Scaling out architecture for DRAM-based processing unit (DPU)
A processor includes a plurality of memory units, each of the memory units including a plurality of memory cells, wherein each of the memory units is configurable to operate as memory, as a computation unit, or as a hybrid memory-computation unit.
1. A memory device comprising:
a memory unit comprising a programmable memory cell configured to be reconfigured to:
operate from a memory for performing storage operations to operate as a computation unit based on a computational requirement; and
operate from the computation unit to operate as the memory for performing storage operations based on a storage requirement for an extension of memory.
2. The memory device of claim 1 , wherein the memory unit is further configured to be reconfigured to operate as both the memory and the computation unit as a hybrid memory-computation unit.
3. The memory device of claim 1 , wherein the memory unit is configured to receive a job from a host.
4. The memory device of claim 1 , wherein the memory unit is configured to be controlled by a host configured to perform at least one of job partitioning for, data distribution to, data collection from, or job distribution to, the memory unit.
5. The memory device of claim 1 , wherein the memory unit comprises DRAM.
6. The memory device of claim 1 , wherein the memory unit comprises a plurality of memory units arranged in a scalable cluster architecture.
7. The memory device of claim 6 , wherein a first memory unit from among the memory units is configured to perform a first portion of a job and a second memory unit from among the memory units is configured to perform a second portion of the job, in response to the first and second memory units being unavailable to perform an entirety of the job or being incapable of performing the entirety of the job.
8. The memory device of claim 6 , wherein at least one of the memory units is configured as the computation unit, and at least another one of the memory units is configured as the memory.
9. A distributed computation system comprising:
a first memory unit; and
a second memory unit,
wherein at least one of the first and second memory units comprises a programmable memory cell configured to be reconfigured to:
operate from a memory for performing storage operations to operate as a computation unit based on a computational requirement; and
operate from the computation unit to operate as the memory for performing storage operations based on a storage requirement for an extension of memory.
10. The system of claim 9 , wherein at least one of the first or second memory units is further configured to be reconfigured to operate as both the memory and the computation unit as a hybrid memory-computation unit.
11. The system of claim 9 , wherein the first memory unit is configured to perform at least a portion of a job according to a workflow, and to forward a remainder of the workflow to the second memory unit after completion of the at least the portion of the job.
12. The system of claim 9 , wherein the first and second memory units are configured to communicate with each other using an asynchronous communication protocol.
13. The system of claim 9 , wherein the first memory unit is configured to perform a first job corresponding to a first request received by a host, and to provide a result of the first job to the host.
14. The system of claim 13 , wherein the second memory unit is configured to receive the result of the first job from the host.
15. The system of claim 13 , wherein the second memory unit is configured to receive a second request from the host, and to be reconfigured according to at least one of the storage requirement or the computational requirement corresponding to the second request.
16. The system of claim 9 , wherein the first memory unit is configured as the memory, and the second memory unit is configured as the computation unit.
17. A method of controlling work flow in a distributed computation system, the method comprising:
receiving a work flow comprising a job to be performed by a first memory unit;
performing, by the first memory unit, at least a portion of the job according to the work flow; and
receiving, by a second memory unit, a remainder of the work flow,
wherein at least one of the first and second memory units comprises a programmable memory cell configured to be reconfigured to:
operate from a memory for performing storage operations to operate as a computation unit based on a computational requirement of the workflow; and
operate from the computation unit to operate as the memory for performing storage operations based on a storage requirement for an extension of memory.
18. The method of claim 17 , further comprising:
forwarding, by the first memory unit, the remainder of the workflow to the second memory unit.
19. The method of claim 17 , further comprising:
providing, by the first memory unit, a result of the of the at least the portion of the job to a host; and
receiving, by the second memory unit, the result and the remainder of the workflow from the host.
20. The method of claim 19 , further comprising:
reconfiguring the second memory unit according to at least one of the storage requirement or the computational requirement corresponding to the remainder of the workflow.
21. The memory device of claim 1 , wherein the programmable memory cell is configured to be reconfigured through an interface of a library.