Massively parallel processing core with plural chains of processing elements and respective smart memory storing select data received from each chain

• US 8,583,896 B2
• Filed: 07/26/2010
• Issued: 11/12/2013
• Est. Priority Date: 11/13/2009
• Status: Active Grant

• Current Assignee: NEC Corporation
• Original Assignee: NEC Laboratories America Inc (NEC Corporation)
• Inventors: Cadambi, Srihari, Majumdar, Abhinandan, Becchi, Michela, Chakradhar, Srimat, Graf, Hans Peter
• Primary Examiner(s): KIM, KENNETH S
• Application Number:
  US12/843,579
• Publication Number:
  US 20110119467A1
• Time in Patent Office: 1,205 Days
• Field of Search: None
• US Class Current: 712/27
• CPC Class Codes: G06F 15/7821 Tightly coupled to memory, e.g. computational memory, smart memory, processor in memory
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

Systems and methods for massively parallel processing on an accelerator that includes a plurality of processing cores. Each processing core includes multiple processing chains configured to perform parallel computations, each of which includes a plurality of interconnected processing elements. The cores further include multiple of smart memory blocks configured to store and process data, each memory block accepting the output of one of the plurality of processing chains. The cores communicate with at least one off-chip memory bank.

Claims

1. An accelerator system, comprising:

a plurality of processing cores, each processing core comprising;

a plurality of processing chains configured to perform parallel computations, each comprising a plurality of interconnected processing elements; and

a plurality of smart memory blocks configured to selectively store data based on an operation, each memory block accepting the output of one of the plurality of processing chains, wherein each smart memory block comprises;

a memory configured to store data elements; and

a filter configured to compare input data elements to a threshold and to store those elements in the memory only if the elements meet the threshold; and

at least one off-chip memory bank connected to each of the processing cores.

2. The accelerator system of claim 1, wherein the smart memory blocks further comprise a scanner configured to scan the memory to determine a new threshold and further configured to update the filter with said new threshold.

3. The accelerator system of claim 1, wherein each processing core is connected to a plurality of off-chip memory banks via a switch, wherein the switch is configured to rotate accesses between the memory banks.

4. The accelerator system of claim 1, wherein the smart memory blocks are configured to issue a stall signal that halts input to the processing chain when a new element is added to the memory.

5. The accelerator system of claim 4, wherein the stall signal is configured to delay storage of data but to allow the chain to continue computation.

6. The accelerator system of claim 4, wherein the stall signal is a global stall signal that halts input to all of the processing chains.

7. The accelerator system of claim 1, wherein the processing elements in each processing chain are arranged with a bi-directional, nearest neighbor interconnect that is configured to propagate inputs in one direction and outputs in the other.

8. The accelerator system of claim 1, wherein the off-chip memory banks share an interconnection bus configured to allow communication between processor cores.

9. The accelerator system of claim 1, wherein the processing cores further comprise an input store that receives input data from a host and passes said input data to the processing chains.

10. The accelerator system of claim 9, wherein each processing chain is connected to a different input store.

11. The accelerator system of claim 9, wherein the processing elements are arranged into a first and a second partition wherein one partition computes while the other partition concurrently communicates with the host.