mrkeyoor.comAsk

Zagitta/gpu-unlock

gpu-unlock review

Esequiel378 reports 51% faster Z-Image Turbo and 61% faster BF16 prefill on one RTX 3090, while our sandbox stopped before execution.

Not run yet

AI Toolsevaluationupdated 17 Nov 2026
+456 / 4dstars / 7d

Zagitta/gpu-unlock is a Linux kernel module that removes a throughput limit on selected NVIDIA GeForce tensor operations used by AI software. It runs before the normal NVIDIA driver, targeting faster matrix work rather than more video memory or faster gaming.

Verdict

Esequiel378 reports 51% faster Z-Image Turbo and 61% faster BF16 prefill on one RTX 3090, while our sandbox stopped before execution. Use gpu-unlock if you control a Linux AI host, have an affected workload and can recover through a full power-off. Skip it for gaming, guaranteed quantized-chat gains or a machine you cannot power-cycle.

Infographic on how gpu-unlock works, its uses, what we have not run yet, common pitfalls and who should use it
We have not run gpu-unlock yet; the middle band lists what our review has not tested.

We ran it

Screenshot of gpu-unlock (github.com/Zagitta/gpu-unlock)
github.com/Zagitta/gpu-unlock

Answers from our run

Did you run gpu-unlock yourself?

No. Its code is C, and it carries no manifest our lab installs from, and no Dockerfile, so there was nothing standard to install, build or test. This review is written from the repository's own documentation.

Who should not use gpu-unlock?

Gamers seeking an FPS upgrade: the upstream workload table lists no gaming gain.

What are the alternatives to gpu-unlock?

turing-unlock, Windows-GPU-UNLOCK, vgpu_unlock. Use gpu-unlock if you control a Linux AI host, have an affected workload and can recover through a full power-off.

SetupN/ANot assessed: installation was not reproduced in a supported environment.
Docs4/5The README has an install section (“How to install”) and usage examples in code; we did not run its quickstart ourselves.Editor’s note: The README covers boot ordering, explicit targets and packaging. We have not reproduced those instructions on a GPU host.
Community5/5Last push on 2026-11-17, 20,054 stars and maintainers answered or closed 7 of the 10 most-discussed issues we sampled.Editor’s note: Sampled issues contain maintainer replies and detailed user reports; the sample cannot establish adoption or support response times.
Maturity1/5No tagged GitHub release, repository created 2026-10-18 and no licence file GitHub recognises.Editor’s note: Turing and Ada are labeled experimental, and board-specific failures appear in the sampled issues. Production reliability was not measured.

How we score

Discussed on

  1. hnGPU-unlock: FP/BF16 instruction issue rate on Nvidia 3000's cards353 points

Who it’s for

Linux AI users with an affected tensor workload and control of the host boot configuration.
Image-generation and BF16 prefill users prepared to compare their exact application against stock.
Hardware experimenters who can recover the machine through a full power-off and retain a stock boot entry.

Who it’s NOT for

Gamers seeking an FPS upgrade: the upstream workload table lists no gaming gain.
Quantized-chat users expecting every model to improve: grumd reports unchanged Q6_K prefill and decode.
Remote operators without power-cycle access: a failed attempt can require full power-off recovery.
Windows users seeking a mainline supported installation: the maintainer points to a separate port and does not maintain Windows.

Setup reality

We did not run gpu-unlock: the 2026-11-17 record at 7eb51bc says no supported ecosystem for C and no Dockerfile. The configured sandbox had 3 CPUs and 8 GB RAM, with image n/a. No install, build or test timing was recorded.

The README describes a Linux host with kernel headers and distribution-specific packaging, an explicit target in the kernel command line, and an initramfs that runs gpu_unlock before NVIDIA binds the card. This requires host administration rather than application credentials.

Custom initramfs configurations, laptop runtime power management and exact board compatibility affect the documented path. Keep a stock boot entry and access to full power-off recovery. The sandbox image label supplies no actual runtime or operating-system version.

gpu-unlock has credible reported gains for selected AI workloads, including 51% faster Z-Image Turbo in Esequiel378's RTX 3090 report. Use it if you run tensor-heavy AI on a Linux machine you control and can recover with a full power-off. Skip it if your goal is faster gaming or a guaranteed improvement to quantized local chat. Our sandbox did not execute the project: its C code had no supported harness ecosystem and no Dockerfile. RTX 3090 results

What people use gpu-unlock for

Esequiel378 reports 51% faster Z-Image Turbo, 38% faster SD 1.5 and 61% faster BF16 vLLM prefill on one RTX 3090 with the unlock alone. Those are application results worth reading before the README's peak tensor figures. Prefill means processing the prompt before generation begins; decode means generating the answer token by token. A faster first stage needn't make a long streamed answer proportionately faster. Benchmark report

