MULTI-BIT DOT PRODUCT ENGINE

Abstract

A method of obtaining an in-memory vector-based dot product is disclosed, which includes providing a matrix of memory cells having M rows, each memory cell in each row holding a value and having dedicated read transistors T1 and T2, where T1 is controlled by the value held in the associated memory cell and T2 is controlled by a row-dedicated source (vin) for each row, the combination of the T1 and T2 transistors for each cell selectively (i) couple a reference voltage with a column-dedicated read bit line (RBL) for each column for an in-memory vector-based dot product operation or (ii) couple ground with the column-dedicated read bit line (RBL) for each column for a memory read operation, where total resistance of the read transistors (R) for each cell in each row is based on Rmax/2(M-1), . . . Rmax, where Rmax is the resistance of the least significant cell in each row.

Description

This invention was not made with government support.

TECHNICAL FIELD

The present disclosure generally relates to electronic memory cells, and in particular, to electronic memory cells capable of providing in-memory arithmetic operation.

BACKGROUND

This section introduces aspects that may help facilitate a better understanding of the disclosure. Accordingly, these statements are to be read in this light and are not to be understood as admissions about what is or is not prior art.

Traditionally, a computer's architecture has included (i) a processor, (ii) memory, and (iii) other peripheral components. This architecture is often referred to as the von-Neumann architecture, after its inventor. This architecture is based on decoupling the memory from the processor, and is found in millions of computers worldwide. A schematic of this architecture 1 is shown in FIG. 1, where a memory 2 is decoupled from the processor 4, and where the memory 2 and the processor 4 communicate via data exchange bus 6. While there has been an explosion of computing powers using this architecture, it is now failing to answer the call of data exchange in such applications as big-data, artificial intelligence, Internet-of-things (IoT), and other data intensive operations. The shortcoming of this architecture is mainly due to a bottleneck 8 found between the memory 2 and the processor 4. This bottleneck 8 mainly arises from a limited data transfer rate between these two physically decoupled units. There are two major challenges with this bottleneck: (1) cycle time needed for data transfer (i.e., throughput), and (2) energy consumption associated with data transfer. To combat these two drawbacks associated with the bottleneck 8, one approach investigated in the prior art is to utilize in-memory vector computations. In such cases, not only does the decoupled memory maintain data it also provides rudimentary logic operations. These in-memory operations, thus, remove some of the issues associated with the bottleneck 8 as (1) data transfer between the memory 2 and the processor 4 is reduced, and (2) consequently energy consumption is also reduced.

With respect to the CMOS technology, Boolean in-memory operations have been in various architectures. For example, in one such architecture, a vector Boolean operations is realized using six transistor static random access memory (6T SRAM) cells. In another such architecture, 8T SRAM cells were used as vector compute primitives, specifically, owing to their decoupled read and write ports. While these architectures are based on vector Boolean operations, the most frequent and compute intensive function required for numerous applications such as machine learning is the dot product operation. Few works based on analog computations in SRAM cells can be found in the prior art. These works use 6T SRAM cells and rely on the resultant accumulated voltage on the bit-lines (BLs). Not only 6T SRAMs are prone to read-disturb failures, the failures are also a function of the voltage on the BLs. This leads to a tightly constrained design space for the proposed 6T SRAM based analog computing. Moreover, multi-bit, highly parallel dot products have yet not been attained in existing SRAM based in-memory works.

Therefore, there is an unmet need for a novel approach using traditional SRAM memory cells that provide more robust approaches for dot product.

SUMMARY

A multi-bit dot product engine-based memory for carrying out vector-based dot products is disclosed. The engine includes a matrix of memory cells having M rows (M-1, M-2, . . . 0) and N columns (N-1, N-2, . . . 0), each memory cell (C) in each of the rows ((CN-1, CN-2, . . . , C₀)_M-1, (CN-1, CN-2, . . . , C₀)_M-2, . . . (CN-1, CN-2, . . . , C₀)₀) holding a value and having dedicated read transistors T1 and T2, where T1 is controlled by the value held in the associated memory cell and T2 is controlled by a row-dedicated read word line RWL_i (RWL_M-1, RWL_M-2, . . . RWL₀), the combination of the T1 and T2 transistors for each cell selectively (i) couple a row-dedicated source (v_in) for each row ((v_in)_M-1, (v_in)_M-2, . . . (v_in)₀) with a column-dedicated read bit line (RBL) for each column (RBL_N-1, RBL_N-2, . . . RBL₀) for an in-memory vector-based dot product operation or (ii) couple ground with the column-dedicated read bit line (RBL) for each column (RBL_N-1, RBL_N-2, . . . RBL₀) for a memory read operation, where total resistance of the T1 and T2 transistors (R) for each cell (R_N-1, R_N-2, . . . , R₀) in each row (M-1, M-2, . . . 0) is based on Rmax/2^(M-1), Rmax/2^(M-2), . . . Rmax, where Rmax is the resistance of the least significant cell in each row and where current in each RBL (I_RBL) for each column ((I_RBL)_N-1, (I_RBL)_N-2, . . . (I_RBL)₀) is summed as I_out. The engine also includes a sensing circuit coupled to the matrix of memory cells and adapted to selectively (i) receive and convert the current I_out to a digital value representing an output of the vector-based dot product, or (ii) sense voltage at each of the RBLs (RBL_N-1, RBL_N-2, . . . RBL₀) to read value of the corresponding cells.

