# Linux module targets higher AI computation rates on Nvidia RTX 3000 cards

_Published Saturday, October 24, 2026 at 10:18 PM EDT · Products · Latest · Tier 2 — Notable_

![Unlock FP/BF16 instruction issue rate on Nvidia 3000's cards to their professional equivalents - Zagitta/gpu-unlock — Primary](https://opengraph.githubassets.com/e1a4a48c00e7cbfd2ebc7bd1e1926c77571c0cd93c887dda0c1d90112c113afd/Zagitta/gpu-unlock)

The gpu-unlock project offers a Linux kernel module that its developer says doubles the instruction issue rate for specified floating-point operations on Nvidia GA102, GA103, GA104 and GA106 cards. It changes speed selectors before the unmodified Nvidia driver takes control, avoiding manual patches with each driver release.

User reports cited by the project show LLM prefill gains of 14-36%. Memory bandwidth limits mean LLM decoding stays about the same, as does INT8 work. Reported matrix benchmark gains are not directly comparable because tests and power limits differ.

## Sources

- [github.com](https://github.com/Zagitta/gpu-unlock)

---
Canonical: https://techandbusiness.org/newswire/or_r5b6tf10i73W7NaAaUI
Published: 2026-10-25T02:18:41.327Z
Story chronology: 2026-10-24T09:58:27.000Z
Retrieved: 2026-10-25T06:20:33.005Z
Publisher: Tech & Business (techandbusiness.org)