Apple patents method to speed up AI calculations across multiple chips
Apple has filed a patent for a way to run large machine learning models across several processors at once. Instead of waiting for complete data from other chips, each processor sends a simplified, compressed version of its work while continuing its own calculations. This allows the system to keep processing without idle time, making the overall operation faster and more efficient.
In the filing’s own words
The subject technology provides for distributed training and inferencing of a machine learning model across multiple processing resources. During a forward pass, each processing resource generates a portion of an update for its corresponding portion of a feature set. While performing this computation, each resource can approximate, compress, and transmit its portion of the features to other resources, and can receive an approximated and compressed portion generated elsewhere. Each resource then decompresses the received representation and updates its local feature portion without waiting for full-precision data from other nodes. By allowing approximation, compression, communication, and computation to occur in parallel, the system may reduce communication overhead that typically causes idle compute periods. This structure may improve sustained utilization and responsiveness when models operate across multiple distributed compute nodes.
What exactly they claim (claim 1)
1. A method comprising: inferencing a machine learning model comprising a set of adjustable weights, wherein the inference comprises multiple iterations of processing input data, each iteration comprising: processing, by a first processing resource, a first portion of the set of adjustable weights and a first portion of a set of features to generate a first portion of an update for the set of features, wherein the first portion of the set of features is derived from a first portion of the input data; while generating the first portion of the update for the set of features: receiving, by the first processing resource and from a second processing resource, an approximated and compressed representation of a second portion of the update for the set of features, wherein the compressed representation of the second portion of the update for the set of features is derived from a second portion of the input data; decompressing, by the first processing resource, the compressed representation of the second portion of the update for the set of features; and updating the first portion of the set of features, by the first processing resource, based at least in part on the first portion of the update for the set of features and the decompressed second portion of the update for the set of features.
Facts
- Company
- Apple (Apple Inc.)
- Official title
- PARALLEL PROCESSING FOR MODEL EXECUTION
- Publication number
- US20260311057A1
- Application number
- 19578867
- Filed
- Mar 25, 2026
- Published
- Oct 8, 2026
- Inventors
- 1
- Claims
- 20
Classified under
G06N 20/00 Physics & computingG06N 5/04 Physics & computingMore from Apple in AI & software
- Apple patents AI that acts on voicemails for you Oct 15, 2026
- Apple patents letting users teach custom hand gestures to their devices Oct 22, 2026
- Apple patents AI that acts on voicemails for you Sep 1, 2026
- Apple patents video tool that swaps focus between subjects after filming Oct 20, 2026
- Apple patents system that changes room ambience based on what you look at Sep 3, 2026
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.