Another multi-bit dot product engine-based memory for carrying out vector-based dot products is also disclosed. The engine includes a matrix of memory cells having M rows (M-1, M-2, . . . 0) and N columns (N-1, N-2, . . . 0), each memory cell (C) in each of the rows ((C_N-1, C_N-2, . . . , C₀)_M-1, (C_N-1, C_N-2, . . . , C₀)_M-2, . . . (C_N-1, C_N-2, . . . , C₀)₀) holding a value and having dedicated read transistors T1 and T2, where T1 is controlled by the value held in the associated memory cell and T2 is controlled by a row-dedicated source (v_in) for each row ((v_in)_M-1, (v_in)_M-2, . . . (v_in)₀), the combination of the T1 and T2 transistors for each cell selectively (i) couple a reference voltage (V_bias) with a column-dedicated read bit line (RBL) for each column (RBL_N-1, RBL_N-2, . . . RBL₀) for an in-memory vector-based dot product operation or (ii) couple ground with the column-dedicated read bit line (RBL) for each column (RBL_N-1, RBL_N-2, . . . RBL₀) for a memory read operation, where total resistance of the T1 and T2 transistors (R) for each cell (R_N-1, R_N-2, . . . , R₀) in each row (M-1, M-2, . . . 0) is based on Rmax/2^(M-1), Rmax/2^(M-2), . . . Rmax, where Rmax is the resistance of the least significant cell in each row and where current in each RBL (I_RBL) for each column ((I_RBL)_N-1, (I_RBL)_N-2, . . . (I_RBL)₀) is summed as I_out. The engine also includes a sensing circuit coupled to the matrix of memory cells and adapted to selectively (i) receive and convert the current I_out to a digital value representing an output of the vector-based dot product, or (ii) sense voltage at each of the RBLs (RBL_N-1, RBL_N-2, . . . RBL₀) to read value of the corresponding cells.

A method of obtaining an in-memory vector-based dot product is also disclosed. The method includes providing a matrix of memory cells having M rows (M-1, M-2, . . . 0) and N columns (N-1, N-2, . . . 0), each memory cell (C) in each of the rows ((C_N-1, C_N-2, . . . , C₀)_M-1, (C_N-1, C_N-2, . . . , C₀)_M-2, . . . (C_N-1, C_N-2, . . . , C₀)₀) holding a value and having dedicated read transistors T1 and T2, where T1 is controlled by the value held in the associated memory cell and T2 is controlled by a row-dedicated source (v_in) for each row ((v_in)_M-1, (v_in)_M-2, . . . (v_in)₀), the combination of the T1 and T2 transistors for each cell selectively (i) couple a reference voltage (V_bias) with a column-dedicated read bit line (RBL) for each column (RBL_N-1, RBL_N-2, . . . RBL₀) for an in-memory vector-based dot product operation or (ii) couple ground with the column-dedicated read bit line (RBL) for each column (RBL_N-1, RBL_N-2, . . . RBL₀) for a memory read operation, where total resistance of the T1 and T2 transistors (R) for each cell (R_N-1, R_N-2, . . . , R₀) in each row (M-1, M-2, . . . 0) is based on Rmax/2^(M-1), Rmax/2^(M-2), . . . Rmax, where Rmax is the resistance of the least significant cell in each row and where current in each RBL (I_RBL) for each column ((I_RBL)_N-1, (I_RBL)_N-2, . . . (I_RBL)₀) is summed as I_out. The method also includes selectively (i) receiving and converting the current I_out to a digital value representing an output of the vector-based dot product, or (ii) sensing voltage at each of the RBLs (RBL_N-1, RBL_N-2, . . . RBL₀) to read value of the corresponding cells.

BRIEF DESCRIPTION OF DRAWINGS

FIG. 1 is a schematic of a computing architecture according to the prior art.

FIG. 2 is a generalized schematic of a memory cell.

FIG. 3 is the generalized schematic of memory cell of FIG. 2 with access transistors.

FIG. 4 is a transistor-based schematic of the memory cell of FIG. 3 depicting a six transistor SRAM.

FIG. 5 is the generalized schematic of memory cell of FIG. 2 with dedicated read access transistors, depicting a generalized eight transistor (8T) SRAM.

FIG. 6 is a transistor-based schematic of the memory cell of FIG. 5.

FIG. 7 is a schematic of a first configuration (Config-A) modifying the dedicated read access transistors of the 8T SRAM of FIG. 6, according to the present disclosure.

FIG. 8 is a schematic of a second configuration (Config-B) modifying the dedicated read access transistors of the 8T SRAM of FIG. 6, according to the present disclosure.

FIG. 9a is a schematic of a matrix of cell having rows and columns, according to the present disclosure, employing the Config-A configuration, wherein current from each column is summed to determine a resultant from a vector-based dot product based on an input voltage (V_in) and the value held in the corresponding memory cells, where each memory cell in each row is given a proportional weight.

FIG. 9b is a schematic of a matrix of cell having rows and columns, according to the present disclosure, employing the Config-B configuration, wherein current from each column is summed to determine a resultant from a vector-based dot product based on an input voltage (V_in) and the value held in the corresponding memory cells, where each memory cell in each row is given a proportional weight.

FIG. 10a is a graph of the summed current of FIG. 9a vs. V_in.

FIG. 10b is a graph of the summed current of FIG. 9a vs. the corresponding proportional weights.

FIG. 10c is a graph of the summed current of FIG. 9a vs. the number of rows.

FIG. 11a is a graph of the summed current of FIG. 9b vs. V_in.

FIG. 11b is a graph of the summed current of FIG. 9b vs. the corresponding proportional weights.

FIG. 11c is a graph of the summed current of FIG. 9b vs. the number of rows.

FIG. 12 is a schematic of a sensing circuit for Config-A, according to the present disclosure utilizing an operational amplifier (Opamp).

FIG. 13 is a schematic of a sensing circuit for Config-B, according to the present disclosure utilizing an Opamp.

FIG. 14a is a graph of the summed current of FIG. 9a vs. V_in utilizing the sensing circuit of FIG. 12.

FIG. 14b is a graph of the summed current of FIG. 9a vs. the corresponding proportional weights utilizing the sensing circuit of FIG. 12.

FIG. 14c is a graph of the summed current of FIG. 9a vs. the number of rows utilizing the sensing circuit of FIG. 12.

