Systems and methods for performing in-flight computations
AssigneeMaxim Integrated Products, Inc.
InventorsMark Alan Lovell, Robert Michael Muchsel
Life of this patent
Estimating the term…
Abstract
In-flight operations in an inbound data path from a source memory to a convolution hardware circuit increase computational throughput when performing convolution calculations, such as pooling and element-wise operations. Various operations may be performed in-line within an outbound data path to a target memory. Advantageously, this drastically reduces extraneous memory access and associated read-write operations, thereby, significantly reducing overall power consumption in a computing system.
Claims20
A method for reducing power consumption in machine learning hardware accelerators, the method comprising:
- retrieving input data from a source memory;
- using a compute cache to perform one or more arithmetic operations on the input data to obtain a result;
- and
- transferring the result to a hardware accelerator that performs a convolution operation and generates an output without generating, storing, accessing, or retrieving intermediate data, thereby, reducing at least one of read operations or write operations, wherein the hardware accelerator is a separate circuit from the compute cache.
A system for reducing power consumption in machine learning hardware accelerators, the system comprising:
- a source memory configured to store input data;
- a compute cache coupled to the source memory, the compute cache retrieves the input data from the source memory and performs one or more arithmetic operations on the input data to obtain a result;
- and
- a hardware accelerator that is a separate circuit from the compute cache, the hardware accelerator being coupled to the compute cache and to a data memory, in response to receiving the result, performs steps comprising:
- performing a convolution operation;
- based on the convolution operation, generating an output without generating, storing, accessing, or retrieving intermediate data to reduce at least one of read operations or write operations;
- and
- writing the output to the data memory.
Dependencies are inferred from the claim text.