 |
| The RTX 3090 still has some hidden potential six years after launch |
Owners of Nvidia’s Ampere-era GeForce cards have a new reason to hold off on an upgrade. An open-source project called gpu-unlock can lift a hidden instruction-issue limit on the RTX 30-series, delivering big gains in certain FP16 and BF16 Tensor Core workloads — with no VBIOS flashing and no patched Nvidia driver required.
The catch? It’s Linux-only, still experimental, and the boost is highly workload-dependent. Gamers, in particular, shouldn’t expect a free frame-rate bump just yet.
What gpu-unlock actually does
The project, developed by GitHub user Zagitta, targets a limitation that throttles FP16 and BF16 Tensor Core operations with FP32 accumulation on consumer Ampere cards. According to the developer, these operations run at roughly half the rate of their professional Nvidia equivalents — a deliberate segmentation between GeForce and workstation GPUs.
The unlock works by writing new values to the hardware’s speed-selection registers (SS0 0x88888888 and the 4-bit SS1 0x8), then performing a function-level reset before the standard Nvidia driver binds the card. Because it runs as a Linux kernel module loaded from the initramfs, it executes ahead of the proprietary driver and hands control back afterwards — no firmware modification, no permanent changes to the card.
It’s a clever piece of engineering that builds on earlier research into unlocking Nvidia’s CMP mining GPUs, combining the CMP unlocker’s V67/PLM technique with the pre-driver timing of a UEFI approach. The practical benefit is that you don’t have to re-patch anything every time Nvidia ships a new driver release.
The performance numbers
Community compatibility reports paint an impressive picture for matrix-heavy work. One RTX 3090 hit 129.6 TFLOPS in BF16 matrix multiplication, up from 77.5 TFLOPS, while TF32 climbed from 40.2 to 69.8 TFLOPS. The RTX 3090 Ti fared even better in the project’s own five-sample testing, with BF16 rising from 87.5 to 157.4 TFLOPS.
| Card | BF16 before | BF16 after | Gain |
|---|---|---|---|
| RTX 3090 Ti | 87.5 TFLOPS | 157.4 TFLOPS | ~+80% |
| RTX 3090 | 77.5 TFLOPS | 129.6 TFLOPS | ~+67% |
| RTX 3080 10 GB | — | — | +70–73% |
| RTX 3070 | — | — | +73–81% |
| RTX 3060 12 GB | — | — | ~+80% |
The project’s tested-card list now covers the RTX 3090 Ti, 3090, 3080 10 GB, 3080 20 GB memory-mod cards, 3070 and 3060 12 GB, all confirmed unlocked. One oddity: the OEM-only RTX 3080 Ti 20 GB (GA102-250) completes the unlock but stubbornly stays at half BF16 rate.
Power limits matter a lot here. The weakest reported gain — a modest +25% — came from a card capped at 225 W.
Where you'll actually see the difference
The gains show up in specific places. Local LLM users have reported 14–36% faster prompt prefill on long-context inference, and ComfyUI testing on an RTX 3090 Ti showed roughly 47–59% higher throughput across several BF16-heavy image-generation workloads. Diffusion models and FlashAttention also benefit.
What doesn’t improve much: LLM token generation, which is bound by memory bandwidth rather than compute, and INT8 work. The reports that compared outputs found them bit-identical before and after the unlock, which is reassuring for anyone worried about silent corruption.
As for gaming — Nvidia’s DLSS leans on Tensor Cores, so theoretically some AI-assisted rendering could benefit. In practice, nobody has published independently verified FPS numbers. Treat any gaming uplift as unproven for now.
Installation and risks
Setup is straightforward for anyone comfortable with Linux. Clone the repo, run make install, then add gpu_unlock.targets=all to your kernel command line (or specify individual PCI addresses). The module loads from the initramfs before Nvidia’s driver can grab the card, and it re-applies the unlock automatically after suspend-to-RAM — adding about a second to resume time. Hibernation isn’t covered.
It does nothing unless you explicitly opt in, and loading the module manually after boot won’t affect a card Nvidia has already bound. Full instructions and distro-specific notes are in the gpu-unlock GitHub repository.
Windows users are mostly out of luck. A separate experimental port has managed to boot an unlocked RTX 3090 into Windows on exactly one system, but that’s the extent of it.
The bigger caveat is risk. This is early-stage software that pokes at undocumented hardware registers on your daily driver. If you depend on your GPU for work or play, the sensible move is to wait for more compatibility reports — or try it on a spare card first. The project is licensed GPL-2.0-only, and Nvidia’s firmware components remain under Nvidia’s own terms.
Still, for anyone running ageing RTX 30-series hardware for AI workloads, the prospect of extracting professional-tier Tensor Core throughput for free is hard to ignore. If you’re feeling adventurous, the code is public and the community is actively logging results.