FIG. 15a is a graph of the summed current of FIG. 9b vs. V_in utilizing the sensing circuit of FIG. 13.

FIG. 15b is a graph of the summed current of FIG. 9b vs. the corresponding proportional weights utilizing the sensing circuit of FIG. 13.

FIG. 15c is a graph of the summed current of FIG. 9b vs. the number of rows utilizing the sensing circuit of FIG. 13.

DETAILED DESCRIPTION

For the purposes of promoting an understanding of the principles of the present disclosure, reference will now be made to the embodiments illustrated in the drawings, and specific language will be used to describe the same. It will nevertheless be understood that no limitation of the scope of this disclosure is thereby intended.

In the present disclosure, the term “about” can allow for a degree of variability in a value or range, for example, within 10%, within 5%, or within 1% of a stated value or of a stated limit of a range.

In the present disclosure, the term “substantially” can allow for a degree of variability in a value or range, for example, within 90%, within 95%, or within 99% of a stated value or of a stated limit of a range.

In the present disclosure terms switch and transistor are used interchangeably and are intended to refer to the same device.

A novel approach using traditional static read access memory (SRAM) cells that provide more robust approaches for dot product is provided in the present disclosure. An eight transistor (8T) SRAM topology is generally used, herein. Prior to describing operations of the 8T SRAM according to the present disclosure, operations of a simpler 6T SRAM cell is first described.

Referring to FIG. 2, a simple schematic of a memory cell 100 is presented based on two back-to-back inverters 102 and 104. Input port of the inverter 102 is identified as 102_i while its output port is identified as 102_o. Similarly, input port of the inverter 104 is identified as 104_i while its output port is identified as 104_o. The output port 102_o is connected to input port 104_i, and the output port 104_o is connected to input port 102_i. Given the disposition of the two inverters 102 and 104, a value that is at the input port 102_i is maintained (hence, a memory cell). For example, if the memory cell 100 maintains a value of 0, which translates to a value of 0 at the input port 102_i, then the output port 102_0 has a value of 1, which is connected to the input port 104_i, which results in a value of 0 at the output port 104_o, which is connected to the input port 102_i, thereby maintaining the value 0 in the memory cell 100. Therefore, as long as the input port 102_i has a value of 0, that value is maintained.

Now referring to FIG. 3, a schematic of a memory cell 150 is shown which includes the memory cell 100 of FIG. 2 with access switches (or transistors) 152 and 154 designed to read values from and write values to the memory cell 100. Each switch 152 and 154 has a corresponding gate that is coupled to a word line (WL). When the WL is activated, the switches 152 and 154 are allowed to selectively read the value of the input ports 102_i and 104_i, or over-write those values.

To further demonstrate the operations of a standard 6T SRAM, refer to FIG. 4, in which a memory cell 250 is shown where the two transistors (or switches) 152 and 154 of FIG. 3 are shown as 252 and 254 and the two inverters 102 and 104 of FIG. 3 are each provided as two P and N transistors (shown as 202_1, 202_2, and 204_1, 204_2, respectively). The input port (or node) 102_i is now shown as 202_i and the input port (or node) 104 is now shown as 204_i. Furthermore, two capacitors 252_c and 254_c, each connected to a corresponding switches 252 and 254 are shown in FIG. 4 as pre-charge capacitors used to change values as discussed below. In the memory cell 250, the node 202_i is connected to the gates of transistors 202_1 and 202_2 and the node 204_i is connected to the gates of transistors 204_1 and 204_2.

The memory cell 250 of FIG. 4 is now described with respect to four operations: Read 0, Read 1, Write 0, and Write 1. For the Read 0 operation, suppose the node 202_i which is equivalent to the input port 102_i (see FIG. 2) is set to 0. In this case, since node 202_i is connected the gates of transistors 202_1 and 202_2, transistor 202_1 is turned on, while transistor 202_2 is off. This scenario causes node 204_i (which is the output of the inverter designated as 202_1 and 202_2) to be raised to VDD (i.e., the source of transistor 202_1). With node 204_i raised to VDD, since node 204_i is connected to the gates of transistors 204_1 and 204_2, transistor 204_1 is off while transistor 204_1 is turned on, thereby lowering node 202_i (i.e., the output of the inverter represented by transistors 204_1 and 204_2) to ground. This scheme is maintained until a new value is written to the memory cell 250, as discussed below. In order to read the value of the memory cell 250, pre-charge capacitors 252_c and 254_c are both charged to VDD and the line identified as WL is turned on, thereby turning on access transistors 252 and 254. In this scenario, lines BL and BL are treated as outputs. By pre-charging the two capacitors, 252_c and 254_c, the voltage at BL and BL start at VDD. In this case, since the node 204_i is at the same voltage as BL (i.e., at VDD), there is no change, however, node 202_i is at 0 and BL is at VDD, thus the voltage at the pre-charge capacitor 254_c begins to discharge through the transistor 204_2. By comparing the value of BL and BL using a sense amplifier, and seeing a decrease in BL as compared to BL, it can be deduced that the value of node 202_i (i.e., the value held in the memory cell 250) is 0. Conversely, if the value in the memory cell 250 is 1 (i.e., node 202_i is at VDD, node 204_i is at ground), pre-charging capacitors 252_c and 254_c to VDD will cause both node 202_i and BL to be at VDD while node 204_i is at ground (i.e., 0) and pre-charge capacitor 252_c (i.e., BL) is at VDD. This scenario causes the pre-charge capacitor 252_c to discharge. By comparing the value of BL and BL using a sense amplifier, and seeing a decrease in BL as compared to BL, it can be deduced that the value of node 202_i (i.e., the value held in the memory cell 250) is 1.

