An open-source Linux tool is giving Nvidia’s six-year-old RTX 3090 a second wind, even though GeForce RTX 5090 is now the newest GeForce RTX desktop flagship on sale. Notebookcheck reported on October 24 that the free, open-source gpu-unlock project, built by GitHub user Zagitta, can lift some Tensor Core workloads by up to 80% without flashing the card’s VBIOS. Mezha.net and Wccftech also covered the story, putting the RTX 3090 GPU unlock at the center of this week’s hardware news.
The headline numbers are real but narrow. In one community report, BF16 throughput on an RTX 3090 rose from 77.5 TFLOPS to 129.6 TFLOPS, and TF32 climbed from 40.2 to 69.8 TFLOPS. Notebookcheck’s headline says performance “nearly doubles,” while its own text frames the gain as up to 80% in certain compute benchmarks.
Then comes the catch. The unlock is experimental, it targets Linux, and the boost depends heavily on the workload. Nobody has published verified gaming benchmarks. If you plan to try it, watch your thermals and power draw first, and our walkthrough on how to check GPU temperature covers the sensors worth logging.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the RTX 3090 GPU Unlock Actually Does
The gpu-unlock project goes after a specific limit on consumer Ampere cards. According to the developer, FP16 and BF16 Tensor Core operations that use FP32 accumulation run at roughly half the rate available on comparable professional Nvidia GPUs. That is an instruction-issue rate limit, not a hardware defect, and the project removes it by changing the registers that control how fast those instructions issue.
The method matters as much as the result. Notebookcheck says Zagitta’s implementation is a Linux kernel module that runs before Nvidia’s graphics driver loads. It modifies the relevant registers, resets the GPU, then hands control back to the unmodified Nvidia driver. Because of that sequence, the card needs no permanent VBIOS flash and no patched driver.
That distinction is the main reason the story spread. Other tweaks to Nvidia silicon have relied on firmware writes or driver patches. A pre-driver kernel module avoids both, at least on paper. The project also reportedly covers the RTX 3090 Ti, the sibling card built on the same GA102 silicon.
The Numbers Behind the “Nearly Doubles” Headline
Three separate data points anchor the story. The first is the community report that Notebookcheck and Mezha covered, showing BF16 at 129.6 TFLOPS against a 77.5 TFLOPS baseline. The second is a GitHub compatibility report on Issue #55 that lists unlocked results at two power caps. The third is an individual social-media claim from an X user identifying as Md Ismail Šojal, who wrote that a 3090 reached 135 TFLOPS from 75 TFLOPS with a 360-watt limit on commit a9b0c75. Nobody has independently confirmed that last claim.
| Report | Metric | Before unlock | After unlock | Change |
|---|---|---|---|---|
| Community report (RTX 3090) | BF16 | 77.5 TFLOPS | 129.6 TFLOPS | About +67% |
| Community report (RTX 3090) | TF32 | 40.2 TFLOPS | 69.8 TFLOPS | About +74% |
| Issue #55 (RTX 3090, 370 W) | FP16 / BF16 / TF32 | Pending | 120.73 / 123.80 / 66.66 TFLOPS | No baseline listed |
| Issue #55 (RTX 3090, 250 W) | FP16 / BF16 / TF32 | Pending | 80.98 / 85.58 / 49.46 TFLOPS | No baseline listed |
| Single X post (RTX 3090, 360 W) | Unspecified | 75 TFLOPS | 135 TFLOPS | About +80%, unconfirmed |
| Project data (RTX 3090 Ti, 450 W) | BF16 | 87.50 TFLOPS | 157.41 TFLOPS | About +80% |
Read the table with care. The percentage changes in the right-hand column are my arithmetic from the published figures, not numbers any outlet printed. The Issue #55 rows mark the “without unlock” values as Pending, so that report shows what an unlocked card can do but not how much it gained. Only the first two rows pair a before value with an after value from the same report.
