Reconfigurable AI system

Abstract

A system in package platform includes a processor chip having a runtime processor core, an accelerator core and a processor-memory interface exposed on a chip-to-chip bonding surface, a first memory chip such as 3D NAND flash memory storing a collection of executable models of inference engines, and a second memory chip storing weights of a selected executable model. The second memory chip can comprise a nonvolatile, random access memory, such as phase change memory. Direct vertical connections such as via-to-via connections, are provided between the processor chip and the second memory chip.

Description

BACKGROUND

Field

The present invention relates to computation platforms for performing inference operations using artificial intelligence models, and models generated using machine learning, and more particularity to such platforms suitable for use in edge devices.

Description of Related Art

Systems executing computation models that are developed using machine learning, including artificial intelligence models, involve executing large numbers of arithmetic operations across input arrays using large arrays of coefficients. The coefficients are often referred to as weights. In a platform executing these models, off-chip memory access can be a limiting power and performance issue. Because of the size of the arrays of coefficients used in these models, on-chip memory can be insufficient, particularly in systems in which it is desirable to utilize more than one model.

It is desirable to provide a platform for performing inference operations addressing these issues.

SUMMARY

A reconfigurable inference platform is described suitable for implementation using a system in package “SiP” configuration. A platform as described herein can comprise a processor chip, a first memory chip suitable for storing arrays of weights, and a second memory chip suitable for storing collections of executable models. The platform can be implemented as a multichip module in a single package. The package can be mounted on a circuit board or other type of substrate, and connected to sensors and other components that can generate data consumed by the execution of the models, and consume data generated by execution of the models.

A processor chip in implementations of the platform can include a runtime processor core, an accelerator core and a processor-memory interface exposed on a chip-to-chip bonding surface of the processor chip. A first memory chip in implementations of the platform can include a nonvolatile, high capacity memory, such as 3D NAND flash memory. The first memory chip can store a collection of executable models of inference engines, where each executable model includes a set of weights to be applied in execution of the model, and in some cases a computation graph for the inference engine. A second memory chip can store at least the set of weights of a selected executable model. The second memory chip can comprise a nonvolatile, random access memory, such as phase change memory. The second memory chip can include a memory-processor interface exposed on a surface of the second memory chip, and complementary to the processor-memory interface on the processor chip. Direct vertical connections such as via-to-via connections, between the processor-memory interface and the memory-processor interface are provided, which enable low power, high throughput, and low latency transfer of data between the chips in support of execution of the selected model.

In an example described herein, the processor chip and the second memory chip are stacked and disposed on an interposer. The first memory chip is also disposed on the interposer which includes interconnection wiring forming at least part of a data path between the first memory chip and the second memory chip. The processor chip can include an input/output interface in addition to the processor-memory interface, and the data path can include a connection from the interconnection wiring of the interposer to the input/output interface of the processor chip.

In an example described herein, the processor chip has access to instruction memory, which can be included on the processor chip or accessible in off-chip storage, storing instructions to perform a runtime procedure. The runtime procedure can include selecting an executable model from the collection of executable models stored in the first memory chip, loading a computation graph for the selected model including configuring the accelerator core on the processor chip, transferring the set of weights of the selected model to the second memory chip, and executing the selected model. Also, the runtime procedure can include changing the model in response to a control event in the field. Thus, the runtime procedure can include changing the selected model to a different model in the collection of executable models, loading a computation graph for the different model including configuring the accelerator core, transferring the set of weights of the different model to the second memory chip, and executing the different model.

An example of a reconfigurable inference method is described comprising providing a processor chip including a runtime processor core, an accelerator core, on-chip memory and a processor-memory interface exposed on a chip-to-chip bonding surface of the processor chip, storing a collection of executable models of an inference engine for a model implemented by machine learning in a first memory chip accessible by the processor chip, each model including a set of weights to be applied in execution of the model, selecting in response to a control event an executable model from the collection of executable models stored in the first memory chip, loading a computation graph for the selected model including configuring the accelerator core, and transferring the set of weights of the selected executable model from the first memory chip to a second memory chip, the second memory chip including a memory-processor interface disposed on a surface the second memory chip and complementary to the processor-memory interface, and executing the selected executable model using direct vertical connections between the processor-memory interface and memory-processor interface.

Other aspects and advantages of the present invention can be seen on review of the drawings, the detailed description and the claims, which follow.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is an illustration of a multichip module including an inference platform as described herein.