In contrast to the read operation, the write operation requires treatment of BL and BL as inputs (as compared to outputs, as in the case of reading the memory cell 250). This time suppose, the value at node 202_i is 1. By forcing BL to ground, the node 202_i begins to discharge from 1 to 0. When the voltage becomes smaller than the threshold voltage of the transistor 202_2, it will be turned off, and at the same time 202_1 is turned on, which raises the node 204_i to 1. By doing so, since the node 204_i is connected to the gates of the transistors 204_1 and 204_2, transistor 204_1 is turned off and transistor 204_2 is turned on. This situations results in the node 202_i to be at 0, which means the memory cell 250 has been successfully changed its value from a 1 to a 0.

In the read situation, however, there is a challenge with the 6T SRAM in which the voltage at the gate of transistors 202_2 or 204_2, as the pre-charge capacitors are being discharged may be high enough to unintentionally turn on these transistors. For example, in the case where node 204_i is a 0 and the pre-charge capacitors 252_c and 254_c are charged to VDD, then when WL is activated, it is possible that as the pre-charge capacitor 252_c is charging, the voltage at node 204_i (i.e., the same voltage at the gate for the transistor 204_2) may be higher than the threshold voltage of the transistor 204_2, which would unintentionally turn on that transistor causing node 202_i to also become 0, which is an unwanted condition since nodes 204_i and 202_i are intended to always be opposite one-another.

This situation is typically called read-disturb failure. To overcome this unintended situations, the width of transistor 202_2 and 204_2 and 252 and 254 must be carefully chosen. Alternatively, the read and write operations can be decoupled in order to isolate the read operation. This alternative approach is known as the 8T SRAM with two additional transistors for said decoupling. Referring to FIG. 5, a schematic of an 8T SRAM 350 is provided which includes the memory cell 100 of FIG. 2 with access transistors (or switches) 352 and 354 designed to write values to the memory cell 100; and access transistors (or switches) 320 and 322 designed to read values from the memory cell 100. Each switch 352 and 354 has a corresponding gate that is coupled to a write word line (WWL), while switch has a gate coupled to a read word line (RWL) and switch 322 has a gate coupled to the input port 104_i. When WWL is activated, the switches 352 and 354 are allowed to over-write the value of the input ports 102_i and 104_i. Similarly, when RWL is activated, the switches 320 and 322 are allowed to read the value of the input port 104_i. One difference with the complementary FIG. 3 (depicting the 6T SRAM schematic) is that in FIG. 5 the BL lines are now WBL and WBL for write bit line and RBL for read bit line. Since the write and read are now decoupled, the robustness of the result 8T SRAM bit-cell is better that the 6T bit-cells in terms of read and write noise margins. Note, the source of transistor 322 is always connected to ground for 8T cells in prior-art, which as will be shown later represents one difference between the memory cell of the present disclosure and the memory cell used in the prior art.

To further demonstrate the operations of an 8T SRAM, refer to FIG. 6, in which a memory cell 450 is shown where the four transistors (or switches) 352, 354, 320, and 322 of FIG. 5 are shown as 452, 454, 420, and 422, respectively, and the two inverters 102 and 104 of FIG. 5 are each provided as two P and N transistors (shown as 402_1, 402_2, and 404_1, 404_2, respectively). The input port 102_i is now shown as 402_i and the input port 104 is now shown as 404_i. In the memory cell 450, the node 402_i is connected to the gates of transistors 402_1 and 402_2 and the node 404_i is connected to the gates of transistors 404_1 and 404_2.

The 8T SRAM cell shown in FIG. 6 is more robust as compared to the 6T cells due to isolated read port (i.e., the transistors 420, 422 coupled to the RBL and RWL). According to the present disclosure, it is possible to configure the 8T cell SRAM for in-memory dot product computations. It should be noted that in sharp contrast to the previous works related to in-memory computing with the CMOS technology, according to the present disclosure, a current based, analog-like dot product computations is enabled using the robust digital 8T bit-cells.

In particular, in a typical 8T SRAM, in a read operation, the line identified as RBL is capacitively pre-charged to a high value (VDD), when RWL is activated (i.e., a high voltage placed on RWL) and the source of the transistor 422 is connected to ground, if a value “0” is held in the memory cell, then the transistor 422 remains in the off state, thereby the voltage at the drain of the transistor 422 remains at high which indicates a “0” in the cell. If however, the memory cell has a “1”, then the transistor 422 is turned on, thereby the voltage at the drain of the transistor 422 discharges to ground through the transistor 422 which indicates a “1” in the cell. In either of these cases, the topology for a typical 8T SRAM is based on changes of voltage (no change at the drain of the transistor 422, indicating a “0” in the cell, and a reduction from a high voltage to about 0 indicating a “1” in the cell), with little or no current passing through the output transistors 420 and 422.

Configuring an 8T SRAM cell in particular configurations, according to the present disclosure and as described below allows a network of these cells to be used as a multi-bit dot product engine. In one embodiment, in these configurations, the determination of whether the cell stores a “1” or a “0” is not based on resultant voltage on the RBL but based on currents flowing through the RBL. A dot product here refers to a vector-based dot product summation of products of two vector components. For example, a vector v, having components v₀, v₁, v₂, v₃ . . . can be applied to a vector w having components w₀, w₁, w₂, w₃, . . . in a dot product operation resulting in Y which is a scalar with a value of v₀·w₀+v₁·w₁+v₂·w₂+v₃·w₃+ . . . . In a matrix, each w can be held in a row of cells, and v_i can be applied to each row, with the currents from each of the rows summed to obtain the resultant. For example, if w₀ is a 4-bit word, then four cells in the same row can be used to store w₀. More on the dot product will be provided later in the present disclosure.

It should be appreciated that the in-memory dot product operation is a particular implementation of the 8T SRAM, in which the source lines of the access transistor T1 is coupled to v_i or V_bias (as described below) depending on which of the below-described two configurations is chosen and the current through RBL is sensed. However, for a typical memory operation, the source lines are coupled to ground and the voltage at RBL is sensed as described above to read the cell values.