Why Power Limits Change the Result
The Issue #55 report is useful because it shows the same RTX 3090 at two power caps. At 250 W, the unlocked card posted 80.98 TFLOPS in FP16, 85.58 in BF16 and 49.46 in TF32. At 370 W, which the report lists as the factory-default cap, the same card reached 120.73, 123.80 and 66.66. The unlock does not remove the power ceiling. It removes an issue-rate ceiling, and a card that cannot feed its Tensor Cores with enough watts will not use the extra headroom.
| Precision | Unlocked at 250 W | Unlocked at 370 W | Gain from extra power (my arithmetic) |
|---|---|---|---|
| FP16 | 80.98 TFLOPS | 120.73 TFLOPS | About +49% |
| BF16 | 85.58 TFLOPS | 123.80 TFLOPS | About +45% |
| TF32 | 49.46 TFLOPS | 66.66 TFLOPS | About +35% |
A second RTX 3090 report, run at a 280 W limit, showed smaller gains: 43% in BF16, 38% in FP16 and 47% in TF32. The same test recorded roughly 50% higher FlashAttention 2 performance and 29% higher masked cuDNN SDPA performance. So the “up to 80%” figure comes from favorable setups, and a power-capped card should expect less. GSMGoTech’s roundup of the project’s reports adds that the weakest gain it found, about 25%, came from a card capped at 225 W.
Check Your Power Cap Before You Compare
Anyone who wants to repeat these tests should record the card’s configured limit first. Two runs at different caps are not comparable, and a hidden limit would explain most of the spread between reports. The standard nvidia-smi query below prints the active and default limits side by side.
nvidia-smi --query-gpu=name,power.limit,power.default_limit,temperature.gpu --format=csvHow the Register Change Works
The project’s documentation, as relayed by Notebookcheck and other coverage on October 24, 2026, describes a three-step routine. The kernel module writes new values to the GPU speed-selection registers that govern Tensor Core instruction issue. It then resets the GPU. Finally, it lets the stock Nvidia driver bind to the card as usual. The driver never sees a modified binary.
GSMGoTech adds that the module loads from the initramfs, so it runs ahead of the proprietary driver at boot. It also says the approach builds on earlier research into unlocking Nvidia’s CMP mining GPUs. The practical upside, according to that outlet, is that owners do not have to re-patch anything when Nvidia ships a new driver release. GSMGoTech also says the reports that compared outputs found them bit-identical before and after the unlock, which suggests the change affects speed and not accuracy.
Still, “should not corrupt results” is a weaker promise than “has been validated.” The compatibility reports are volunteer submissions, they cover a small number of cards, and they run on different software stacks. A single-GPU test rig in someone’s garage is not a qualification lab.
The Catch: Linux, Experimental, Workload-Dependent
Three limits define the “catch” in the headline. First, the project is a Linux kernel module, so Windows users are out of luck for now. Notebookcheck says a separate experimental Windows port has booted an unlocked RTX 3090 into Windows on one specific system, but that is a single data point, not a supported path.
Second, the unlock is experimental. Notebookcheck warns that installing it on a primary gaming PC carries risk, and that advice is fair. A tool that resets the GPU early in boot can fail in ways that leave a machine unbootable until you remove the module. No source in this story describes permanent damage to a card, and none rules out system instability either.
Third, the gains vary by workload. One RTX 3090 user reported a 16% improvement in long-context prompt processing for local LLM inference, while token generation speed stayed practically unchanged. Matrix-heavy jobs that sit on the throttled Tensor Core path benefit most. Everything else sees little or nothing.
Gaming Gains Remain Unproven
The word “performance” in the headline invites a gaming reading, and the evidence does not support it. Notebookcheck says no independently verified gaming benchmarks show any frame-rate improvement. Nvidia’s DLSS uses Tensor Cores, so some AI-assisted rendering steps could in theory benefit, but theory is all anyone has so far.
Most game rendering does not lean on FP16 or BF16 Tensor Core math with FP32 accumulation, which is the exact operation the unlock targets. Rasterization and shader work run through other paths. Treat any claim that this tool will “double your FPS” as unsupported. The sources in this story report compute gains, and any claim that gaming performance nearly doubles remains unconfirmed.
That gap between headline and data is worth repeating because it drives clicks. A reader who owns a 3090 for gaming and installs a Linux kernel module hoping for 60 extra frames will likely end up with a risky install and no benefit.