FIG. 2 is an illustration of another embodiment of a multichip module including an inference platform as described herein.

FIG. 3 is an illustration of yet another embodiment of a multichip module including an inference platform as described herein.

FIG. 4 is a simplified functional block diagram of an inference platform as described herein.

FIG. 5 is a flowchart of a runtime procedure which can be executed by an inference platform as described herein.

DETAILED DESCRIPTION

A detailed description of embodiments of the present technology is provided with reference to the FIGS. 1-5.

FIG. 1 illustrates a reconfigurable inference platform that includes a processor chip 101, a first memory chip 103 (model collection), and the second memory chip 102 (weight memory). In this example, the processor chip 101 and the second memory chip 102 are stacked, and the combination of the processor chip 101 stacked with second memory chip 102, and the first memory chip 103 is mounted on an interposer 110. The assembly is configured as a multichip module 120 in a single package.

The processor chip 101 can include a runtime processor core (e.g. CPU) and an accelerator core, such as an artificial intelligence accelerator (e.g. AIAcc) or a neuron processing unit. The processor chip 101 includes a chip-to-chip bonding surface on which a processor-memory interface 131 is exposed for connection to the second memory chip 102. The second memory chip includes a memory-processor interface 132 exposed on a surface of the second memory chip, and complementary to the processor-memory interface 131 on the processor chip 101. In this example, direct vertical connections are provided between the processor-memory interface and the memory-processor interface. The direct vertical connections can comprise copper via-to-via conductors or other chip-to-chip contact technologies suitable for high speed, low latency, and low power communication between the chips.

In this example, processor chip 101 includes an input/output interface 113 disposed on the surface of the chip 101. The input/output interface 113 is connected to vertical connectors such as through silicon via TSV connections to interconnection wiring 111 on the interposer 110.

The first memory chip 103 includes an interface 112 for connection to the interconnection wiring 111 on the interposer 110.

Thus, interconnection wiring 111 provides part of the data path between the first memory chip and the second memory chip through the processor chip 101.

In the example illustrated in FIG. 1, the processor chip 101 includes another input/output interface 122 for connection to external contact structures 121 of the multichip module 120.

FIG. 2 illustrates another configuration of an inference engine as described herein. This configuration includes a processor chip 201, a first memory chip 203 (model collection), and the second memory chip 202 (weight memory). In this example, the processor chip 201 and the second memory chip 202 are stacked, and the combination of the processor chip 201 stacked with the second memory chip 203 and the first memory chip 202 is mounted on an interposer 210. The assembly is configured as a multichip module 220 in a single package.

The processor chip 201 can include a runtime processor core (e.g. CPU) and an accelerator core, such as an artificial intelligence accelerator (e.g. AIAcc) or a neuron processing unit. The processor chip 201 includes a chip-to-chip bonding surface on which a processor-memory interface 231 is exposed for connection to the second memory chip 202. The second memory chip includes a memory-processor interface 232 exposed on a surface of the second memory chip, and complementary to the processor-memory interface 231 on the processor chip 201. In this example, direct vertical connections at the surfaces are provided between the processor-memory interface and the memory-processor interface. The direct vertical connections can comprise very short length copper via-to-via conductors or other chip-to-chip contact technologies suitable for high speed, low latency, and low power communication between the chips.

In this example, processor chip 201 includes an input/output interface 213 disposed on the surface of the chip 201. The input/output interface 213 is connected to vertical connectors such as through silicon via TSV connections to interconnection wiring 211 on the interposer 210.

Also, the second memory chip 202 includes an input/output interface 241 exposed on the surface opposite the processor chip 201, which connects to complementary interface 240 on the interposer 210, for connection to the interconnection wiring 211 of the interposer 210.

The first memory chip 203 includes an interface 212 for connection to the interconnection wiring 211 on the interposer 210.

Thus, interconnection wiring 211 provides part of the data path between the first memory chip and the second memory chip through the interposer interconnection wiring 211, as an alternative to a data path through the processor chip 201.

In the example illustrated in FIG. 2, the processor chip 201 includes another input/output interface 222 for connection to external contact structures 221 of the multichip module 220.

