Modders remove half-speed AI throttle from Nvidia RTX 20, 30 and 40-series GPUs

Firmware exploit unlocks full tensor performance, but gains vary and hardware damage is possible

By LineZotpaper AI staff

Published26 October 2026

Updated26 October 2026

Read Time2 min

Modders have discovered a way to remove Nvidia’s imposed half-speed AI performance limits on GeForce RTX 20, 30 and 40-series graphics cards using a firmware exploit, allowing peak tensor rates to double on some models, though real-world gains in AI workloads vary significantly and the unofficial mods carry warnings of potential hardware damage.

Three separate projects, gpu-unlock, gpu-unlock-ada and turing-unlock, each offer unofficial Linux modules that exploit a stack buffer overflow in Nvidia’s signed SEC2 “Booter” firmware to write to speed-selector registers that the host driver cannot normally touch. The vulnerability was first discovered on a CMP 170HX mining card and described by researcher Jon Pry in a paper published in June.

Performance improvements depend heavily on the card, power limits and workload. The RTX 4090 saw its FP16 tensor rate jump from 165.6 to 331.6 TFLOPS, but in one owner’s matrix-multiply benchmark cuBLAS GEMM medians improved only 12–34%. An RTX 4070 Ti on Windows fared better, with matrix-multiply gains of about 80%. Ampere cards such as the RTX 3090 Ti showed a 35% FP16 improvement and up to 80% on BF16. The gpu-unlock README notes that a 225-watt power cap produced only a 25% BF16 gain on an RTX 3090.

LLM prefill improved 14–36% on Ampere cards while decode and INT8 performance stayed “about the same.” Tokens per second in chatbots did not change because that metric depends on memory bandwidth. On Turing-based RTX 2080 Ti modded to 22 GB of memory, FP16 matrix-multiply gains ranged from 36% to 64%.

The projects explicitly caution users. The turing-unlock README states the mod “may damage your hardware,” and KohakUnlock, a separate 2080 Ti unlock, warns it “can hang a GPU or host, prevent boot, corrupt data, or damage hardware.” Experimental Windows builds exist but are not as mature as the Linux modules.

Coverage remains incomplete. The Ada fork supports the RTX 4090 and, experimentally, the 4090 D, 4070 Ti Super and 4060 Ti 16 GB. GA103 Ampere cards and most Turing GPUs are untested. Blackwell (RTX 50-series) support is “out of scope,” per gpu-unlock’s maintainer.

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Analysis

Why This Matters

Owners of older GeForce cards can substantially improve AI inference and image-generation speeds, potentially extending the useful life of already-purchased hardware.
The exploit exposes a firmware vulnerability that Nvidia may need to patch, raising questions about security in the GPU ecosystem.
The performance gains are not uniform; users must weigh potential hardware risk against modest real-world speedups for many workloads.

Background

Nvidia has long restricted tensor-core performance in consumer GeForce cards compared to its professional lineup, a design choice that helps segment the market. The half-speed limit on FP32-accumulate operations was documented in Nvidia’s published specs. The exploit relies on a stack buffer overflow in signed firmware, discovered after Nvidia released a debug Booter encrypted with a known public test key.

Key Perspectives

gpu-unlock maintainer Zagitta: said on Hacker News that the exploit “just flips the 2 bits” needed and that the vulnerability was not their discovery; they also stated that Hopper and Blackwell cards are “out of scope for this approach.” KohakUnlock README: warns that the mod “can hang a GPU or host, prevent boot, corrupt data, or damage hardware” and cautions that it is “not a claim of a 2× application/game speedup.” turing-unlock README: notes that the modification “may damage your hardware.”

What to Watch

Whether Nvidia releases a firmware update that closes the buffer-overflow exploit.
Expansion of Windows-native unlock tools, which remain experimental and may broaden the user base.
Reports from users on untested cards (GA103, most Ada and Turing GPUs) that could fill gaps in compatibility data.

Sources

Nvidia RTX 20, 30, and 40-series GPUs hacked to remove a half-speed AI throttle · Tom's Hardware

Checked

12 claims in this story checked against 1 source: 9 supported.

Show each claim

SupportedThe headline says modders removed Nvidia’s half-speed AI throttle from RTX 20-, 30- and 40-series GPUs.Latest from Tom's Hardware: “Modders have discovered a way to remove imposed half-speed AI performance limits on Nvidia RTX 20, 30, and 40-series GPUs with a firmware exploit”
SupportedThe opening says the firmware exploit can double peak tensor rates on some models, while real-world AI gains vary.Latest from Tom's Hardware: “Without the limitation, the RTX 4090 can achieve double its peak tensor rate [...] Real workloads gain significantly less than the RTX 4090’s doubled peak would imply”
SupportedThe story says the unofficial mods carry warnings of potential hardware damage.Latest from Tom's Hardware: “The turing-unlock README notes that the mod “may damage your hardware,” while KohakUnlock adds that it “can hang a GPU or host, prevent boot, corrupt data, or damage hardware””
SupportedThree projects provide unofficial Linux modules; the exploit uses a stack buffer overflow in Nvidia’s signed SEC2 “Booter” firmware to enable writes to speed-selector registers the host driver cannot write.Latest from Tom's Hardware: “Three separate projects offer generation unlocks, each with its own unofficial Linux module [...] the rate limits are in two “speed selector” registers that the host driver can’t write, but Nvidia’s signed SEC2 “Booter” firmware has a stack buffer overflow”
Partly supportedThe vulnerability was first discovered on a CMP 170HX mining card.Latest from Tom's Hardware: “The stack overflow was found on the CMP 170HX in the Heavy Secure Falcon”
SupportedResearcher Jon Pry described the vulnerability in a paper published in June.Latest from Tom's Hardware: “The underlying bug is described in Jon Pry’s “A Canary in the Crypto Mine: Defeating Stack Protection in a GPU Secure Coprocessor,” published on Zenodo in June.”
SupportedThe RTX 4090’s FP16 tensor rate rose from 165.6 to 331.6 TFLOPS.Latest from Tom's Hardware: “the Ada fork developer’s RTX 4090 moving from 165.6 to 331.6 TFLOPS on FP16”
SupportedIn one owner’s benchmark, RTX 4090 cuBLAS GEMM medians improved by 12–34%.Latest from Tom's Hardware: “in one independent owner’s matrix-multiply benchmark, cuBLAS GEMM medians only saw improvements of 12–34%”
Not in the sourcesExperimental Windows builds exist but are not as mature as the Linux modules.Latest from Tom's Hardware: “Three separate projects offer generation unlocks, each with its own unofficial Linux module, while experimental Windows builds pursue a different approach.”
SupportedAmpere LLM prefill improved 14–36%, while decode and INT8 performance stayed about the same.Latest from Tom's Hardware: “For Ampere, LLM prefill improved 14–36%, with diffusion and FlashAttention speeding up, but LLM decode and INT8 stayed “about the same.””
SupportedThe turing-unlock README says the mod “may damage your hardware.”Latest from Tom's Hardware: “The turing-unlock README notes that the mod “may damage your hardware,””
Partly supportedThe story says Blackwell support is “out of scope,” per gpu-unlock’s maintainer.Latest from Tom's Hardware: “Zagitta, the maintainer of gpu-unlock, said on Hacker News that [...] Hopper and Blackwell cards are “out of scope for this approach” at least.”

Correction

26 October 2026An earlier version of this story identified GA103 cards as Ada cards. They are Ampere cards.