Back to results
US 2025/0094365 A1Published application
Kernel mapping to nodes in compute fabric
AssigneeMicron Technology, Inc.
InventorsGongyu Wang, Jason Eckhardt
Life of this patent
Estimating the term…
Abstract
A reconfigurable compute fabric can include multiple nodes, and each node can include multiple tiles with respective processing and storage elements. Compute kernels can be parsed into directed graphs and mapped to particular node or tile resources for execution. In an example, a branch-and-bound search algorithm can be used to perform the mapping. The algorithm can use a cost function to evaluate the resources based on capability, occupancy, or power consumption of the various node or tile resources.
Claims20
An apparatus comprising:
- a plurality of memory-compute tiles of a compute-near-memory system, wherein each tile in the compute-near-memory system is coupled to one or more adjacent tiles using a synchronous compute fabric and is coupled to one or more non-adjacent tiles using an asynchronous compute fabric, and wherein each tile comprises a respective functional unit and a memory circuit;
- and
- a non-transitory processor-readable storage medium, the processor-readable storage medium including instructions that, when executed by a processor, cause the processor to:
- receive an acyclic graph that represents multiple operations of a compute kernel for execution using at least a portion of the compute-near-memory system;
- perform a resource search to identify candidate tiles in the compute-near-memory system to perform respective operations of the multiple operations using their respective functional units;
- and
- assign the multiple operations to respective tiles of the identified candidate tiles based on results of the resource search and on timing information from the acyclic graph.
4 more elements
A compute near memory system comprising:
- a plurality of memory-compute nodes, wherein each of the nodes comprises tiles that are coupled using an asynchronous compute fabric and a synchronous compute fabric, wherein the tiles comprise respective functional units that are configured to perform operations in parallel to provide a final compute result based on a compute kernel;
- and
- a controller configured to:
- receive the compute kernel;
- parse the compute kernel to provide an acyclic graph that represents multiple operations;
- perform a resource search to identify candidate tiles in the system to perform respective operations of the multiple operations using their respective functional units;
- and
- assign the multiple operations to respective tiles of the identified candidate tiles based on results of the resource search and based on timing information from the acyclic graph.
5 more elements
An apparatus comprising:
- a plurality of memory-compute tiles of a compute-near-memory system, wherein each tile in the compute-near-memory system is coupled to one or more adjacent tiles using a first compute fabric and is coupled to one or more non-adjacent tiles using a second compute fabric, and wherein each tile comprises a respective functional unit and a memory circuit;
- and
- a non-transitory processor-readable storage medium, the processor-readable storage medium including instructions that, when executed by a processor, cause the processor to:
- receive an acyclic graph that represents multiple operations of a compute kernel for execution using at least a portion of the compute-near-memory system;
- perform a resource search to identify candidate tiles in the compute-near-memory system to perform respective operations of the multiple operations using their respective functional units;
- and
- assign the multiple operations to respective tiles of the identified candidate tiles based on results of the resource search and on timing information from the acyclic graph.
4 more elements
Dependencies are inferred from the claim text.