FIG. 3 illustrates another configuration of an inference engine as described herein. This configuration includes a processor chip 302, a first memory chip 303 (model collection), and the second memory chip 301 (weight memory). In this example, the processor chip 302 and the second memory chip 301 are stacked, and the combination of the processor chip 302 stacked with the second memory chip 301 and the first memory chip 303 is mounted on an interposer 310. In this example, the processor chip 302 is between the second memory chip 301 and the interposer 310. The assembly is configured as a multichip module 320 in a single package.

The processor chip 302 can include a runtime processor core (e.g. CPU) and an accelerator core, such as an artificial intelligence accelerator (e.g. AIAcc) or a neuron processing unit. The processor chip 302 includes a chip-to-chip bonding surface (top surface) on which a processor-memory interface 332 is exposed for connection to the second memory chip 301. The second memory chip 301 includes a memory-processor interface 331 exposed on a surface of the second memory chip (bottom surface), and complementary to the processor-memory interface 332 on the processor chip 302. In this example, direct vertical connections are provided between the processor-memory interface and the memory-processor interface. The direct vertical connections can comprise copper via-to-via conductors or other chip-to-chip contact technologies suitable for high speed, low latency, and low power communication between the chips.

In this example, processor chip 302 includes an input/output interface 313 disposed on the bottom surface of the chip 302. The input/output interface 313 is connected to vertical connectors, which connect to interconnection wiring 311 on the interposer 310.

Also, the processor chip 302 includes an input/output interface 361 exposed on the bottom surface opposite the second memory chip 301, which connects to a complementary interface 362 on the interposer 310, for connection to the interconnection wiring 350 of the interposer 310.

The first memory chip 303 includes an interface 312 for connection to the interconnection wiring 311 on the interposer 310.

Thus, interconnection wiring 311 provides part of the data path between the first memory chip and the second memory chip through the interposer interconnection wiring 311, and through the processor chip 302.

The interposer 310 includes an interface 352 for connection of the interconnection wiring 350 of the interposer (which can be connected to or part of the interconnection wiring 311 of the interposer). Wiring connections are provided from the interface 352 to external contact structures 351 of the multichip module 320.

In other embodiments, the interface 352 can be replaced or supplemented by an interface or interfaces on the side or bottom surfaces of the interposer.

FIGS. 1-3 provide example arrangements of a platform as described herein, showing varieties of configurations of the chips and connections among the chips, the interposer and external contacts of the package. Other arrangements can be implemented as suits a particular need.

FIG. 4 is a simplified functional block diagram of a platform implemented as described with reference to FIGS. 1-3. The platform includes a processor chip 401, a first memory chip 403, and a second memory chip 402. The processor chip 401 in this example includes a CPU or processor core 410, accelerator core 411, on-chip memory 412, such as SRAM which can be used as working memory and as a cache memory, a first I/O interface 413 and a second I/O interface 414. A bus system 420 provides for intra-chip communications among the components.

The first memory chip 403 in this example comprises a high capacity, nonvolatile memory 440 such as 3D NAND implemented using charge trapping storage technology, for example. The first memory chip 403 includes a first memory I/O interface 441 for off-chip communications. The first memory I/O interface 441 can comprise a high-speed serial port, such as an SPI compatible port, or a parallel port, depending on the particular implementation of the memory chip utilized. A data path 451 is provided in this example between the first memory I/O interface 441, and the first I/O interface 413 on the processor chip 401.

The second memory chip 402, in this example, comprises a high-speed, random-access nonvolatile memory 430, such as 3D phase change storage technology. In other examples, the second memory chip 402 can comprise NOR flash memory using charge trapping storage technology, or other suitable random-access technologies like resistive RAM (e.g. metal oxide memory), magnetic RAM, Ferroelectric RAM and so on.

The second memory chip 402 includes a memory I/O interface 431, for off-chip communications directly connected by vertical interconnections 450 to the second I/O interface 414 on the processor chip 401.

DRAM is an option to bond into the SiP in case on-chip SRAM is not big enough. Thermal (heat) management can used to guarantee data retention.

An accelerator core (e.g. accelerator core 411), as the term is used herein, is a configurable logic circuit including components designed or suitable for execution of some or all of the arithmetic operations of an inference model. Configuration of the accelerator core can include loading a set of weights used in the inference model, or parts of the set of weights. In some embodiments, configuration of the accelerator core can include loading some or all of the of the computation graphs of the inference model to define the sequence and architecture of the operation of the inference model. The inference model can comprise a computation graph of a deep learning neural network, in some examples having a plurality of fully connected and partially connected layers, activation functions, normalization functions and so on.