Two configurations—Config-A and Config-B are provided, according to the present disclosure, to convert the 8T SRAM from a voltage-based single memory readout cell to a current-based cell multi-bit dot product engine. Config-A is shown in FIG. 7 which shows a schematic of the first proposed configuration (Config-A) for implementing the dot product engine using the 8T-SRAM bit-cell. Comparing FIGS. 7 and 6, T1 and T2 in FIG. 7 correspond to the transistors 422 and 420, respectively, in FIG. 6. In the schematic of Config-A shown in FIG. 7, the source of transistor T1 is coupled to the input analog voltage v_i, and the RWL is activated. The current I_RBL which runs through the RBL line is sensed which is proportional to the dot product v_i·g_i, where g_i as discussed above is the ON/OFF conductance of the transistors T1 and T2. While a sense resistor is shown in FIG. 7 as part of a sense circuit (or can be considered as the effective input resistance of the sensing-circuit), a number of current sensing approaches, known to a person having ordinary skill in the art are available. The second configuration (Config-B) is shown in FIG. 8, wherein the input analog voltage is applied to the gate of T2, while the source of transistor T2 is supplied with a constant voltage V_bias, thereby generating I_RBL. which is sensed in the same way as in Config-A.

In Config-A (FIG. 7), a static current flow from the v_i to the RBL, exists which is controlled by the value of the bit held in the bit cell of the SRAM (i.e., a ‘1’ or a ‘0’), where when the bit cell is ‘1’, I_RBL is proportional to the input v_i and the conductance of the two transistors T1 and T2. If the bit-cell (see FIG. 6) stores ‘0’, the transistor T1 is OFF, and the output current through the RBL is close to 0. Whereas if the bit-cell stores a ‘1’, the current is proportional to v_i·g_ON, where g_ON is the series ‘ON’ conductance of the transistors T1 and T2. When placing the circuit in an array of cells, similar inputs v_i are applied on the source lines of transistors T1 for each row of the memory array. Since the RBL is common throughout the column, the currents from all the inputs v_i are summed into the RBL. Moreover, since the source line is common throughout each row, the same inputs v_i are supplied to multiple columns. Thus, the final output current through RBL of each column is proportional to:

I ^j _RBL=Σ(v_i·g^j_i)  (1)

where g^j_i is the ‘ON’ or ‘OFF’ conductance of the transistors, depending on whether the bit-cell in the i-th row and j-th column stores a ‘1’ or ‘0’, respectively. The output current vector thus resembles the vector-matrix dot product, where the vector is v_i in the form of input analog voltages and g^j_i, where g^j_i stored as digital data in the SRAM.

To provide a clearer picture of the dot product in a matrix of cells, refer to FIG. 9a where a schematic of a 4-bit matrix is shown (note the write transistors and the transistors associated with 6T SRAM are not shown), based on Config-A (see FIG. 7). In the above-equation (Eq. 1), if conductance is replaced with weight (w^j_is corresponding to weights of the i^th row and j^th column) these weights can be written as W^ji=w^j₃w^j₂w^j₁w^j₀, resulting in the vector matrix dot product: Σ(v_i·W^j_i)=[v_i·(2³w₃+2²w₂+2¹w₁+2⁰w₀)] which translates to Σ(v_i·2³w₃)+Σ(v_i·2²w₂)+Σ(v_i·2¹w₁)+Σ(v_i·2⁰w₀) for column j. If the read transistors T1 and T2 of the SRAM bit-cells in column 1 through 4 are sized in the ratio 2³: 2²: 2¹: 2⁰, as shown in FIG. 9a, the conductance of transistors in the ‘ON’ state would also be in the ratio 2³: 2²: 2¹: 2⁰ associated with the columns.

Thus, summing the currents through the RBLs of the four columns yields the required dot product in accordance to the equation shown above. An example is now provided to better illustrate the dot product operation. Suppose, a first 4-bit vector w₁ held at a first row has a value of 1001 and its dot product with a vector v₁ is desired. Suppose further, a second 4-bit vector w₁ held at a second row has a value of 0110 and it's dot product with v₁ is also desired. From the above relationship we know that v₁ is applied to both the first row and the second row, however, in each row the overall resistance from T1 and T2 (see FIGS. 7 and 9a) changes from R/8 to R (R represents the overall resistance based on conductance of T1 and T2) where R is the total resistance of T1 and T2 of the least significant bit. Therefore, application of v₁ to each of the first and second rows (holding 1001 and 0110, respectively) results in a current of 1×v₁/(R/2³)+0×v₁/(R/2²)+0×v₁/(R/2¹)+1×v₁/(R/2⁰) for the first row and 0×v₁/(R/2³)+1×v₁/(R/2²)+1×v₁/(R/2¹)+0×v₁/(R/2⁰) for the second row, where for all rows, 8 (2³) is the weight of the most significant bit, . . . and 1 (2⁰) is the weight of the least significant bit. Rewriting these, the current from each row is 8v₁/R+1v₁/R (or 9v₁/R) for the first row and 4v₁/R+2v₁/R (or 6v₁/R) for the second row. Adding these will result in 15v₁/R which is in agreement with the dot product of v₁ with 1001 and 0110 (i.e., v₁·1001+v₁·0100) with unit R for the total resistance of T1 and T2.

This sizing pattern can be repeated throughout the array. In addition, one could also use transistors having different threshold voltages to mimic the required ratio of conductances as 2³:2²:2¹:2⁰. Thus, affecting conductance or threshold voltages can be used to make the needed ratios. It should be noted that the currents through the RBLs of the four consecutive columns are summed together, thus we obtain one analog output current value for every group of four columns. In other words, the digital 4-bit word stored in the SRAM array is multiplied by the input voltage v_i and summed up by analog addition of the currents on the RBLs. This one-go computation of vector multiplication and summation in a digital memory array would result in high throughput computations of the dot products.