Where Local AI Users Will Feel the Difference
The people with the most to gain are the ones running models on a home workstation card. GSMGoTech reports that local LLM users saw 14% to 36% faster prompt prefill on long-context inference, and that ComfyUI testing on an RTX 3090 Ti showed roughly 47% to 59% higher throughput across several BF16-heavy image-generation workloads. Notebookcheck’s single-user figure of 16% sits inside that prefill range.
| Workload | Reported effect | Outlet or report |
|---|---|---|
| BF16 matrix multiplication (RTX 3090) | 77.5 to 129.6 TFLOPS | Notebookcheck, community report |
| TF32 matrix multiplication (RTX 3090) | 40.2 to 69.8 TFLOPS | Notebookcheck, community report |
| FlashAttention 2 (RTX 3090, 280 W) | About +50% | GitHub compatibility report |
| Masked cuDNN SDPA (RTX 3090, 280 W) | About +29% | GitHub compatibility report |
| Long-context prompt processing | +16% (one user) | Notebookcheck |
| Prompt prefill, local LLMs | +14% to +36% | GSMGoTech |
| ComfyUI BF16 image generation (RTX 3090 Ti) | About +47% to +59% | Notebookcheck, GSMGoTech |
| LLM token generation | Practically unchanged | Notebookcheck |
| Gaming frame rates | No verified data | Notebookcheck |
The pattern in that table is consistent. Compute-bound work that feeds the Tensor Cores gains the most, while memory-bandwidth-bound work stays flat. GSMGoTech explains the flat token generation by pointing out that decoding is limited by memory bandwidth and not by compute. Anyone setting up a local model stack should read our guide to running an LLM locally with llama.cpp and Ollama and then decide whether prefill speed or generation speed matters more for their use.
Historical Context: From CMP Mining Cards to GeForce
The idea of unlocking hardware that Nvidia deliberately slowed is not new. Both Notebookcheck and GSMGoTech note that gpu-unlock builds on earlier research used to unlock capabilities on Nvidia’s CMP mining GPUs. Those cards were sold to crypto miners with limits on non-mining features, and researchers found ways around them. Zagitta’s project borrows from that technique and combines it with pre-driver timing, which GSMGoTech compares to a UEFI-based approach.
GSMGoTech frames the Tensor Core limit as deliberate segmentation between GeForce and workstation GPUs, and the project’s own description matches that reading, since it compares the throttled rate to professional cards. Nvidia has long separated consumer and professional lines through firmware and driver rules. This story shows how long such limits can survive, since the RTX 3090 is six years past launch and the restriction only now has a public workaround.
Nvidia has not commented in any source I reviewed. There is no confirmed official endorsement of gpu-unlock, and no on-record statement from Nvidia or from Zagitta appears in the coverage. That silence is typical for projects that sit in a gray area between tinkering and circumvention.
Market Impact for Used Ampere Cards
I have no verified price data for used RTX 3090 cards, so I will not quote one. The direction of the effect is still easy to reason about. Local AI hobbyists already use the 3090 for inference, and a free boost to the compute side makes that card more attractive against newer, pricier options such as GeForce RTX 5090, the newest GeForce RTX desktop flagship on sale. The same logic applies to the 3090 Ti, where the project’s own data reportedly shows BF16 rising from 87.5 to 157.4 TFLOPS at 450 W.
Nvidia’s wider business is not at risk from this. As our earlier look at Nvidia’s sub-17x forward earnings multiple showed, the company’s valuation debate centers on data center demand and custom silicon, not on the aftermarket for six-year-old consumer cards. A Linux kernel module will not dent that story. It may, however, nudge a few buyers to delay an upgrade, which matters more to board partners and resellers than to Nvidia itself.
Nvidia’s newer consumer launches face a quieter version of the same question. Our coverage of the NVIDIA RTX Spark launch at $2,599 shows how the company prices new hardware against Apple Silicon. Against that backdrop, any trick that stretches an old card’s life changes the buy-or-wait math for someone who only runs inference at home.
Competitive Comparison: Unlock vs. Upgrade vs. Alternatives
An unlocked RTX 3090 competes with three other choices. The first is buying a newer GeForce card. The second is moving to a unified-memory machine built for local AI, such as the one described in our piece on the AMD Ryzen AI Halo with 192GB for 300B-class models. The third is doing nothing and living with the stock rate.
The unlock wins on cost, since the sources describe it as free, and on simplicity, since it needs no VBIOS flash. It loses on support, platform reach and certainty. A newer card ships with a warranty and a driver path from Nvidia. A unified-memory box can hold much larger models than a single consumer GPU. The unlock offers neither of those, so it suits an owner who already has the card, already runs Linux, and already accepts that this is a hobbyist hack.