Who used it Work performed Reported outcome
Esequiel378 Z-Image Turbo, SD 1.5 and BF16 vLLM on an RTX 3090 under NixOS Unlock-only throughput gains of 51%, 38% and 61%, respectively. Results
GitHub user xielidun Qwen3.8 Flash Next IQ4_XS in Strata across four RTX 3060 12GB cards Prefill rose from 450 to 600; decode barely changed. The opening report supplies no unit for those figures. Issue #31
Reddit user grumd Qwen 27B Q6_K after confirming the RTX 3080 20GB unlock No measured prefill or decode improvement through the default quantized execution path. LocalLLaMA discussion
GitHub user ilikenwf Mainline gpu-unlock on a Clevo RTX 4090 Laptop in NVIDIA mode Reports approximately doubled instruction throughput and 175 W maximum draw. The suggested role of the earlier compute mode is the reporter's hypothesis. Laptop report

The Strata and Q6_K reports make a useful pair. Both involve quantized models, yet their prefill outcomes differ. A model's filename doesn't tell you which arithmetic kernels dominate its runtime. Choose a representative prompt and execution backend before deciding that another user's result applies to your machine. These are individual reports, with their authors and configurations attached, rather than a fleet-wide success rate.

What happened when we ran it

Our 2026-11-17 sandbox record at commit 7eb51bc says NOT RUN: no supported ecosystem for C and no Dockerfile. The recorded configuration was 3 CPUs and 8 GB RAM, with image listed as n/a. That is a harness coverage finding, not a failed GPU benchmark or evidence that gpu-unlock itself is broken.

Recorded item Result What it tells you
Source snapshot 7eb51bc, 2026-11-17 Identifies the recorded clone
Sandbox configuration 3 CPUs, 8 GB RAM, image n/a Describes the allocated sandbox, not measured resource consumption
Execution NOT RUN: C ecosystem unsupported; no Dockerfile No installation, build or test result to score
Separate hands-on safety step FAILED: none of compose.yaml, compose.yml, docker-compose.yaml or docker-compose.yml existed The Compose-only safety rule stopped the attempted route.
Teardown ok Cleanup status only.

The hands-on run record also contains a generic sentence saying the project was deployed with its own Docker Compose. That conflicts with the specific failed safety check, which says the required file was absent. We reject that deployment label as evidence of a running system. There is no application result behind it to report.

The rest of this review evaluates documentation and attributed community results. Installation time, test totals and GPU speed from our own machine are unavailable. The image label supplies no operating-system or runtime version. Buying hardware on the assumption that MrKeyoor reproduced the RTX 3090 numbers would misread this review.

Where people get stuck: the module must beat NVIDIA to the card

GitHub issue #93 reports error -16 after manually loading the module once NVIDIA already owned the GPU. The reporter, maksakovaa, later says reinstalling from fix-mkinitcpio-systemd-hook worked. That sequence is more useful than treating every failure as a firmware incompatibility: a package can be present while the early-boot step never happens. Issue #93

Symptom and source What the record says Action supported by that record
Only the boot argument appears in the journal, #7 Zagitta says the module isn't loading at all Inspect the actual initramfs and module-loading path; a boot flag alone isn't success. Installation thread
Manual load returns -16, #93 The attempted load occurred after NVIDIA bound the GPU; the reporter later credits the hook-fix branch Repair early loading rather than repeatedly loading the module after boot. Hook report
Turing changes from -19 to -110, #242 The loader-halt branch alone left reporters timing out; additional work produced successes Read the exact patch and card reports. The open thread records a workaround, not a universal fix. Turing thread
Laptop unlock disappears, #227 pantha704 says a modprobe.d setting wasn't honored for the initramfs-loaded NVIDIA module Their workaround put nvidia.NVreg_DynamicPowerManagement=0x00 on the kernel command line. Laptop thread

In #7 the user had already added the Limine argument and could see it in the generated configuration. Their journal still showed only the command line. Zagitta explained that a real attempt should log either an unlock or a failure. The thread eventually contains a success report, but the supplied extract doesn't provide a complete reproducible resolution. Copying that user's final status without the missing steps would be poor installation advice.

