Skip to main content
GatiFlowIntelligence Platform
Deep DiveLoginRegister

GatiFlow Intelligence Platform

TermsPrivacyComplianceOpt-outMethodologyAcademyChangelogStatus

← All Deep Dives
Saturday Deep Dive

A Kernel Module Lifts the Half-Speed Cap on Older GeForce Cards, and the Gains Are Real but Narrow

Saturday edition · October 31, 2026 · 2478 words · 13 min read

A Linux kernel module called gpu-unlock removes a limit that Nvidia builds into GeForce tensor cores. FP16, BF16 and TF32 matrix math with FP32 accumulation, the arithmetic behind most AI training and unquantized inference, runs on these cards at half the rate that the same silicon reaches in Nvidia's professional products. GatiFlow first logged the repository on October 25 at 5,322 stars and last read it on October 28 at 12,301, a 131.1 percent gain in under three and a half days, and Trendshift ranked it first overall on October 24 and 25. The headline promise is up to double the speed. The evidence from the project's documentation, from users and from GatiFlow's tracking describes something narrower: real gains on specific workloads, stability on one GPU generation, a method that rests on a firmware exploit, a spike that has started to cool, and a used-card market that was already firm.

The mechanism is a firmware exploit rather than a driver patch. The rate limit sits in two speed-selector registers that the host driver cannot write. The project's own write-up describes a stack-overflow bug in Nvidia's signed Booter firmware, found after Nvidia published a debug build encrypted with a known public test key, and already used to unlock the CMP 170HX mining card. The bug lets the module run privileged code that opens both registers to writes, and the module then sets them and resets the card, so the unmodified Nvidia driver starts on hardware already running at full rate. The write-up credits earlier researchers with finding the flaw and building the payload. There is no patched driver and no VBIOS flash, and because the work finishes before the driver loads, driver updates do not undo it. The module is licensed GPL-2.0, runs only on Linux, and stays off unless a kernel parameter switches it on. The README explains the cap as market segmentation: Nvidia wants businesses to buy professional cards at higher prices.

The headline figure is a peak rate. In the project's table of dense matrix multiplication, a 3090 Ti at its 450-watt limit rises from about 88 to about 157 TFLOPS in BF16, an 80 percent gain. On the RTX 4090, the Ada fork's own matched test shows the raw instruction rate doubling, from about 160 to about 320 TFLOPS, but matrix-multiply gains of only 12 to 34 percent in FP16 and 16 to 20 percent in BF16, and its authors call it one run pair rather than an application speedup. Real workloads show smaller gains again. The README's headline lists 25 to 55 percent for prompt processing, 25 to 50 percent for fine-tuning and 30 to 55 percent for image generation, although Tom's Hardware and TechSpot cited 14 to 36 percent for Ampere prompt processing when they reported on October 26 and 27. The README lists 0 to 15 percent for quantized models in llama.cpp or Ollama, and no change for token generation, which is limited by memory bandwidth, or for INT8, FP8 and INT4 models. Independent reports follow the same pattern: a Reddit tester cited by wccftech measured BF16 diffusion models in ComfyUI running 56 to 59 percent faster on a 3090 Ti while an INT8 model gained 0.3 percent. For a team that serves models, the unlock lowers the cost of reading prompts, tuning models and generating images. It does not move the cost of memory-bound token generation.

Stability depends on the generation. The README rates Ampere cards (RTX 30) as stable, backed by more than 100 user reports, and labels Ada (RTX 40) and Turing (RTX 20) experimental. Blackwell (RTX 50) is listed as unsupported, and the maintainer calls support unlikely. Results are not guaranteed even on Ampere: one report on an OEM RTX 3080 Ti with 20 GB described a completed unlock with no BF16 gain. Derivative projects are filling the Ada and Turing gaps. gpu-unlock-ada, a fork that adds the RTX 4090, has more than four thousand stars, and a Turing port, turing-unlock, targets the RTX 2080 Ti. The cards that gain are therefore earlier-generation cards. As long as Blackwell stays out of the tool's reach, the pool of unlockable cards is the installed base of earlier generations, and it shrinks as those cards retire.

