FULLY INDICATIVE THINLY MATRIX REPRESENTATIONS WITH LIMITED METADATA
A fully meaningful sparse matrix representation with limited metadata compresses dense matrices into sparse matrices based on importance, addressing inefficient storage and computation in neural networks by reducing metadata by 33% while maintaining expressiveness.
DE102024132011A1Pending Publication Date: 2026-01-22NVIDIA CORP
Patent Information
- Application Number
- DE102024132011
- Authority / Receiving Office
- DE · DE
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-07-18
- Filing Date
- 2024-11-04
- Publication Date
- 2026-01-22
Smart Images
Figure 00000000_0000_ABST
Abstract
Systems and methods for compressing a dense matrix into a fully meaningful sparse matrix representation with limited metadata are disclosed. The methods include generating a sparse matrix with appropriate metadata based on a dense matrix. Generating the sparse matrix with appropriate metadata involves identifying a first number M of elements to be compressed, a second number N of elements to be retained, and a third number B specifying the number of bits used by each metadata value; determining a metadata value for each of the N elements of the dense matrix; packing a first metadata value with more than B bits into a second metadata value with B bits; and generating the sparse matrix containing the N elements of the dense matrix.The methods involve storing the sparse matrix and the corresponding metadata, with the corresponding metadata comprising the second metadata value.
Need to check novelty before this filing date? Find Prior Art
Citation Information
Patent Citations
Systems and methods for sparse matrix multiplication
WO2023196039A1