Back to results
US 10,706,498 B2Granted
Machine learning sparse computation mechanism
AssigneeIntel Corporation
InventorsEriko Nurvitadhi, Balaji Vembu, Tsung-Han Lin, Kamal Sinha, Rajikshore Barik, Nicolas C. Galoppo Von Borries
Life of this patent
Estimating the term…
Abstract
An apparatus to facilitate processing of a sparse matrix is disclosed. The apparatus includes a plurality of processing units each comprising one or more processing elements, including logic to read operands, a multiplication unit to multiply two or more operands and a scheduler to identify operands having a zero value and prevent scheduling of the operands having the zero value at the multiplication unit.
Claims12
An apparatus comprising:
- one or more processors including one or more processing units;
- and
- a buffer for storage of data;
- wherein the one or more processing units are to:
- detect sparse matrices, wherein detecting sparse matrices includes the one or more processing units to determine whether detected sparse matrices are frequently utilized in neural network operation;
- compress one or more detected sparse matrices to generate one or more compressed matrices, including the one or more processing units to compress a first sparse matrix to generate a first compressed matrix;
- store the one or more compressed matrices in the buffer;
- and
- upon receiving an instruction for a neural network operation including the first sparse matrix, obtain the first compressed matrix from the buffer, decompress the first compressed matrix to generate the first sparse matrix, and utilize the decompressed first sparse matrix in the neural network operation.
6 more elements
A graphics processing unit comprising:
- one or more streaming multiprocessors;
- and
- a shared memory for storage of data;
- wherein the one or more streaming multiprocessors are to:
- detect sparse matrices, wherein detecting sparse matrices includes the one or more streaming multiprocessors to determine whether detected sparse matrices are frequently utilized in neural network operation;
- compress one or more detected sparse matrices to generate one or more compressed matrices, including compressing a first sparse matrix to generate a first compressed matrix;
- store the one or more compressed sparse matrices in the shared memory;
- and
- upon receiving an instruction for a neural network operation including the first sparse matrix, obtain the first compressed matrix from the shared memory, decompress the first compressed matrix to generate the first sparse matrix, and utilize the decompressed first sparse matrix in the neural network operation.
6 more elements
A method comprising:
- detecting, by one or more processing units of a processor, one or more sparse matrices, wherein detecting sparse matrices includes determining whether detected sparse matrices are frequently utilized in neural network operation;
- compress the one or more sparse matrices to generate one or more compressed matrices, including compressing a first sparse matrix to generate a first compressed matrix;
- storing the one or more compressed sparse matrices in a buffer;
- and
- upon receiving an instruction for a neural network operation including the first sparse matrix, obtaining the first compressed matrix from the buffer, decompressing the first compressed matrix to generate the first sparse matrix, and utilizing the decompressed first sparse matrix in the neural network operation.
2 more elements
Dependencies are inferred from the claim text.