Firmware exploit lifts Nvidia's half-speed AI throttle on GeForce GPUs

An RTX 4090's peak tensor rate doubles — from 165.6 to 331.6 TFLOPS on FP16 — once a firmware exploit strips Nvidia's imposed half-speed AI throttle, according to modders whose work VideoCardz reports on. The catch for practitioners: real workloads gain far less, and chatbot token speeds barely move.

Nvidia ships GeForce cards across the RTX 20, 30, and 40 generations with FP32-accumulate math capped at half each card's FP16-accumulate rate, per its own published specs. Its Turing and Ampere professional cards don't carry that limit. The unlock returns native matrix-math throughput to consumer owners, which can speed up prompt processing and image generation — but chatbot token generation stays about the same because it depends on memory bandwidth, not raw compute.

Which cards does the unlock cover?

Three separate projects each ship an unofficial Linux module, while experimental Windows builds take a different route. Coverage varies by generation:

• turing-unlock (adapted from gpu-unlock): RTX 2080 Ti with 22GB memory mods; other Turing IDs added but untested. FP16 only — Turing lacks BF16 hardware.
• gpu-unlock: Ampere, listing GA102/GA103/GA104/GA106 IDs. Confirmed on RTX 3090 Ti, 3090, 3080 10GB, a 20GB memory-modded 3080, 3070, and 3060 12GB. The 3070 Ti and GA103 cards remain untested; one RTX 3080 Ti 20GB OEM card stays at half rate.
• gpu-unlock-ada: RTX 4090 confirmed; the RTX 4090 D, 4070 Ti Super, and 4060 Ti 16GB are experimental, with one card each.

Blackwell (RTX 50) isn't supported. gpu-unlock's developer calls it out of scope.

What do the numbers actually look like?

The gpu-unlock-ada README records one independent owner's cuBLAS GEMM benchmark on an RTX 4090: median matrix-multiply improvements of just 12–34% for FP16 and up to 20% for BF16. A separate Windows data point, an MSI RTX 4070 Ti, posted larger gains of +80.1% and +76.4% — power limits may explain the difference.

An RTX 3090 Ti tested by one user with a modified Nvidia driver scored +35% FP16, +80% BF16, and +94% TF32 on matrix multiply. Other Ampere user reports, which gpu-unlock notes aren't directly comparable:

• RTX 3090: +25–67% BF16 (the low end came from cards capped at 225 W)
• RTX 3080 10GB: +70–73%; a memory-modded 20GB variant: +56–59%
• RTX 3070: +73–81%; RTX 3060 12GB: around +80%

For Ampere, LLM prefill improved 14–36%, and diffusion and FlashAttention sped up. LLM decode and INT8 stayed "about the same." Power limit "matters a lot," per the README.

On the Turing side, a Dell OEM RTX 2080 Ti with a 22GB memory mod, limited to 250 W, gained +36% to +64% on FP16 matrix multiply depending on size. A user's 260 W variant hit +59%. A separate project, KohakUnlock, ran on a Leadtek 2080 Ti 22GB and saw an 85% improvement on an instruction-rate test, plus a best matrix-multiply median of +53% against a second, locked card of the same model. A stock ASUS 2080 Ti 11GB reached +77% via a pre-boot loader from haohervchb.

KohakUnlock's README says plainly: "not a claim of a 2× application/game speedup." The warnings are sterner than that. turing-unlock notes the mod "may damage your hardware," and KohakUnlock adds it "can hang a GPU or host, prevent boot, corrupt data, or damage hardware."

How does the exploit work?

The gpu-unlock explainer says the rate limits live in two "speed selector" registers the host driver can't write. Nvidia's signed SEC2 "Booter" firmware carries a stack buffer overflow that lets code run in "heavy secure" mode — a bug exposed after Nvidia published a debug Booter encrypted with a known public test key. A single 4-byte write with the V67 payload opens the privilege mask, the host sets the full rate, and the module resets the card before Nvidia's driver binds it.

Jon Pry documented the underlying bug in "A Canary in the Crypto Mine: Defeating Stack Protection in a GPU Secure Coprocessor," published on Zenodo in June. The overflow was found on the CMP 170HX's Heavy Secure Falcon, and the V67 constants come from WildFlash1st's cmp90hx-unlock. The pre-driver timing gpu-unlock uses appears to come from WildFlash1st's Windows EFI bootloader, while the best-known application of the exploit is the CMP Unlocker for the CMP 170HX.

Zagitta, gpu-unlock's maintainer, said on Hacker News the project is their work but "the exploits … aren't my ideas by any means" — the unlock "just flips the 2 bits" needed. Zagitta confirmed RTX 40 support is in progress, with forks already unlocking those cards, but said Hopper and Blackwell are "out of scope for this approach" at least.

What's still unverified?

Plenty. The Ada fork says other boards "remain unverified" and that shared firmware "does not establish support for untested cards." turing-unlock keeps its own list of untested Turing IDs, and gaps persist for GA103 and most Ada and Turing silicon. No project has published a roadmap, but gpu-unlock is collecting compatibility reports from GPU owners — so expect the coverage table to keep shifting.