Your old gaming GPU was doing AI math with the handbrake on

One of the most-shared GPU jokes on X this weekend was a greentext from someone who bought a used RTX 3090 in 2023, got told for three years to upgrade, and then watched "some guy drop a github repo." It is funny because it is mostly true.

What happened. Over the past week, two open-source Linux kernel modules appeared on GitHub:

• gpu-unlock (Zagitta, created October 17 PT) targets RTX 30 cards. It runs before the normal NVIDIA driver loads, writes the GPU's tensor "speed selector" registers, then hands the card to the unmodified driver. It is not a VBIOS flash or a patched driver. Its README says it "doubles the instruction issuing speed of FP/BF16 FMA into FP32 operations."
• gpu-unlock-ada (JustAResearcher, created October 23 PT, GPL-2.0) ports the idea to the RTX 4090, with experimental RTX 4070 Ti SUPER support. VideoCardz and TechPowerUp report a similar unlock now also working on the RTX 2080 Ti.

The numbers. The limit only applies to one kind of math: FP16 or BF16 inputs with FP32 accumulation, which a lot of AI training and inference uses.

• RTX 4090, the project's issue-rate test: FP16 with FP32 accumulate 165.6 → 330.8 TFLOPS, BF16 165.5 → 330.3. FP16 with FP16 accumulate was already at about 326 TFLOPS, and INT8 and plain FP32 did not change.
• RTX 3090 Ti, dense 8192³ GEMM at 450 W: BF16 87.5 → 157.4 TFLOPS, FP16 87.5 → 118.4.
• User reports on other RTX 30 cards range from +25% (power-capped 3090s) to about +80% (RTX 3060 12 GB) in BF16 GEMM.

The half rate itself was never a secret. NVIDIA's own Ada whitepaper lists the RTX 4090 at 165.2 TFLOPS for FP16 tensor math with FP32 accumulate, half its FP16-accumulate figure. What is new is that a register write appears to remove it on cards people already own.

The catch. Matrix math is not the same as everything getting twice as fast:

• LLM prefill, the step where a model reads your prompt, sped up 14-36% in the gpu-unlock user reports. Token generation, the part you watch typing out, is limited by memory bandwidth and stayed about the same. Reports that compared outputs found them bit-identical.
• Games are unaffected. No extra frames.
• It is Linux-first (experimental Windows ports exist), needs a kernel boot parameter, and is experimental: unsupported setups may fail to boot, and TechPowerUp notes any modification like this voids the warranty. FP8 on a 4070 Ti SUPER measured 1.0029x, so not every tensor path was limited.
• We did not find a public NVIDIA response.

Our take. For people running local models on a used 24 GB card, this is a real, free speed-up for the prompt-reading half of the job and for image models. It also hands the "just buy a new GPU" crowd an awkward week. If you try it, benchmark your own workload before and after: the 2x is a lab number, and the honest figure for your chatbot is probably smaller.

What we did and didn't do: we read both GitHub projects (README files and repository data), the GPU Unlock benchmarks page, VideoCardz and TechPowerUp coverage, and NVIDIA's Ada whitepaper on October 25, 2026. We have not run the unlock on our own hardware; all performance numbers are the projects' or users' own measurements. Not sponsored; no affiliate links.

MODEL Editors