Referring to FIG. 9b, the schematic of FIG. 9a is shown for CONIG-B, where the sources of T1 transistors is coupled to a reference voltage (V_bias) that can be a global reference voltage and the gates of transistors T2 is coupled to V_in. I_OUT represents an analog output of the vector-based dot product according to the present disclosure for Config-B.

It should also be noted that the static current flowing through the ON transistors T1 and T2 of the present disclosure would typically be much higher as compared to memristor topology known to a person having ordinary skills in the art. In order to reduce the static current flow, the supply voltage of the SRAM cell can be scaled down. In this regard, it should be noted that in 8T cells these are known to retain their robust operation even at highly scaled supply voltages.

In the yet another approach to reduce the static current, a VDD lower than a nominal VDD of 1V is used as discussed with respect to CONFIG-B (see FIG. 8). With reference to FIG. 8, the source of each T1 are coupled to a constant voltage V_bias. The input vector v_i is coupled to RWLs, i.e., the gate of T2. Similar to Config-A (see FIG. 7), the output current I_RBL is proportional to v_i. We will later show from our simulations that for a certain range of input voltage values, a linear relationship between I_RBL and v_i is obtained, which can be exploited to calculate the approximate dot product. To implement multi-bit precision, the transistor sizing is accomplished in the same way as Config-A (see FIG. 9a), so that the I_RBL is directly proportional to the transistor conductances. Key features of the proposed Config-B are as follows. V_bias can be adjusted to reduce the current flowing through the RBLs. The input voltages v_i have a capacitive load, as opposed to a resistive load in Config-A (see FIG. 7). This relaxes the constraints on the input voltage generator circuitry, and is useful while cascading two or more stages of the dot product engine. However, as discussed below, Config-B (see FIG. 8) has a small non-zero current corresponding to zero input as compared to Config. A (see FIG. 7) that has substantially zero current for zero input.

In order to sense the output current at the RBLs, a current to voltage converter can be used. This can most simply be a resistor, as shown in FIGS. 7 and 8. However, there are a few constraints. As the output current increases, the voltage drop across the output resistor increases, which in turn changes the desired current output. A change in the voltage on the RBL would also change the voltage across the transistors T1 and T2, thereby making their conductance a function of the voltage on the RBL. Thus, at higher currents corresponding to multiple rows of the memory array, the I_RBL does not approximate the vector-matrix dot product, but deviates from the ideal output. This dependence of the RBL voltage on the current I_RBL is further discussed below with different approaches.

The operation of the proposed configurations (Config-A (see FIG. 7) and Config-B (see FIG. 8)) for implementing a multi-bit dot product engine was simulated using HSPICE on the 45 nm PTM technology. For the entire analysis, a scaled down VDD of 0.65V for the SRAM cells was chosen. The main components of the dot-product engine implementation are the input voltages and conductances of the transistors for different states of the cells. A summary of the analysis for the two configurations is presented in FIG. 10a, 10b, and 10c for Config-A (see also FIG. 7), and FIGS. 11a, 11b, and 11c for Config-B (see also FIG. 8). FIGS. 10a and 11a show I_RBL in μA vs. V_in in volts for different weights (1111, 1010, and 0100). FIGS. 10b and 11b show I_RBL in μA vs. weight levels (from 0 to 15) for V_in of 0.05, 0.1, and 0.15V for Config-A (see also FIG. 7), and for V_in of 0.6, 0.55, and 0.5V for Config-B (see also FIG. 8), showing a linear relationship between I_RBL and weight levels at various voltage levels. FIGS. 10c and 11c show I_RBL in mA vs. number of rows from 1 to 65 for V_in of 0.1 V and weight of 1111 for Config-A (see also FIG. 7) and for V_in of 0.55 V and weight of 1111 for Config-B (see also FIG. 8), showing a significant deviation from ideal output (I_N=N×I₁ with increasing number of rows for both, where I₁ is the current corresponding to one row and N is the number of rows). In all these figures, a sensing resistance of 50-ohms connected to the RBL is used, however, as discussed above, other techniques besides a sense resistor (e.g., an op-amp with a comparator) can also be used to sense current. It should be noted that a small sense resistance is required to ensure that the voltage across the sensing resistance is not high enough to drastically alter the conductances of the connected transistors T1 and T2.

In FIGS. 10a and 11a, results presented are for a single 4-bit cell. To preserve the accuracy of a dot-product operation, it is advantageous to operate the cell in the voltage ranges such that the current is a linear function of the applied voltage v_i. These voltage ranges are marked as linear region in FIGS. 10a and 11a. The slope of the linear section I_RBL versus V_in plot varies with weight, thus signifying a dot product operation. Further, at the left voltage extremity of the linear region, I_RBL tends to zero irrespective of the weight, thus satisfying the constraint that the output current is zero for zero V_in. It is to be noted that the two configurations show significantly different characteristics due to the different point-of-application of input voltages.

In FIGS. 10b and 11b, results from the dependence of the current I_RBL on the 4-bit weight levels for Config-A at constant voltages V_in=0.05V, 0.1V, 0.15V and configuration B at V_in=0.5V, 0.55V, 0.6V, respectively, are provided. Different voltages were chosen so as to ensure the circuit operates in the linear region as depicted by FIGS. 10a and 11a. Desirably, I_RBL shows a linear dependence on weight levels and tends to zero out for weight=‘0000’. The choice of any voltage in the linear regions of FIGS. 10a and 11a does not alter the linear dependence of the I_RBL on weight levels.