Attention is cooling. Trendshift's daily tally shows star gains peaking at 3,709 on October 25 and falling to 1,041 on October 28, with 12,287 stars gained since October 18. GatiFlow's own readings show the same decay: the hourly pace fell by about four-fifths between the first readings on October 25 and the last reading on October 28. deepseek-ai/DeepInfer, DeepSeek's inference framework, followed an earlier version of the same shape. gpu-unlock overtook it on GatiFlow's velocity list on October 28, but DeepInfer's stars roughly doubled in the eight days after its first reading on October 20, to 23,894, and it now adds tens of stars an hour. The two repositories attack the cost of inference from opposite sides, one through software efficiency and the other through hardware headroom, and both spiked and then cooled within the same eight days. Software gains are also arriving on the same cards: wccftech cited a developer's report of a 27-billion-parameter model going from 82 to 381 tokens per second on a single 3090 through generation-side optimizations, the decode gain that the unlock does not provide. Stars measure interest. The better evidence of use is in the forks, the derivative projects and the user reports behind the Ampere rating.

The project's description changed within days of the peak. Through the afternoon of October 27 UTC, its GitHub tagline spoke of unlocking the FP/BF16 instruction issue rate on 3000-series cards. By early October 28 UTC it opened with "Your GeForce runs AI at half speed" and named the 20, 30 and 40 series. GatiFlow's report of October 25 had filed it alongside consumer device-unlock tools such as free-my-tv and nxnative. The coverage and user reports since then concern local inference and image generation, not gaming, and wccftech noted that it found no convincing evidence of a gaming gain.

The used market was already firm before the tool appeared. Price trackers put a used RTX 3090 at about $1,199 on eBay on October 1, 27 percent above its twelve-month average, and GPUDojo's weekly table has the lowest used 3090 Ti offer on eBay at $1,250 on August 7 and $1,599.99 on October 9, amid strong local-AI demand. The signal after that is mixed. The same table has that lowest offer at $2,050 on October 23, a day before wccftech's first report, and back at $1,699.99 on October 29, while GPUDojo lists the 3090 at about $1,300. On October 24, wccftech reported that RTX 3090 prices had climbed from about a thousand dollars to more than two thousand since the AI-gain reports began, and illustrated it with eBay listings of brand-new OEM 3090 Ti cards from sellers in China, a different product from a used retail card. New hardware is no easier to buy. Nvidia senior vice president Jeff Fisher reportedly said in mid-October that the RTX 5090 has not been discontinued while acknowledging that supply is constrained, and Tom's Hardware found the card missing from Nvidia's own store and listed by third-party sellers at about $7,300 on Newegg and $8,500 on Amazon, roughly four times its $1,999 list price. For buyers, that means used Ampere cards at elevated prices after a short spike and a new flagship at several times list. For investors in local-inference and edge-AI businesses, the used-card price is a fast but noisy read on demand for owned compute.

The incentive behind the cap shows in Nvidia's numbers. Since its May earnings report, Nvidia no longer reports graphics cards as a separate category and splits revenue into Data Center and Edge Computing, the latter covering PCs, game consoles, workstations, robotics and automotive. In the quarter that ended April 26, Data Center brought in $75.2 billion, up 92 percent from a year earlier, and Edge Computing $6.4 billion, up 29 percent. CNBC reported in April that compute and networking had averaged a 69 percent operating margin over three years against 40 percent for graphics. That gap fits the maintainer's explanation of the cap as segmentation: consumer silicon that can do professional tensor work at consumer prices competes with the product that carries the higher margin.

The risks start with the method. It depends on a vulnerability in Nvidia's boot firmware, so its durability depends on what Nvidia changes in future firmware and silicon, and the README expects no unlock on Blackwell. Derivative projects carry stronger warnings than the original: the Turing port says the mod may damage hardware, and KohakUnlock warns that it can freeze the card or the host machine, block booting, corrupt data or damage the hardware. The gpu-unlock maintainer calls damage very unlikely because the card's normal throttling and power management stay in place, and reports that outputs are bit-identical before and after. The unlock is reversible by removing one kernel parameter.

