gpu-unlock
A kernel module that doubles RTX 3090 Ti's BF16 throughput without patching Nvidia's driver
gpu-unlock is a Linux kernel module that raises the FP16/BF16 matrix-multiply throughput of GeForce RTX 3000 cards (GA102/103/104/106) toward their professional-card rate, without patching the NVIDIA driver each time it updates. On a 3090 Ti at 450W, reported dense GEMM throughput goes from 87.5 to 157.4 TFLOPS in BF16; real gains on other cards run from +25% to +81% depending on power limit, and reports say outputs stay bit-identical before and after. It helps matrix-heavy work like LLM prefill, diffusion and FlashAttention, but memory-bandwidth-bound LLM decode and INT8 work see no change.
What holds up
- +Doubles BF16 throughput on a 3090 Ti (87.5 to 157.4 TFLOPS), with bit-identical output reported
- +No driver patching needed after each NVIDIA release, unlike the older UEFI unlock method
- +Builds verify SHA-256 checksums of every pinned input before compiling
Mind the limits
- −Gains depend heavily on power limit — cards capped at 225W saw only +25%, not +80%
- −Doesn't help LLM decode or INT8 work, since those are memory-bandwidth bound, not compute bound
- −Only benefits GA102/103/104/106 chips — other Nvidia generations aren't covered
Picked here? Add the badge
Put it on your site. It links back to this page. Free for every listed product.
<a href="https://stillworks.dev/products/p/gpu-unlock/"><img src="https://stillworks.dev/badge/gpu-unlock.svg" alt="Picked by StillWorks" width="230" height="54"></a>