Turing has a different story. In #242, celisej567 reports that mainline 7eb51bc failed with -19, while the loader-halt branch progressed to -110 and left NVIDIA unable to bind. RakkiPoni's additional patch then drew success reports for TU102, TU104 and TU106. Ledzpg reports one failed cold boot followed by success with the identical patched tree. Keep the actual boot logs with a compatibility report; changing code isn't the only thing that changed between those two outcomes. Patch progression

A matching GPU name isn't a matching board

Issue #28 contains different outcomes for cards sold as GA102-250 RTX 3080 Ti 20GB. The original board reported a successful selector change without the expected BF16 uplift. Other contributors subsequently reported full-rate results on their own GA102-250 cards. This is a reason to collect board identity before making a compatibility judgment. OEM compatibility thread

The original reporter lists PCI ID 10de:2205, board ID 0x300 and VBIOS 94.02.5A.00.32. A successful comparison board has ID 0x151E and VBIOS 94.02.5A.00.02. Downgrading the original system's driver from 615 to 610 improved its measured throughput only partially. A later Windows UEFI test also improved that same board without reaching the comparison result. The thread leaves its underlying constraint unresolved; it doesn't prove that every OEM card fails or that one driver downgrade fixes the class.

For a useful report, record the PCI and subsystem IDs, board ID, VBIOS, kernel, NVIDIA driver, gpu-unlock commit, power cap and benchmark precision. Separate the selector readback from the workload outcome. The board in #28 is the concrete example of why an apparently successful register operation and the expected application gain require separate evidence.

Kernel privileges and a stock boot entry are part of the decision

The upstream README says gpu-unlock runs before the stock NVIDIA driver and requires an explicit kernel-command-line target. It changes firmware state in memory and resets the card before driver handoff. That puts the experiment in the boot path of the machine you rely on, so a recoverable test host is a sensible requirement. Upstream documentation

Esequiel378's report says a failed unlock can leave the GPU unusable until a full power-off and that a reboot can preserve the unlock. Keep a stock boot entry and access to a complete power cycle. Removing the target option controls future attempts; restoring a known stock hardware state can also require cutting power. The Ada guide separately warns against force-unloading a module while firmware may still use retained DMA buffers. Recovery report Ada recovery guidance

Use Zagitta/gpu-unlock as the project identity. Its FAQ explicitly disclaims associated websites, coins and installers. An unofficial domain can offer a useful explanation without becoming the developer's documentation, and a fork repeating the README isn't another independent performance test. The kernel module's GPL-2.0-only terms also sit alongside separate NVIDIA firmware terms in the upstream notices. Upstream FAQ and notices

Quantized chat and board energy need different readings

Reddit tester grumd reports no prefill or decode improvement for Qwen 27B Q6_K through llama.cpp's default MMQ path. Their explanation is specific: the quantized path uses INT8-to-INT32 arithmetic rather than the throttled floating-point accumulation path. Retaining some BF16 tensors therefore doesn't, by itself, predict a material speedup for the whole model. Execution-path discussion

That discussion also compares MMQ with cuBLAS accumulation paths. grumd reports a speed tradeoff, extra VRAM use and lower measured divergence from a BF16 reference for the cuBLAS paths in their numerical comparison. Those observations concern different execution choices; they aren't evidence that enabling gpu-unlock automatically improves model accuracy. Before switching a backend, compare useful output, memory fit and speed together. A path with a smaller numerical difference can still be the slower choice for the job.

Esequiel378's roughly 40% lower energy per output belongs to the combined unlock and clock-capped curve-offset configuration. The report measures GPU board energy, excludes the rest of the computer and infers lower core voltage rather than measuring it. That makes the result useful for assessing board work per output, but it doesn't supply a household electricity-saving percentage. Keep unlock-only and combined-tuning results separate when comparing them with your own baseline. Energy method

The same author reports stock-matching hashes across 540 image batches in a two-hour soak and zero mismatches across 260 hashed benchmark runs. This is stronger correctness evidence than merely reaching a higher tensor rate. It is still attached to the tested configuration: the report explicitly warns that an unstable clock offset can produce wrong results without crashing. Treat output checking as part of a tuning experiment, rather than assuming a completed job was correct. Correctness and tuning

gpu-unlock vs turing-unlock, Windows UEFI ports and vgpu_unlock

Choose gpu-unlock for the documented Linux tensor-throughput approach; choose a different project when you need a different operating-system route or GPU virtualization. These names overlap enough to mislead searchers. The table separates what people are trying to accomplish from claims about which implementation is faster.