Against professional Nvidia cards, the unlock narrows a gap rather than closing it. The project says the throttled operations ran at about half the professional rate, and the best reported RTX 3090 results fall well short of a clean doubling in most setups. I read the result as a partial recovery of locked performance, not a new capability.
Risks and Sensible Testing Practice
The safest way to evaluate gpu-unlock is on a secondary machine. Keep the card at its default power limit, record a baseline with your actual workload, and only then load the module. Without a baseline, you will repeat the same gap seen in Issue #55, where the “without unlock” values are still Pending.
What to Log During a Test
Log temperature, power draw, clock speed and throughput for at least two runs before and two runs after. A free tool like HWiNFO can capture sensor data to CSV, and our guide to HWiNFO sensor logging for throttling checks shows the setup for Windows machines. On Linux, the nvidia-smi query shown earlier covers the same ground. Higher Tensor Core throughput means more heat, so a card that ran cool at the stock rate may throttle once the unlock lets it work harder.
When to Walk Away
Skip the unlock if the card is your only gaming machine, if you cannot recover a Linux boot problem, or if your jobs depend on memory bandwidth rather than matrix math. In those cases the downside is real and the upside is close to zero. Also remember that no official Nvidia support exists for this change.
Predictions: What to Expect Next
These are my forecasts, not reported facts. I base them on how similar community projects have moved and on the early traction of the compatibility reports.
- More benchmark reports will land within weeks. The Issue #55 report still lists locked baselines as Pending, and volunteers will fill those gaps. Expect the headline number to settle somewhere below “nearly doubles” for power-capped cards.
- Gaming tests will arrive and show little. Someone will run DLSS-heavy titles on an unlocked card. Given how narrow the targeted operation is, I expect small or no frame-rate gains.
- The Windows port will get more attention. A single system booting an unlocked 3090 into Windows is enough to draw testers, and a stable Windows path would widen the audience sharply.
- Nvidia will stay quiet or respond through drivers. A formal statement seems unlikely. A driver or firmware change that blocks the early-boot reset is possible, though I have no evidence one is planned.
- Local AI forums will adopt it as a tuning step. Prefill gains of 14% to 36% are enough to matter for long-context work, so expect it to appear in home-lab guides alongside power-limit tuning.
For more hardware coverage on this site, browse the hardware category and check back as new compatibility data arrives.
Frequently Asked Questions
Does the RTX 3090 unlock double gaming performance?
No verified evidence says so. Notebookcheck states that no independently verified gaming benchmarks show any frame-rate gain. The reported improvements come from compute workloads that use FP16, BF16 and TF32 Tensor Core math.
Does gpu-unlock require flashing the VBIOS or patching the Nvidia driver?
No, according to Notebookcheck. The project runs as a Linux kernel module that changes registers before the Nvidia driver loads, resets the GPU, and then hands control back to the unmodified driver.
How much faster does BF16 get on an RTX 3090?
One community report measured BF16 matrix multiplication rising from 77.5 TFLOPS to 129.6 TFLOPS, about 67%. Other setups differ. The Issue #55 report lists 85.58 TFLOPS at 250 W and 123.80 TFLOPS at 370 W for an unlocked card.
Does the unlock work on Windows?
The original project targets Linux. Notebookcheck says a separate experimental Windows port has booted an unlocked RTX 3090 into Windows on one specific system, but it is not a supported or broadly tested path.
Will local LLM token generation speed up?
Probably not by much. Notebookcheck cites one user who saw a 16% gain in long-context prompt processing while token generation stayed practically unchanged. GSMGoTech attributes that to token generation being bound by memory bandwidth.
Which cards does the project cover?
The project targets Ampere GeForce cards, including the GA102-based RTX 3090 and RTX 3090 Ti. GSMGoTech says its tested list also includes the RTX 3080, 3070 and 3060 12 GB, though the OEM-only RTX 3080 Ti 20 GB completes the unlock without gaining BF16 rate.
Has Nvidia endorsed gpu-unlock?
No official Nvidia endorsement appears in any source reviewed for this story, and no on-record statement from Nvidia or Zagitta has surfaced. Treat the tool as an unsupported community project.
