公开(公告)号：US20230244746A1

公开(公告)日：2023-08-03

申请号：US17674470

申请日：2022-02-17

Applicant: THE REGENTS OF THE UNIVERSITY OF MICHIGAN

Inventor： Dennis SYLVESTER ,  David BLAAUW ,  Yu CHEN ,  Pierre ABILLAMA ,  Hun-Seok KIM

IPC: G06F17/14 ,  G06F17/16 ,  G06N3/08

CPC classification number: G06F17/145 ,  G06F17/16 ,  G06N3/08

Abstract: A computer-implemented method is presented for performing a computation with a neural network. The method includes: receiving a first input patch of data; applying a Walsh-Hadamard transform to the input patch to yield a transformed input patch in a transformed domain; computing an element-wise product of the transformed input patch and a kernel of the neural network; applying an inverse Walsh-Hadamard transform to the element-wise product to yield an intermediate matrix; and creating a first output patch from the intermediate matrix, where the size of the first output patch is smaller than the intermediate matrix.

公开(公告)号：US20240028900A1

公开(公告)日：2024-01-25

申请号：US17872715

申请日：2022-07-25

Applicant: THE REGENTS OF THE UNIVERSITY OF MICHIGAN

Inventor： Hun-Seok KIM ,  David BLAAUW ,  Dennis SYLVESTER ,  Yu CHEN ,  Pierre ABILLAMA ,  Hyochan AN

IPC: G06N3/08 ,  G06F17/16 ,  G06F7/544

CPC classification number: G06N3/082 ,  G06F17/16 ,  G06F7/5443

Abstract: Recent advances in model pruning have enabled sparsity-aware deep neural network accelerators that improve the energy efficiency and performance of inference tasks. SONA, a novel transform-domain neural network accelerator is introduced in which convolution operations are replaced by element-wise multiplications and weights are orthogonally structured to be sparse. SONA employs an output stationary dataflow coupled with an energy-efficient memory organization to reduce the overhead of sparse-orthogonal transform-domain kernels that are concurrently processed while maintaining full multiply-and-accumulate (MAC) array utilization without any conflicts. Weights in SONA are non-uniformly quantized with bit-sparse canonical-signed-digit (BS-CSD) representations to reduce multiplications to simpler additions.