The legal and operational risks come next. Nvidia's GeForce driver license says the software is not licensed for datacenter deployment, with an exception for blockchain processing, and the clause itself does not say what counts as a data center. A team running GeForce cards in a colocation facility or a rented rack should have counsel read it against that deployment whether or not the cards are unlocked. The module loads at boot, ahead of the GPU driver, and generates its firmware from files in Nvidia's driver download, which the README says remain under Nvidia's own terms. The project has three contributors on Trendshift's count. The README warns that Windows installers circulating online are not from the project and urges great care with them, and Windows support exists only as third-party experimental ports.

For the decision most teams face, the arithmetic splits three ways. Renting capacity avoids the license, stability and exploit questions and leaves the hardware cost with the provider. Buying new means paying about four times list for the consumer flagship, when it can be found at all. Buying used Ampere cards to unlock them means paying prices that were already elevated in October and spiked briefly around the news, in exchange for 25 to 55 percent more throughput on prompt processing, tuning and image work and no gain on memory-bound token generation.

A break-even estimate shows how much the rental tier matters. Take a used 3090 at about $1,300, a 350-watt draw and electricity at an assumed $0.16 per kilowatt-hour, and compare it with renting the same card. At the $0.22 an hour from which RunPod lists a 3090, buying pays for itself after nearly a year of round-the-clock use without the unlock, and after six to eight months if the unlock speeds the whole workload by 25 to 55 percent. At the roughly $0.50 an hour of on-demand listings on the same tracker, the purchase pays back in about four months without the unlock and in two and a half to three months with it. The unlock shortens the payback by about a fifth to two fifths, while the choice of rental tier moves it by more than a factor of two. The estimate ignores resale value, setup time and the license clause, and it assumes every hour of work is the kind the unlock speeds up, so for bursty or decode-bound use renting wins. The unlock is strongest for teams that already own Ampere cards: for them it is a reversible gain on fine-tuning and image jobs, and the license clause is the main item to check. It is weakest for teams that would have to buy cards to get it.

Three developments will show how durable the opportunity is. The first is Nvidia's third-quarter call on Tuesday, November 17, the next scheduled occasion for commentary on GeForce supply and pricing. The second is whether Nvidia changes the boot firmware or silicon that the module depends on, and whether the Ada and Turing ports move from experimental to stable. The third is whether public analysis shows how Blackwell enforces its tensor cap, given that the README expects no unlock there. The unlock changes the arithmetic for teams that already own Ampere cards. It does not change it for teams that would have to buy those cards at today's prices.

Sources:

- GitHub - Zagitta/gpu-unlock: Your GeForce runs AI at half speed. This fixes it for 20/30/40 series cards. Up to 2× FP16 / BF16 / TF32 tensor throughput · GitHub (https://github.com/Zagitta/gpu-unlock)

- How Nvidia Ampere generation GPUs got unlocked (https://github.com/Zagitta/gpu-unlock/blob/main/docs/How%20Ampere%20was%20unlocked.md)

- Nvidia RTX 20, 30, and 40-series GPUs hacked to remove a half-speed AI throttle — firmware exploit doubles RTX 4090's peak tensor rate, lifts RTX 3090 Ti matrix math up to 80% (https://www.tomshardware.com/pc-components/gpus/nvidia-rtx-20-30-and-40-series-gpus-hacked-to-remove-a-half-speed-ai-throttle-firmware-exploit-doubles-rtx-4090s-peak-tensor-rate-lifts-rtx-3090-ti-matrix-math-up-to-80-percent)

- Modders unlock hidden AI power on Nvidia RTX 20, 30, and 40 series GPUs, nearly doubling some benchmarks (https://techspot.com/news/114351-hackers-unlock-hidden-ai-performance-nvidia-rtx-20.html)

- Zagitta/gpu-unlock — GitHub trending stats & insights (https://trendshift.io/repositories/314044)

- NVIDIA GeForce RTX 30 GPUs Get Up To 2X Faster AI Compute Performance With New GPU Unlock Mod, Real-World Gains Of Up To 60% (https://wccftech.com/nvidia-geforce-rtx-30-gpus-get-up-to-2x-faster-ai-compute-performance-with-new-gpu-unlock-mod-real-world-gains-of-up-to-60/)