Alternative Reason to choose it Evidence and boundary
Separate turing-unlock Investigate the earlier dedicated Turing implementation and its reports Ph0rk0z reports it working alongside gpu-unlock for 3090s, with moving synthetic benchmarks but no noticeable ComfyUI gain. Mainline later added experimental Turing support. #30 #64
Windows-GPU-UNLOCK You need the Windows UEFI route named by upstream Zagitta links the port and says they don't maintain Windows. The laptop thread records an ineffective GA107 Optimus port. FAQ #227
vgpu_unlock or vgpu_unlock-rs You want one GPU shared across virtual machines The supplied unofficial comparison describes vGPU functionality, a different goal from restoring tensor throughput. Feature comparison
Clock tuning or undervolting You want to adjust the frequency/power tradeoff Esequiel378 tests a combined configuration, with better image and prefill results than unlock alone on that card. Matched results

Patched NVIDIA drivers are another route people discuss. On Hacker News, Zagitta describes this module as a less annoying package than patching the driver. In issue #203, the maintainer says the required VBIOS portions cannot be edited to perform this unlock. Vix Loader's ability to load a modified VBIOS therefore doesn't demonstrate an interchangeable solution. Author's explanation VBIOS discussion

What the other reviews get wrong

MakeUseOf's remove-the-flag-and-reboot rollback advice needs the cold-power-off qualification from the detailed RTX 3090 report. Other broad claims also need the original workload and card attached. An instruction-rate result, a working desktop and a shorter generation job answer different questions. MakeUseOf guide Recovery evidence

The unofficial gpu-unlock.com FAQ describes separate Ampere, Ada and Turing modules. Zagitta later announced experimental Turing and Ada support in mainline 2.0 in #64. Older coverage saying RTX 40 testing is unconfirmed also misses the independent RTX 4090 reports and ilikenwf's mainline laptop result. Those updates improve the evidence for particular configurations; the experimental label still matters. Older FAQ Maintainer update Laptop result

The upstream table's categorical no-change entries deserve scrutiny too. Esequiel378 reports modest decode gains, while the Ada guide distinguishes an FP8 instruction-rate gain from a matched FP8 GEMM that barely changed. Neither supports promising large improvements to every decode or FP8 workload. The observed operation matters more than the precision label alone. Decode results FP8 distinction

The X post by itsharmanjot says there is no BIOS risk. That framing omits the reported need for full-power-off recovery. Meanwhile, the video titled I Tested the GPU Unlock Hack: Is It Worth the Risk? opens by contrasting RTX 3090 arithmetic rates and discusses different accumulation modes. The supplied transcript excerpt supports that explanation, not a measured gaming result. X post Video transcript

Questions people ask

Mixed-precision quantized models can show no speed gain: grumd reports unchanged Q6_K prefill and decode despite a confirmed unlock. The execution path determines whether the newly available throughput helps. LocalLLaMA report

Does gpu-unlock help mixed-precision quantized models that retain some BF16 tensors?

Some quantized workloads improve prefill and others don't. xielidun reports a prefill increase in Strata, whereas grumd reports no gain for Q6_K through MMQ. Benchmark the complete model and backend you intend to use; a retained BF16 tensor doesn't settle the question. Strata report MMQ report

Does gpu-unlock increase GPU temperature?

The upstream FAQ says additional compute may raise temperature slightly while normal power management remains active. Reddit users grumd and andy2na report no temperature change in their unaffected workloads. Those reports describe different loads, so they don't supply one temperature prediction for your card. FAQ User observations

Can gpu-unlock improve gaming performance?

The upstream workload table lists no change for gaming and FP32 CUDA. Choose it for an affected AI computation rather than as a gaming upgrade. The available tensor-rate demonstrations don't turn that documented boundary into an FPS claim. Upstream workload table

Why does NVIDIA limit these tensor operations on GeForce cards?

Zagitta's FAQ attributes the limitation to market segmentation between consumer and professional cards. That is the developer's explanation of NVIDIA's motive, rather than an NVIDIA statement supplied in this evidence. The narrower technical distinction concerns floating-point tensor operations with FP32 accumulation. Developer FAQ

How do I install gpu-unlock on CachyOS with Limine?

The #7 reporter added the target argument in /etc/default/limine and regenerated the configuration with limine-update, but that alone didn't load the module. Pair the bootloader argument with the correct early initramfs loading path. Preserve existing boot arguments and use your card's full PCI address; the sample address isn't a device selector for every machine. Limine installation thread

What should the verification log show after a successful unlock?

Zagitta says targets=all should produce either gpu_unlock: unlocked or gpu_unlock: unlock failed (...). Seeing only the kernel-command-line text means the module isn't loading. NVIDIA remaining the driver in use afterward is expected. Follow the unlock message with a healthy driver check and a matched workload comparison. Maintainer explanation Ada checks

Which benchmark should I use to verify real performance gains?

