/Patent Tracker
NvidiaGranted patentAI & softwareInterest 6/10

Nvidia patents method to shrink neural networks using codebooks

Patent drawing
Patent drawing from the official filing.

Nvidia describes a technique for reducing the size of neural networks by finding the best way to round their internal numbers. The system tests different scaling settings to match data values to a fixed list of reference points, minimizing the error introduced by this simplification.

Why it mattersThis improves how efficiently AI models can be stored and run, which is useful for anyone using Nvidia hardware for machine learning tasks.

In the filing’s own words

Apparatuses, systems, and techniques for calculating data compression parameters using codebook entry values. In at least one embodiment, one or more circuits is to calculate one or more data compression parameters based, at least in part, on at least on one or more values of the data to be compressed in relation to at least two codebook entry values.

What exactly they claim (claim 1)

1. One or more processors, comprising circuitry to: calculate a plurality of sets of candidate quantization scale parameters and one or more assignment parameters indicating an assignment of one or more data values to one or more codebook entry values based, at least in part, on minimization of one or more errors of quantizing the one or more data values to the one or more codebook entry values; select, from the plurality of sets of candidate quantization scale parameters, a set of quantization scale parameters that minimizes the one or more errors of quantizing the one or more data values; and generate a quantized version of a neural network using the selected set of quantization scale parameters.

Facts

Company
Nvidia (NVIDIA Corporation)
Official title
Calculating data compression parameters
Patent number
US12762751B2
Application number
16952893
Filed
Nov 19, 2020
Granted
Oct 13, 2026
Inventors
5
Claims
33

Classified under

G06T 9/002 Physics & computingG06T 1/20 Physics & computing

More from Nvidia in AI & software

Summaries are simplified from the filing’s own abstract and first claim. A patent filing is not a product announcement.
Patent data and drawings: U.S. Patent and Trademark Office (USPTO). Nebula is not affiliated with or endorsed by the USPTO.