- NVIDIA RTX 30 & RTX 40 GPUs Get Impressively Fast In AI LLMs As Compute Doubles With New Unlock Tool But Prices Surge Too (https://wccftech.com/nvidia-rtx-30-40-gpus-get-faster-in-ai-llms-with-new-unlock-tool-but-prices-surge-too/)

- GitHub - JustAResearcher/gpu-unlock-ada: GPU tensor unlock source for RTX 4090 and experimental RTX 4070 Ti SUPER; firmware generated locally · GitHub (https://github.com/JustAResearcher/gpu-unlock-ada)

- GitHub - nlzy/turing-unlock: Unlock full speed FP16 tensor core of NVIDIA Turing GPUs · GitHub (https://github.com/nlzy/turing-unlock)

- GitHub - KohakuBlueleaf/KohakUnlock: Experimental NVIDIA tensor-compute unlock research: guarded TU102 initramfs port, offline tests, and Ampere reproduction notes · GitHub (https://github.com/KohakuBlueleaf/KohakUnlock)

- GitHub - deepseek-ai/DeepInfer: DeepInfer: High-performance, large-scale inference framework for DeepSeek models · GitHub (https://github.com/deepseek-ai/DeepInfer)

- NVIDIA RTX 3090 Ti - 24GB usable (https://gpudojo.com/rtx-3090-ti)

- RTX 3090 pricing on Runpod (https://computeprices.com/providers/runpod/gpus/rtx3090)

- NVIDIA GeForce RTX 3090 24GB (used) (https://aiindigo.com/hardware/nvidia-rtx-3090-24gb)

- Nvidia reportedly insists the RTX 5090 isn't dead despite severe shortages and removed listings at multiple retailers — flagship GPU is missing from Nvidia's store and commands $8,500 on Amazon (https://www.tomshardware.com/pc-components/gpus/nvidia-reportedly-insists-the-rtx-5090-isnt-dead-despite-severe-shortages-and-removed-listings-at-multiple-retailers-flagship-gpu-is-missing-from-nvidias-store-and-commands-usd8-500-on-amazon)

- Nvidia denies RTX 5090 discontinued claims amid ongoing supply shortages (https://arynews.tv/nvidia-denies-rtx-5090-discontinued-claims-amid-ongoing-supply-shortages)

- NVIDIA Announces Financial Results for First Quarter Fiscal 2027 (https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-first-quarter-fiscal-2027)

- Nvidia no longer reports gaming GPU sales as a separate segment — posts eye-watering $81.6 billion Q1 profit thanks to AI boom (https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-no-longer-reports-sales-of-graphics-solutions-as-a-separate-segment-posts-eye-watering-usd81-6-billion-q1-profit-thanks-to-ai-boom)

- Nvidia faces backlash from gamers who feel abandoned for AI (https://www.cnbc.com/2026/04/18/nvidia-ai-backlash-gamers-geforce-gpu.html)

- License For Customer Use of NVIDIA GeForce Software (https://www.nvidia.com/en-au/drivers/geforce-license/)

- NVIDIA Sets Conference Call for Third-Quarter Financial Results (https://nvidianews.nvidia.com/news/nvidia-sets-conference-call-for-third-quarter-financial-results-6928873)

Disclaimer: This article is generated by GatiFlow Intelligence for informational purposes only. It does not constitute investment advice, recruitment recommendations, or legal guidance. All data is derived from public sources and AI analysis — verify independently before making decisions. Past trends do not guarantee future results.

Where this came from

Every Deep Dive starts from GatiFlow's own pipeline: 13 public developer sources, collected every six hours, with a confidence score and the evidence behind each signal. The same signals, filtered to the topics you follow, are a JSON API.

No credit card required.

Get the next one by email

One article every Saturday morning in your time zone. No account needed, and nothing else is sent to the address.

Double opt-in: you confirm by email first. What we store, and for how long, is in the privacy policy.

Tell me I am wrong

Corrections, the version of this you have lived through, or what you would like covered next. It reaches me directly and is never published. It is kept for two years so it can be read and answered; the privacy policy has the details.

0/2000