An accelerator core can be implemented using configurable logic, like arrays of configurable units used in field programmable gate arrays for example, in which compiled computation graphs are configured using bit files. An accelerator core can be implemented using a hybrid of data flow configurable logic and sequential processing configurable logic.

The runtime processor core (e.g. CPU 410) can execute a runtime program to coordinate operation of the accelerator core to accomplish real time inference operations, including data input/output operations, loading computation graphs, moving the set of weights to be applied in the inference operation into and out of the accelerator core, delivering input data to the accelerator core, and performing parts of the computation graphs.

FIG. 5 is a flowchart illustrating an example of logic of a procedure executed by an inference platform, such as platforms described with reference to FIGS. 1-4. The logic can be implemented using computer programs stored in memory, such as the SRAM on-chip memory 412, or other memory accessible by the CPU 410. In this example, the procedure includes downloading a collection of executable artificial intelligence models from an external source, such as a network, and loading the collection in the high capacity NAND flash memory on the platform (501). During runtime, the procedure waits for a control event (502). The control event can include a reset, an expiration of a timer, a message received from a communication network or other external source, data generated by execution of an inference engine in the processor chip itself, or other signals. As long as no control event is detected, the procedure loops.

When the control event is detected, the procedure includes selecting an artificial intelligence model from the collection stored in the NAND flash memory (503). The selected model, or at least a set of weights of the selected model, is then transferred from the NAND flash memory to the weight memory (504). The procedure includes configuring the accelerator core using parameters of the selected model read from the NAND flash memory (505). After loading the weights and configuring the accelerator core, the procedure includes executing an inference procedure using the parameters of the selected models stored in the weight memory, including transferring parameters such as weights, using the direct vertical connections between the processor chip 401, and the second memory chip 402 (506).

Thus, the procedure of FIG. 5 includes a procedure to select an executable model from the collection of executable models stored in the first memory chip to load a computation graph for the selected model including configuring the accelerator core, to transfer the set of weights of the selected model to the second memory chip, and to execute the selected model. Also, as shown in FIG. 5, after executing or beginning to execute the selected model, the process loops to step 502, to wait for a next control event. Upon detection of the next control event, the steps 503 to 506 are traversed, and can include changing the selected model to a different model in the collection of executable models, loading a computation graph for the different model including configuring the accelerator core, transferring the set of weights of the different model to the second memory chip, and executing the different model.

It will be appreciated with reference to FIG. 5, that many of the steps can be combined, performed in parallel or performed in a different sequence without affecting the functions achieved. In some cases, as the reader will appreciate, a rearrangement of steps will achieve the same results only if certain other changes are made as well. In other cases, as the reader will appreciate, a rearrangement of steps will achieve the same results only if certain conditions are satisfied. Furthermore, it will be appreciated that the flow charts herein show only steps that are pertinent to an understanding of the invention, and it will be understood that numerous additional steps for accomplishing other functions can be performed before, after and between those shown.

An SiP platform described, in which one or more 3D NAND chip(s) store a collection including multiple different AI models (computation graph and weights), a weight memory chip(s) stores the weights of a selected AI model, and a processor chip which can be a special purpose AI logic chip (CPU+AI accelerator) is included with the memory system to execute the selected AI model parameters (e.g. weights), hyperparameters (e.g. neural network computation graphs or architectural details) needed by the CPU/NPU (e.g. layers, normalization functions, activation functions, etc.)

Inter-chip bonding between the AI logic chip and the weight memory chip can be Via-to-Via Cu bonding or other 3D (2.5 D) bonding technologies.

While the present invention is disclosed by reference to the preferred embodiments and examples detailed above, it is to be understood that these examples are intended in an illustrative rather than in a limiting sense. It is contemplated that modifications and combinations will readily occur to those skilled in the art, which modifications and combinations will be within the spirit of the invention and the scope of the following claims.

Claims

1. A reconfigurable inference platform, comprising: a processor chip including a runtime processor core, an accelerator core, on-chip memory and a processor-memory interface exposed on a chip-to-chip bonding surface of the processor chip;a first memory chip accessible by the processor chip to store a collection of executable models of an inference engine, each model including a set of weights to be applied in execution of the model;a second memory chip to store the set of weights of a selected executable model, the second memory chip including a memory-processor interface exposed on a surface of the second memory chip and complementary to the processor-memory interface; anddirect vertical connections between the processor-memory interface and memory-processor interface,wherein the processor chip and the second memory chip are stacked and disposed on an interposer, and the first memory chip is disposed on the interposer, the interposer including interconnection wiring forming part of a data path between the first memory chip and the second memory chip, andwherein the processor chip includes a second input/output interface, the data path including a connection from the interconnection wiring of the interposer to the second input/output interface on the processor chip.