To expand the dot-product functionality to multiple rows, an analysis for up to 64 rows in the SRAM array was also performed, driven by 64 input voltages. In the worst case condition, when the 4-bit weight stores ‘1111’, maximum current flows through the RBLs, thereby increasing the voltage drop across the output resistance. In FIGS. 10c and 11c, the total current I_RBL indicates to deviate from its ideal value with increasing number of rows, in the worst case condition. The deviation in these figures (FIGS. 10c and 11c) is because the output current is sensed with an equivalent sensing resistance (R_sense) and hence the final voltage on the bit-line (V_BL) is dependent on the current I_RBL. At the same time, I_RBL is also dependent on V_BL and as a result the effective conductance of the cell varies as V_BL changes as a function of the number of rows. It was also observed that the deviation reduces with decreasing sensing resistance as expected. The total summed up current reaches almost 6 mA for 64 rows for the worst case condition (all the weights are ‘1111’).

There are several ways to circumvent the deviation from ideal behavior with increasing number of simultaneous row accesses and also reduce the maximum current flowing through the RBLs. One possibility is to use an operational amplifier (Opamp) at the end of each 4-bit column, where the negative differential input of the Opamp is fed by the bit-line corresponding to a particular column. FIGS. 12 and 13 show schematics of these Opamps for Config-A and Config-B, respectively. The positive input of the Opamps are supplemented by a combination of the Opamp offset voltage and any desired voltage required for suitable operation of the dot-product as shown in FIGS. 12 and 13. Opamp provides capability of sensing the summed up current at the RBL while maintaining a constant voltage at the RBL

The same analysis as discussed above for the case of no Opamp (see FIGS. 10a-10c and 11a-11c) for Config-A and Config-B was repeated but this time with the bitline terminated by associated Opamps. For this analysis, V_pos was set to 0.1 V for the positive input of the Opamp and thus analysis is limited to input voltages above V_pos to maintain the unidirectional current. It should be noted that ideal Opamps were considered for these simulations, where the voltage V_pos can be accounted for both the non-ideal offset voltage of the Opamp and a combination of an externally supplied voltage. In FIGS. 14a and 15a, plots of I_RBL versus input voltage V_in for the two configurations, are provided. Similar behavior is seen as in the case of no Opamp (see FIGS. 10a and 11a). However, it should be noted that the current ranges have decreased since RBL is now clamped at V_pos. Further, the dot-product operation is only valid for V_in>V_pos and thus the acceptable input range is shifted in the presence of an Opamp. FIGS. 14b and 15b show the behavior of I_RBL versus weight levels for the two configurations and desirably, linearity is preserved.

FIGS. 14c and 15c show the current through the RBL as a function of the number of rows. As expected, due to the high input impedance of the Opamp, and the clamping of V_BL at voltage V_pos the deviation of the summed up current from the ideal value have been mitigated to a large extent. Although, the current levels have reduced significantly as compared to the FIG. 10c and 11c, the resultant current for 64 rows would still be higher than the electro-migration limit for the metal lines constituting the RBL. According to one embodiment, access is made sequentially to a smaller section of the crossbar (e.g., 16 rows at a time), convert the analog current into its digital counterpart each time and finally add all accumulated digital results. In addition use of high threshold transistors for the read port of the SRAM would also help to reduce the maximum current values. Further, the maximum current is obtained only when all the weights are ‘1111’, which is usually not true due to the sparsity of matrices involved in various applications.

Those having ordinary skill in the art will recognize that numerous modifications can be made to the specific implementations described above. The implementations should not be limited to the particular limitations described. Other implementations may be possible.

Claims

1. A multi-bit dot product engine-based memory for carrying out vector-based dot products, comprising: a matrix of memory cells having M rows (M-1, M-2, . . . 0) and N columns (N-1, N-2, . . . 0), each memory cell (C) in each of the rows ((CN-1, CN-2, . . . , C0)M-1, (CN-1, CN-2, . . . , C0)M-2, . . . (CN-1, CN-2, . . . , C0)0) holding a value and having dedicated read transistors T1 and T2, where T1 is controlled by the value held in the associated memory cell and T2 is controlled by a row-dedicated read word line RWLi (RWLM-1, RWLM-2, . . . RWL0), the combination of the T1 and T2 transistors for each cell selectively (i) couple a row-dedicated source (vin) for each row ((vin)M-1, (vin)M-2, . . . (vin)0) with a column-dedicated read bit line (RBL) for each column (RBLN-1, RBLN-2, . . . RBL0) for an in-memory vector-based dot product operation or (ii) couple ground with the column-dedicated read bit line (RBL) for each column (RBLN-1, RBLN-2, . . . RBL0) for a memory read operation, where total resistance of the T1 and T2 transistors (R) for each cell (RN-1, RN-2, . . . , R0) in each row (M-1, M-2, . . . 0) is based on Rmax/2(M-1), Rmax/2(M-2), . . . Rmax, where Rmax is the resistance of the least significant cell in each row and where current in each RBL (IRBL) for each column ((IRBL)N-1, (IRBL)N-2, . . . (IRBL)0) is summed as Iout; anda sensing circuit coupled to the matrix of memory cells and adapted to selectively (i) receive and convert the current Iout to a digital value representing an output of the vector-based dot product, or (ii) sense voltage at each of the RBLs (RBLN-1, RBLN-2, . . . RBL0) to read value of the corresponding cells.

2. The multi-bit dot product engine-based memory of claim 1, wherein the sensing circuit is based on a sense resistor generating an analog voltage and an analog-to-digital converter generating a digital representation of the analog voltage.

3. The multi-bit dot product engine-based memory of claim 1, wherein the sensing circuit includes an operational amplifier (op-amp) generating an analog voltage and an analog-to-digital converter generating a digital representation of the analog voltage.

4. The multi-bit dot product engine-based memory of claim 3, wherein the op-amp receives the analog current Iout at its negative terminal, a reference voltage (Vpos) at its positive terminal and a feedback resistor Rf between its output and the negative terminal, wherein the op-amp's output is Vout=−Iout·Rf+Vpos.

