Heterogeneous ML accelerator cluster with flexible system resource balance
Aspects of the disclosure are directed to a heterogeneous machine learning accelerator system with compute and memory nodes connected by high speed chip-to-chip interconnects. While existing remote/disaggregated memory may require memory expansion via remote processing units, aspects of the disclosure add memory nodes into machine learning accelerator clusters via the chip-to-chip interconnects without needing assistance from remote processing units to achieve higher performance, simpler software stack, and/or lower cost. The memory nodes may support prefetch and intelligent compression to enable the use of low cost memory without performance degradation.
1. A machine learning accelerator system, comprising:
one or more accelerator clusters comprising a plurality of non-uniformly distributed memory nodes and a plurality of compute nodes connected by a high speed chip-to-chip interconnect that facilitates transfer of machine learning models or data, the non-uniformly distributed memory nodes having varying capacities or latencies; and
one or more processors configured to:
keep a portion of the machine learning models or data uncompressed in faster memory nodes of the plurality of memory nodes and a portion of the machine learning models or data compressed in slower memory nodes of the plurality of memory nodes; and
based on a latency threshold for the non-uniformly distributed memory nodes, prefetch the machine learning models or data from the slower memory nodes of the plurality of non-uniformly distributed memory nodes to on-chip memory in the compute nodes via the high speed interconnect.
2. The system of claim 1 , wherein the high speed interconnect comprises an inter-chip interconnect (ICI) having a plurality of off-chip links and a plurality of on-chip links.
3. The system of claim 1 , wherein the plurality of memory nodes are of varying types, comprising high bandwidth memory (HBM) and at least one of double data rate synchronous dynamic random access memory (DDR) or low power DDR (LPDDR).
4. The system of claim 1 , wherein the plurality of compute nodes are of varying types, comprising at least two of tensor processing units (TPUs), graphics processing units (GPUs), or central processing units (CPUs).
5. The system of claim 1 , wherein the one or more processors are further configured to determine memory nodes of the plurality of memory nodes constitute the slower memory nodes from the latency threshold.
6. The system of claim 1 , wherein the one or more processors are further configured to prefetch the machine learning models or data from remote memory nodes of the plurality of memory nodes to the on-chip memory in the compute nodes via the high speed interconnect.
7. The system of claim 6 , wherein the one or more processors are further configured to determine memory nodes of the plurality of memory nodes constitute the remote memory nodes based on locations of the memory nodes within the one or more accelerator clusters.
8. The system of claim 1 , wherein the one or more processors are further configured to swap the machine learning models or data between memory nodes of the plurality of memory nodes based on a hotness of the machine learning models or data.
9. The system of claim 8 , wherein the one or more processors are further configured to determine a hotness of the machine learning models or data based on at least one of recency or frequency of use.
10. The system of claim 1 , wherein the one or more processors are further configured to compress the portion of the machine learning models or data in the slower memory nodes of the plurality of memory nodes.
11. The system of claim 10 , wherein the one or more processors are further configured to:
determine additional machine learning models or data are needed to perform a computation; and
decompress the additional machine learning models or data from the slower memory nodes of the plurality of memory nodes.
12. The system of claim 1 , wherein keeping a portion of the machine learning models or data uncompressed is based on at least one of recency or frequency of use.
13. A method for a machine learning accelerator system, the machine learning accelerator system including one or more accelerator clusters including a plurality of non-uniformly distributed memory nodes and a plurality of compute nodes connected by a high speed chip-to-chip interconnect that facilitates transfer of machine learning models or data, the non-uniformly distributed memory nodes having varying capacities or latencies, the method comprising:
determining, by one or more processors, memory nodes of the plurality of memory nodes constitute slower memory nodes based on a latency threshold for the non-uniformly distributed memory nodes;
keeping, by the one or more processors, a portion of the machine learning models or data uncompressed in faster memory nodes of the plurality of memory nodes and a portion of the machine learning models or data compressed in the slower memory nodes of the plurality of memory nodes; and
prefetching, by the one or more processors, the machine learning models or data from the slower memory nodes to on-chip memory in the compute nodes via the high speed interconnect.
14. The method of claim 13 , further comprising:
determining, by the one or more processors, memory nodes of the plurality of memory nodes constitute remote memory nodes based on locations of the memory nodes within the one or more accelerator clusters; and
prefetching, by the one or more processors, the machine learning models or data from the remote memory nodes to the on-chip memory in the compute nodes via the high speed interconnect.
15. The method of claim 13 , further comprising:
determining, by the one or more processors, a hotness of the machine learning models or data based on at least one of recency or frequency of use; and
swapping, by the one or more processors, the machine learning models or data between memory nodes of the plurality of memory nodes based on the hotness of the machine learning models or data.
16. The method of claim 13 , further comprising:
determining, by the one or more processors, additional machine learning models or data are needed to perform a computation; and
decompressing, by the one or more processors, the additional machine learning models or data from the slower memory nodes of the plurality of memory nodes.
17. A non-transitory computer readable medium for storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations for a machine learning accelerator system, the machine learning accelerator system including one or more accelerator clusters including a plurality of non-uniformly distributed memory nodes and a plurality of compute nodes connected by a high speed chip-to-chip interconnect that facilitates transfer of machine learning models or data, the non-uniformly distributed memory nodes having varying capacities or latencies, the operations comprising:
determining memory nodes of the plurality of memory nodes constitute slower memory nodes based on a latency threshold for the non-uniformly distributed memory nodes;
keeping, by the one or more processors, a portion of the machine learning models or data uncompressed in faster memory nodes of the plurality of memory nodes and a portion of the machine learning models or data compressed in the slower memory nodes of the plurality of memory nodes; and
prefetching the machine learning models or data from the slower memory nodes to on-chip memory in the compute nodes via the high speed interconnect.
18. The non-transitory computer readable medium of claim 17 , wherein the operations further comprise:
determining memory nodes of the plurality of memory nodes constitute remote memory nodes based on locations of the memory nodes within the one or more accelerator clusters; and
prefetching the machine learning models or data from the remote memory nodes to the on-chip memory in the compute nodes via the high speed interconnect.
19. The non-transitory computer readable medium of claim 17 , wherein the operations further comprise:
determining a hotness of the machine learning models or data based on at least one of recency or frequency of use; and
swapping the machine learning models or data between memory nodes of the plurality of memory nodes based on the hotness of the machine learning models or data.
20. The non-transitory computer readable medium of claim 17 , wherein the operations further comprise:
determining additional machine learning models or data are needed to perform a computation; and
decompressing the additional machine learning models or data from the slower memory nodes of the plurality of memory nodes.