OpenAI patents memory chip design to update AI weights while calculating
OpenAI has filed a patent for a specialized computer chip that performs math and updates data at the same time. The design splits memory into two groups, allowing the system to load new AI settings into one part while the other part continues processing data. This setup aims to reduce idle time in heavy AI training and inference tasks.
In the filing’s own words
The disclosure is related to systems, apparatuses, devices, and methods for parallel operations in computing systems that may be useful for heavy AI training and inference workloads. The disclosed technology provides arrangements for efficient computation. In some implementations, a compute-in-memory (CIM) device or macro is configured to perform vector matrix multiplication (VMM) operations between a vector of activation values and a matrix of weight values. A column cell in the CIM macro has bitcells are configured to store weight values and other bitcells are configured to store weight values that are different from the first weight values. The CIM macro is further configured to perform, in parallel, read/write operations to the first set of bitcells and the VMM operation in the second set of bitcells based on a weight select signal.
What exactly they claim (claim 1)
1. A compute apparatus to minimize stalls using parallelization, the compute apparatus comprising: memory cells comprising first cells and second cells, the first cells configured to store first weight values and the second cells configured to store second weight values different from the first weight values, the first weight values and the second weight values being stored to perform vector matrix multiplication (VMM) operations; a first set of lines and a second set of lines configured to address the first cells and the second cells, respectively; and a signal configured to determine access to at least one of the first cells or the second cells, wherein the compute apparatus is configured to load updated weight values into the first cells with the second cells remaining in use for a VMM operation.
Facts
- Company
- OpenAI (OpenAI Opco, LLC)
- Official title
- PARALLEL OPERATIONS IN COMPUTE SYSTEMS
- Publication number
- US20260279480A1
- Application number
- 19674310
- Filed
- May 12, 2026
- Published
- Sep 17, 2026
- Inventors
- 5
- Claims
- 20
Classified under
G11C 29/32 Physics & computingG11C 16/32 Physics & computingG11C 29/1201 Physics & computingG11C 2029/1802 Physics & computingMore from OpenAI in Chips & hardware
- OpenAI patents chip design that learns while processing data Sep 29, 2026
- OpenAI patents analog circuit that learns by balancing electrical energy Sep 29, 2026
- OpenAI patents chip design that switches between math types to save power Oct 13, 2026
- OpenAI patents chip design that flips math operations without moving data Sep 17, 2026
- OpenAI patents device that shakes to get your attention Oct 8, 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.