Memory device performing parallel arithmetic processing and memory module including the same

• US 10,642,612 B2
• Filed: 07/13/2018
• Issued: 05/05/2020
• Est. Priority Date: 11/15/2017
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd. (Samsung Group)
• Original Assignee: Samsung Electronics Co. Ltd. (Samsung Group)
• Inventors: Kim, Jin-hyun
• Primary Examiner(s): Chappell, Daniel C.
• Application Number:
  US16/034,688
• Publication Number:
  US 20190146788A1
• Time in Patent Office: 662 Days
• CPC Class Codes: G06F 1/10 Distribution of clock signals , e.g. skew
• CPC Class Codes: G06F 9/3001 Arithmetic instructions
• CPC Class Codes: G06N 3/063 using electronic means
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G11C 11/54 using elements simulating biological cells, e.g. neuron
• CPC Class Codes: G11C 5/04 Supports for storage elements , e.g. memory modules; Mounting or fixing of storage elements on such supports
• CPC Class Codes: G11C 7/1006 Data managing, e.g. manipulating data before writing or reading out, data bus switches or control circuits therefor

Abstract

A memory module includes a first memory device configured to receive data and first information from a hardware accelerator, to generate an arithmetic result by performing arithmetic processing using the data and the first information, and to output the arithmetic result through an interface with at least one other memory device; and a second memory device configured to receive the arithmetic result from the first memory device through the interface without using the hardware accelerator, and to store the arithmetic result.

Claims

1. A memory module comprising:

a first memory device configured to receive data and first information from a hardware accelerator, to generate an arithmetic result by performing arithmetic processing using the data and the first information, and to output the arithmetic result through an interface that communicates with at least one other memory device; and

a second memory device configured to receive the arithmetic result from the first memory device through the interface without using the hardware accelerator, and to store the arithmetic result,

wherein;

the first memory device comprises a first die configured to communicate with a device outside the first memory device, and a plurality of second dies,

each of the second dies comprises a memory cell array and an arithmetic circuit configured to perform the arithmetic processing, and

the first die comprises a first interface circuit configured to exchange the data with the hardware accelerator and a second interface circuit configured to exchange the arithmetic result with the second memory device.

2. The memory module of claim 1, wherein arithmetic processing using data received from the hardware accelerator is performed in parallel in the plurality of second dies in a distributed fashion.

3. The memory module of claim 1, wherein the hardware accelerator comprises at least one selected from a field-programmable gate array (FPGA), a massively parallel processor array (MPPA), a graphics processing unit (GPU), an application-specific integrated circuit (ASIC), a neural processing unit (NPU), a tensor processing unit (TPU), and a multi-processor system-on-chip (MPSoC).

4. The memory module of claim 1, wherein the first interface circuit receives the data in response to one of a rising edge and a falling edge of a clock signal, and

the second interface circuit outputs the arithmetic result in response to a rising edge and a falling edge of a clock signal.

5. The memory module of claim 1, wherein the first memory device comprises a first high bandwidth memory (HBM), the second memory device comprises a second HBM, the first die comprises a buffer die, and the second dies comprise core dies stacked on the buffer die.

6. The memory module of claim 1, wherein the first memory device receives a weight as the first information from the hardware accelerator for neural network computation and generates, as the arithmetic result, a neural network computation result obtained using the data and the weight.

7. The memory module of claim 1, wherein the second memory device receives a command and an address from the hardware accelerator and stores the arithmetic result received from the first memory device at a position indicated by the address.

8. The memory module of claim 1, wherein the first memory device comprises a command/address generator and transmits a command and an address to the second memory device, the command requesting to store the arithmetic result, and the address indicating a position at which the arithmetic result is to be stored, and

the second memory device stores the arithmetic result in response to the command and the address from the first memory device.

9. The memory module of claim 1, further comprising a third memory device configured to communicate with the second memory device,

wherein the third memory device receives the arithmetic result from the second memory device and provides the arithmetic result to the hardware accelerator, under control of the hardware accelerator.

10. The memory module of claim 1, wherein the hardware accelerator is provided inside the memory module.

11. A memory module comprising:

an interposer having wiring for signal transmission formed thereon; and

a plurality of high bandwidth memories (HBMs) mounted on the interposer,

wherein each of the HBMs comprises a plurality of core dies, and a buffer die configured to communicate with a device outside each HBM, each of the core dies comprising a memory cell array and an arithmetic circuit configured to perform arithmetic processing,

a first HBM among the plurality of HBMs is configured to receive data from a memory controller through a buffer die and perform arithmetic processing using the data in a plurality of core dies in parallel, and

the memory module is configured such that an arithmetic result of the first HBM is provided to a second HBM among the plurality of HBMs through data communication between the buffer die of the first HBM and a buffer die of the second HBM.

12. The memory module of claim 11, wherein the memory controller is provided outside the memory module and comprises a hardware accelerator configured to control neural network computation.

13. The memory module of claim 12, wherein the first HBM is further configured to receive a plurality of weights from the hardware accelerator through the buffer die thereof, and each of the core dies of the first HBM is configured to perform the neural network computation using a respective piece of the data and a respective weight.

14. The memory module of claim 13, wherein the plurality of HBMs comprise HBMs arranged in A rows and B columns, where A and B are integers of at least 2,

the first HBM is in the same column as the second HBM,

B HBMs in one row are configured to respectively receive different weights from the memory controller in parallel,

the hardware accelerator is configured such that data from the hardware accelerator is transmitted to the B HBMs sequentially in one direction, and

each of the B HBMs is configured to perform the neural network computation using the data and a corresponding weight among the different weights.

15. The memory module of claim 14, wherein each of the different weights from the hardware accelerator is transmitted to A HBMs sequentially in one column.

16. The memory module of claim 11, wherein the buffer die of the first HBM comprises a converter and is configured to perform serial-parallel conversion on at least one among the data and an arithmetic result, the data being provided to the core dies, and the arithmetic result being received from the core dies.

17. A memory device comprising:

a first die configured to interface with an external memory controller and an external memory device;

a plurality of second dies stacked on the first die, each of the second dies comprising a memory cell array; and

a through-silicon via (TSV) region configured to transmit and receive a signal between the first die and the plurality of second dies,

wherein each of the second dies comprises a plurality of channels, each of the channels comprising an arithmetic circuit configured to perform arithmetic processing using data provided from the external memory controller, and

the first die comprises;

a first interface circuit configured to receive the data and an arithmetic command by communicating with the external memory controller and to transmit the received data to the second dies through the TSV region; and

a second interface circuit configured to receive an arithmetic result from the second dies through the TSV region and to output the arithmetic result by communicating with the external memory device.

18. The memory device of claim 17, wherein the memory device is a high bandwidth memory (HBM), the first die comprises a buffer die, and each of the second dies comprises a core die.

19. The memory device of claim 17, wherein the second interface circuit receives the arithmetic result from the external memory device, and

the arithmetic circuit is configured to perform the arithmetic processing using information provided from the external memory controller and the arithmetic result received through the second interface circuit.