2. The platform of claim 1, wherein the direct vertical connections comprise via-to-via connections.

3. The platform of claim 1, wherein the processor core has access to instruction memory, storing executable instructions to perform a procedure including: selecting an executable model from the collection of executable models stored in the first memory chip, loading a computation graph for the selected model including configuring the accelerator core, transferring the set of weights of the selected model to the second memory chip, and executing the selected model.

4. The platform of claim 1, wherein the processor core has access to instruction memory, storing executable instructions to perform a procedure in response to a control event, including changing the selected model to a different model in the collection of executable models, load a computation graph for the different model including configuring the accelerator core, transferring the set of weights of the different model to the second memory chip, and executing the different model.

5. The platform of claim 1, wherein the interposer is below the second memory chip, and the processor chip is disposed above the second memory chip.

6. The platform of claim 1, wherein the interposer is below the processor chip and the second memory chip is disposed above the processor chip.

7. The platform of claim 1, wherein the first memory chip comprises a charge trapping, NAND-architecture memory, and the second memory chip comprises nonvolatile random access memory.

8. The platform of claim 7, wherein the nonvolatile random access memory is phase change memory.

9. The platform of claim 7, wherein the nonvolatile random access memory is a charge trapping, NOR-architecture memory.

10. The platform of claim 1, wherein the processor chip, first memory chip and second memory chip are disposed in a multichip package.

11. A reconfigurable inference method, comprising: providing a processor chip including a runtime processor core, an accelerator core, on-chip memory and a processor-memory interface exposed on a chip-to-chip bonding surface of the processor chip;storing a collection of executable models of an inference engine for a model implemented by machine learning in a first memory chip accessible by the processor chip, each model including a set of weights to be applied in execution of the model;selecting in response to a control event an executable model from the collection of executable models stored in the first memory chip, loading a computation graph for the selected model including configuring the accelerator core, and transferring the set of weights of the selected executable model from the first memory chip to a second memory chip, the second memory chip including a memory-processor interface disposed on a surface the second memory chip and complementary to the processor-memory interface; andexecuting the selected executable model using direct vertical connections between the processor-memory interface and memory-processor interface,wherein the processor chip and the second memory chip are stacked and disposed on an interposer, and the first memory chip is disposed on the interposer, the interposer including interconnection wiring forming part of a data path between the first memory chip and the second memory chip, andwherein the processor chip includes a second processor-memory interface, and including transferring data from the first memory chip to the processor chip on a data path including a connection from the interconnection wiring of the interposer to the second processor-memory interface on the processor chip.

12. The method of claim 11, wherein the direct vertical connections comprise via-to-via connections.

13. The method of claim 11, including changing, in response to a second control event, the selected model to a different model in the collection of executable models, loading a computation graph for the different model including configuring the accelerator core, transferring the set of weights of the different model to the second memory chip, and executing the different model.

14. The method of claim 11, wherein the first memory chip comprises a charge trapping, NAND-architecture memory, and the second memory chip comprises nonvolatile random access memory.

15. The method of claim 14, wherein the nonvolatile random access memory is phase change memory.

16. A reconfigurable inference platform, comprising: a processor chip including a runtime processor core, an accelerator core, on-chip memory and a processor-memory interface exposed on a chip-to-chip bonding surface of the processor chip;a first memory chip accessible by the processor chip to store a collection of executable models of an inference engine, each model including a set of weights to be applied in execution of the model;a second memory chip to store the set of weights of a selected executable model, the second memory chip including a memory-processor interface exposed on a surface of the second memory chip and complementary to the processor-memory interface; anddirect vertical connections between the processor-memory interface and memory-processor interface,wherein the processor chip and the second memory chip are stacked and disposed on an interposer, and the first memory chip is disposed on the interposer, the interposer including interconnection wiring forming part of a data path between the first memory chip and the second memory chip, andwherein (i) the interposer is below the second memory chip and the processor chip, and (ii) the processor chip is disposed above the second memory chip or the second memory chip is disposed above the processor chip.