Benchmarks

Every figure below is labelled by where it came from. This site has not re-run these benchmarks — they are the project maintainer's published measurements and the community reports the project links to, gathered so you don't have to dig through issue threads.

Sources checked 2026-10-25 · Upstream README pinned at commit af28cbb (2026-10-24) · Percentages computed from the published before/after figures

The headline result — RTX 3090 Ti, dense GEMM

Project-reported Dense 8192³ GEMM on an RTX 3090 Ti at a 450 W power limit, five-sample medians, FP32 accumulation. Published in the gpu-unlock README and its linked benchmark report.

Input type	Before (locked)	After (unlocked)	Change
FP16	87.46 TFLOPS	118.39 TFLOPS	+35%
BF16	87.50 TFLOPS	157.41 TFLOPS	+80%
TF32	42.58 TFLOPS	82.61 TFLOPS	+94%

Why BF16 gains the most: the lock halves the FP/BF16 FMA issue rate into FP32 operations, so removing it can at most double that rate — BF16's +80% captures most of the theoretical headroom in this benchmark. FP16 improved less in the same run (+35%); the README doesn't break down the difference, and real workloads mix many operation types anyway, so treat the trio as "measured on this benchmark" rather than a formula. And these are one card, one benchmark, one power limit: "what the unlock can do on a 450 W flagship", not a promise for every card.

Per-card results — community reports

Community-reported BF16 dense-GEMM gains from the user reports in the project's tested-cards list. Each report ran its own benchmark, so rows aren't directly comparable — and power limits explain much of the spread.

Card	Chip	Reported BF16 GEMM gain	Note
RTX 3090 Ti	GA102	+68–80%	Includes the 450 W headline run above.
RTX 3090	GA102	+25–67%	Lowest figure from a card capped at 225 W.
RTX 3080 10 GB	GA102	+70–73%	—
RTX 3080 20 GB (mod)	GA102	+56–59%	Memory-upgraded card.
RTX 3070	GA104	+73–81%	—
RTX 3060 12 GB	GA106	~+80%	—
RTX 3080 Ti 20 GB (OEM)	GA102-250	no gain	Unlock completes; BF16 stays at half rate. The known exception.

Real workloads

Community-reported Collected in the project's README from its issue reports:

LLM prefill: +14–36%. Prefill is compute-bound matrix work, so it benefits — though by less than raw GEMM, because a real model run is never one perfect matrix multiply.
LLM decode: no improvement reported. Decode is limited by memory bandwidth. The unlock raises an issue-rate ceiling decode never hits.
Diffusion (ComfyUI): +47–59% on BF16 image models — one user’s test, fully labelled. Reddit user under-_-null measured an RTX 3090 Ti (Dell Precision 5820, 450 W, ComfyUI 0.42.1, batch 1, median of 15 generations per model): BF16 results were +56.1%, +57.2% and +58.6% across three image models, with MiniMax H3 at +60.2%; Notebookcheck summarised the run as roughly +47–59%. The same test, for balance: LTX-2.5 video generation gained +16.8% and an INT8 model +0.3% — the size of the gain depends entirely on how issue-rate-bound the workload is. Single user, single card, not independently verified — treat it as one well-documented data point, not a promise. FlashAttention is grouped with the matrix-heavy workloads in the README, without a canonical percentage.
INT8: no improvement reported. The selectors this module writes govern FP/BF16 rates (the ComfyUI run above saw +0.3% on its INT8 model — within noise).
Outputs: bit-identical in the reports that compared them. The reports that checked results before and after found exactly the same outputs (the ComfyUI run above compared 150 output pairs pixel-by-pixel, all identical) — the unlock changes how fast instructions issue, not what they compute.

How to read these numbers honestly

"Up to ~2×" applies to one microbenchmark class (dense GEMM issue rate), on tested cards, at generous power limits. End-to-end application gains are smaller — LLM prefill's +14–36% is the realistic shape of it. If a workload of yours is memory-bound, expect nothing. Anyone quoting you a single "speedup percentage" for this unlock without naming the workload is selling something.

Gaming performance

There is no verified gaming benchmark data for gpu-unlock as of this writing — no FPS tables in the README or its linked reports. The lock it removes targets FP/BF16 compute issue rates used by AI and HPC workloads, and every published measurement is a compute benchmark. If gaming numbers matter to you, the honest answer today is "unmeasured", and you should treat any FPS claims circulating without a methodology as unverified.

Primary source: github.com/Zagitta/gpu-unlock (README benchmark table and tested-cards list, with per-report links), pinned at commit af28cbb. ComfyUI figures: Reddit user under-_-null’s RTX 3090 Ti test, as reported by Notebookcheck and WccfTech. RTX 40 series figures live on the RTX 40 page — different project, different evidence base.