5. The multi-bit dot product engine-based memory of claim 1, wherein R is adjusted for each cell based on adjusting conductances of corresponding T1 and T2.

6. The multi-bit dot product engine-based memory of claim 5, wherein the conductances are adjusted for each cell based on adjusting threshold voltages of corresponding T1 and T2.

7. The multi-bit dot product engine-based memory of claim 5, wherein the conductances are adjusted for each cell based on adjusting sizes of corresponding T1 and T2.

8. A multi-bit dot product engine-based memory for carrying out vector-based dot products, comprising: a matrix of memory cells having M rows (M-1, M-2, . . . 0) and N columns (N-1, N-2, . . . 0), each memory cell (C) in each of the rows ((CN-1, CN-2, . . . , C0)M-1, (CN-1, CN-2, . . . , C0)M-2, . . . (CN-1, CN-2, . . . , C0)0) holding a value and having dedicated read transistors T1 and T2, where T1 is controlled by the value held in the associated memory cell and T2 is controlled by a row-dedicated source (vin) for each row ((vin)M-1, (vin)M-2, . . . (vin)0), the combination of the T1 and T2 transistors for each cell selectively (i) couple a reference voltage (Vbias) with a column-dedicated read bit line (RBL) for each column (RBLN-1, RBLN-2, . . . RBL0) for an in-memory vector-based dot product operation or (ii) couple ground with the column-dedicated read bit line (RBL) for each column (RBLN-1, RBLN-2, . . . RBL0) for a memory read operation, where total resistance of the T1 and T2 transistors (R) for each cell (RN-1, RN-2, . . . , R0) in each row (M-1, M-2, . . . 0) is based on Rmax/2(M-1), Rmax/2(M-2), . . . Rmax, where Rmax is the resistance of the least significant cell in each row and where current in each RBL (IRBL) for each column ((IRBL)N-1, (IRBL)N-2, . . . (IRBL)0) is summed as Iout; anda sensing circuit coupled to the matrix of memory cells and adapted to selectively (i) receive and convert the current Iout to a digital value representing an output of the vector-based dot product, or (ii) sense voltage at each of the RBLs (RBLN-1, RBLN-2, . . . RBL0) to read value of the corresponding cells.

9. The multi-bit dot product engine-based memory of claim 8, wherein the sensing circuit is based on a sense resistor generating an analog voltage and an analog-to-digital converter generating a digital representation of the analog voltage.

10. The multi-bit dot product engine-based memory of claim 8, wherein the sensing circuit includes an operational amplifier (op-amp) generating an analog voltage and an analog-to-digital converter generating a digital representation of the analog voltage.

11. The multi-bit dot product engine-based memory of claim 10, wherein the op-amp receives the analog current Iout at its negative terminal, a reference voltage (Vpos) at its positive terminal and a feedback resistor Rf between its output and the negative terminal, wherein the op-amp's output is Vout=−Iout·Rf+Vpos.

12. The multi-bit dot product engine-based memory of claim 8, wherein R is adjusted for each cell based on adjusting conductances of corresponding T1 and T2.

13. The multi-bit dot product engine-based memory of claim 12, wherein the conductances are adjusted for each cell based on adjusting threshold voltages of corresponding T1 and T2.

14. The multi-bit dot product engine-based memory of claim 12, wherein the conductances are adjusted for each cell based on adjusting sizes of corresponding T1 and T2.

15. A method of obtaining an in-memory vector-based dot product, comprising: providing a matrix of memory cells having M rows (M-1, M-2, . . . 0) and N columns (N-1, N-2, . . . 0), each memory cell (C) in each of the rows ((CN-1, CN-2, . . . , C0)M-1, (CN-1, CN-2, . . . , C0)M-2, . . . (CN-1, CN-2, . . . , C0)0) holding a value and having dedicated read transistors T1 and T2, where T1 is controlled by the value held in the associated memory cell and T2 is controlled by a row-dedicated source (vin) for each row ((vin)M-1, (vin)M-2, . . . (vin)0), the combination of the T1 and T2 transistors for each cell selectively (i) couple a reference voltage (Vbias) with a column-dedicated read bit line (RBL) for each column (RBLN-1, RBLN-2, . . . RBL0) for an in-memory vector-based dot product operation or (ii) couple ground with the column-dedicated read bit line (RBL) for each column (RBLN-1, RBLN-2, . . . RBL0) for a memory read operation, where total resistance of the T1 and T2 transistors (R) for each cell (RN-1, RN-2, . . . , R0) in each row (M-1, M-2, . . . 0) is based on Rmax/2(M-1), Rmax/2(M-2), . . . Rmax, where Rmax is the resistance of the least significant cell in each row and where current in each RBL (IRBL) for each column ((IRBL)N-1, (IRBL)N-2, . . . (IRBL)0) is summed as Iout; andselectively (i) receiving and converting the current Iout to a digital value representing an output of the vector-based dot product, or (ii) sensing voltage at each of the RBLs (RBLN-1, RBLN-2, . . . RBL0) to read value of the corresponding cells.

16. The method of claim 15, wherein the sensing circuit is based on a sense resistor generating an analog voltage and an analog-to-digital converter generating a digital representation of the analog voltage.

17. The method of claim 15, wherein the sensing circuit includes an operational amplifier (op-amp) generating an analog voltage and an analog-to-digital converter generating a digital representation of the analog voltage.

18. The method of claim 17, wherein the op-amp receives the analog current Iout at its negative terminal, a reference voltage (Vpos) at its positive terminal and a feedback resistor Rf between its output and the negative terminal, wherein the op-amp's output is Vout=−Iout·Rf+Vpos.

19. The method of claim 15, wherein R is adjusted for each cell based on adjusting conductances of corresponding T1 and T2.

20. The method of claim 19, wherein the conductances are adjusted for each cell based on adjusting one or more of threshold voltages and sizes of corresponding T1 and T2.