Pull requests / #1336

#1336 cuda: compact traversal for read-only Q8 miss fills

open · @CC-David-CC · 0 comments · View on GitHub

NVIDIA / CUDADocumentation

Description

Default-off read-only Q8 cache foundation and compact miss traversal. Synthetic all-hit fill 10.261 to 2.097 us; one miss 207.597 to 198.592 us. Request speed neutral; exact tokens and work. Fixtures and sanitizer passed. 3.74 GiB extra VRAM.

![Results](https://raw.githubusercontent.com/CC-David-CC/Strata-a5500/71501822fcc08a1cb2226f60dd41903796a2ca65/docs/experiments/compact-q8-main/results.png)

Related on strata.com

Editorial links to help you install, pick models, or read release notes — not part of the upstream thread.