Use a tensor benchmark with the affected accumulation mode, then time your actual image or model workload. Esequiel378 separates GEMM, image generation, prefill and decode, and compares output hashes. Ph0rk0z's report of moving synthetic benchmarks without noticeable ComfyUI gains shows why the application check matters. Keep driver, kernel, power cap and workload settings matched. Method ComfyUI report

Is NVIDIA driver 610.43.03 required?

The upstream FAQ says no: that pinned driver supplies firmware inputs. The runtime driver is a separate choice. Issue #28 records a partial improvement after a driver change on one unusual board, which doesn't turn that version into a universal runtime requirement. FAQ Board-specific report

Can RTX 20-series Turing cards use gpu-unlock?

Mainline 2.0 includes experimental Turing support according to Zagitta. Issue #242 shows successful patched configurations alongside failures and inconsistent cold-boot behavior. Treat a particular TU102, TU104 or TU106 report as evidence for that setup, rather than a blanket assurance for the generation. Announcement Patch reports

Does gpu-unlock support the RTX 2000 Ada?

A commenter asks about RTX 2000 Ada in #64, but the supplied thread gives no answer or test result. The professional model's name doesn't make it a Turing RTX 20-series card, and that thread's experimental-generation announcement supplies no model-specific outcome. Unanswered question

Can Turing and Ampere unlocks run together in one machine?

Yes, individual reports describe mixed systems. Ph0rk0z used the separate Turing implementation alongside gpu-unlock for 3090s; JayCTBee later reports three RTX 2060 cards and three RTX 3060 cards unlocked with the additional Turing patch. Keep the implementation and patch identity attached to either result. Earlier arrangement Mixed patched system

Is AD103 support in scope, and what validation is needed?

AD103 appears in the experimental Ada work, and #4 discusses the validation needed for a port. The maintainer asks for boot logs, driver output and before/after benchmarks. The Ada guide additionally requires the exact selected card's unlock message, healthy nvidia-smi and a workload check. Offline firmware similarities alone aren't an application result. Port discussion Experimental checks

What we have not tested yet

We have not reproduced installation or executed gpu-unlock on a supported GPU. Our sandbox stopped before execution, and the separate Compose safety check failed. No local build timing, test count, vulnerability audit, output-correctness result, thermal measurement or runtime version is available. The supplied latest-release lookup is null and does not establish the latest release.

The community reports leave practical questions open: whole-system energy savings, audio-processing models and a family voice-assistant workload lack a measured answer in this pack. The RTX 3090 benchmark disabled resume handling, and upstream excludes hibernation from its documented resume support. The GA107 Windows Optimus report failed to retain an effective unlock. Fleet reliability, every laptop power-management configuration and every board variant remain untested by us.

Our recommendation is to start with a stock reference for the exact job you want to accelerate, plus a tested route back to a powered-off stock boot. If that job already runs through an unaffected quantized kernel, the published no-gain result is a good reason to leave the boot path alone.

Alternatives

ProjectWhat it isPick it when
turing-unlockThe separate Turing implementation discussed in upstream issue #30.Investigate the dedicated Turing reports or earlier mixed-generation setups; mainline 2.0 also advertises experimental Turing support.
Windows-GPU-UNLOCKThe Windows UEFI port linked by Zagitta.Your target machine runs Windows and you accept a separate implementation with its own card-specific limits.
vgpu_unlockA different unlock aimed at NVIDIA vGPU features for virtual machines.You want to split a GPU across VMs rather than change tensor instruction throughput.

What people are saying

  1. [velocity-scout] impacifist/Windows-GPU-UNLOCK
  2. [hackernews] GPU-unlock: FP/BF16 instruction issue rate on Nvidia 3000's cards
  3. [velocity-scout] JustAResearcher/gpu-unlock-ada
  4. [velocity-scout] Zagitta/gpu-unlock

Sources

  1. 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
  2. GitHub - Esequiel378/gpu-unlock-3090-results: GPU Unlock on an RTX 3090: +55% image gen, +68% LLM prefill, 40% less energy, bit-identical outputs. Raw data, scripts, NixOS module.
  3. People who tried gpu-unlock on 3090s and such - does this make any difference on mixed-precision quants where you get some parts of the model quantized to <=Q8 but keep some tensors at BF16?
  4. Compatibility report: GA102-250 (RTX 3080 Ti 20 GB OEM): unlock completes, selector verified held at runtime, but BF16 rate stays at half speed
  5. Unlock failure: EVGA GeForce RTX 2070 SUPER BLACK 8 GB (10de:1e84)
  6. RTX 3050 Ti Laptop (Ampere) GPU unlock compatibility
  7. How to install it?
  8. gpu-unlock-ada/QUICKSTART.md at main · JustAResearcher/gpu-unlock-ada

More ai tools reviews