Nvidia secures tensor conversion patent and publishes distributed neural network methods.
The filing
On November 10, 2026, Nvidia received a US grant for a processor and system that converts between tensor convolution and tensor contraction operations to generate feature maps from image data (USPTO 12,792,902 (opens in a new tab)). Two days later, the company published applications covering distributed computational methods like triangle multiplication and compiler-scheduled triangular update neural networks, alongside multi-resolution hash encoding and distributed attention pair bias (USPTO 20260359009A1, USPTO 20260359389A1 (opens in a new tab), USPTO 20260359632A1 (opens in a new tab), USPTO 20260359633A1 (opens in a new tab), USPTO 20260359641A1 (opens in a new tab), USPTO 20260359726A1 (opens in a new tab), USPTO 20260359924A1 (opens in a new tab)). Concurrently, Mellanox Technologies Ltd. published applications focused on aggregated predictions in multi-modal systems, hierarchical prediction aggregation, and multi-modal prediction attribution (USPTO 20260359626A1 (opens in a new tab), USPTO 20260359658A1 (opens in a new tab), USPTO 20260359659A1 (opens in a new tab)).
This filing’s full text wasn’t read for this issue, so its claims and the applicant’s background aren’t shown. Both are in the record, under Sources.