry located on a single integrated circuit or within a same device. Alternatively, the examples may be implemented as any number of separate integrated circuits or separate devices interconnected with each other in a suitable manner.
However, other modifications, variations and alternatives are also possible. The specifications and drawings are, accordingly, to be regarded in an illustrative rather than in a restrictive sense.
In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word ‘comprising’ does not exclude the presence of other elements or steps then those listed in a claim. Furthermore, the terms “a” or “an,” as used herein, are defined as one or more than one. Also, the use of introductory phrases such as “at least one” and “one or more” in the claims should not be construed to imply that the introduction of another claim element by the indefinite articles “a” or “an” limits any particular claim containing such introduced claim element to inventions containing only one such element, even when the same claim includes the introductory phrases “one or more” or “at least one” and indefinite articles such as “a” or “an.” The same holds true for the use of definite articles. Unless stated otherwise, terms such as “first” and “second” are used to arbitrarily distinguish between the elements such terms describe. Thus, these terms are not necessarily intended to indicate temporal or other prioritization of such elements. The mere fact that certain measures are recited in mutually different claims does not indicate that a combination of these measures cannot be used to advantage.
While certain features of the invention have been illustrated and described herein, many modifications, substitutions, changes, and equivalents will now occur to those of ordinary skill in the art. It is, therefore, to be understood that the appended claims are intended to cover all such modifications and changes as fall within the true spirit of the invention.
This application claims priority from U.S. provisional patent 62/100,967 having a filing date of Jan. 8, 2015 which is incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| 62100967 | Jan 2015 | US |
