AnswersPublished24 sources
What Zagitta’s gpu-unlock changes on RTX 30 GPUs—and what it doesn’t
Zagitta’s Linux only gpu unlock targets selected FP16/BF16 Tensor Core operations on Ampere GPUs. The module applies hardware speed selector changes before NVIDIA’s unmodified driver binds.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Zagitta’s Linux only gpu unlock targets selected FP16/BF16 Tensor Core operations on Ampere GPUs.
- The module applies hardware speed selector changes before NVIDIA’s unmodified driver binds.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is Zagitta’s free, open-source gpu-unlock tool, how does its Linux kernel module raise FP16 and BF16 multiply-accumulate instruction is. Article summary: Zagitta’s gpu-unlock is an open-source Linux kernel module intended to remove selected Tensor Core instruction-issue limits on Ampere GeForce GPUs, bringing those issue rates—not overall GPU performance—to professional-c. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
Zagitta’s open-source gpu-unlock is a Linux kernel module for selected Ampere GeForce GPUs. It targets FP16 and BF16 Tensor Core operations that accumulate into FP32, and its developer says it restores their instruction-issue rates to professional-card levels. The benchmark evidence shows substantial gains in some matrix workloads—but not a uniform speedup for every AI application. 1
8
How gpu-unlock works
The project targets GPUs based on GA102, GA103, GA104 and GA106. It loads from the initramfs before NVIDIA’s stock driver binds to the card, writes hardware speed selectors, and performs a function-level reset (FLR) before handing control to the unmodified driver. The repository identifies SS0 0x88888888 and the four-bit SS1 0x8 IMLA4 selector among the settings it applies. 1
The approach combines the V67/PLM technique used in earlier CMP mining-card work with the early, pre-driver timing of a UEFI approach. It is a Linux kernel module—not a patched NVIDIA driver or a UEFI application—and the described method does not require permanent VBIOS flashing or a hardware modification. 1
4
What the benchmark numbers show
The project’s RTX 3090 Ti results are for dense 8192³ GEMM with FP32 accumulation, a 450 W power limit and five-sample medians. They are developer-reported measurements, not a standardized independent test suite. 1
| RTX 3090 Ti test | Stock | Unlocked | Reported change |
|---|---|---|---|
| FP16 | 87.46 TFLOPS | 118.39 TFLOPS | About +35% |
| BF16 | 87.50 TFLOPS | 157.41 TFLOPS | About +80% |
| TF32 | 42.58 TFLOPS | 82.61 TFLOPS | About +94% |
A separate RTX 3090 user reported FP16 rising from 82 to 126 TFLOPS (+54%), BF16 from 82 to 129 TFLOPS (+57%), and TF32 from 41 to 74 TFLOPS (+80%). These are individual community results, so they should not be treated as guaranteed gains for every card or setup. 13
One stock EVGA RTX 3090 compatibility report describes operation on Ubuntu 24.04, with a 350 W power cap, and roughly twice the FP16/BF16 throughput per MHz. That is a clock-normalized result; it does not establish twice the absolute throughput under the power limit. 30
Why AI application gains can be smaller
A matrix benchmark isolates the type of work gpu-unlock targets. An application may spend only part of its time on those operations, so its overall improvement can be lower. The reported results illustrate that difference:
- On a stock RTX 3090, one vLLM report found about 33% faster prompt processing and essentially unchanged token generation. The same report described eight hours of mixed-workload testing.
10
- A separate report on a power-limited 230 W RTX 3090 found prompt-processing gains of about 22–30% and token-generation gains of about 1–5%. Its results come from a different workload and setup, so they are not directly comparable with the vLLM report.
43
- BF16-heavy ComfyUI graphs on one RTX 3090 Ti were reported to run 31–60% faster, while INT8 performance was essentially unchanged.
19
The practical result depends on whether a workload uses the affected precision and instructions, as well as factors such as its kernels, matrix sizes, batching and power limit. A higher instruction-issue ceiling can produce a large GEMM gain without translating into the same uplift for end-to-end inference. 1
10
19
43
Release status, compatibility and caveats
The repository was public by October 18, 2026. A third-party tracker’s historical figures show about 2,200 stars on October 24. The supplied activity listing also shows commit 4189b17 reverting the branch to a working state at c274e74, a commit associated with reapplying the unlock after suspend. These details are snapshots, not assurances about the repository’s current state. 19
28
42
44
The module is Linux-specific. A separate experimental UEFI-based Windows approach has reported a BF16 improvement on a particular hardware setup, but that does not establish broad Windows compatibility or a supported Windows version of Zagitta’s module. 20
Installation involves kernel and initramfs configuration and a GPU reset at boot. Compatibility reports include distribution-specific setup issues, and the project’s development history includes a rollback. Anyone considering it should check current compatibility information and have a recovery path; brief success reports do not establish long-term reliability across cards, kernels, drivers or suspend-and-resume configurations. 6
42
44
The evidence supports a narrow conclusion: gpu-unlock can substantially accelerate selected Tensor Core matrix workloads, with the project reporting an approximately 80% BF16 gain on its RTX 3090 Ti test. The available results do not establish a comparable improvement across all AI inference—and they do not verify a DLSS or gaming-performance benefit. 1
4
45
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What Zagitta’s gpu-unlock changes on RTX 30 GPUs—and what it doesn’t"?
Zagitta’s Linux only gpu unlock targets selected FP16/BF16 Tensor Core operations on Ampere GPUs.
What are the key points to validate first?
Zagitta’s Linux only gpu unlock targets selected FP16/BF16 Tensor Core operations on Ampere GPUs. The module applies hardware speed selector changes before NVIDIA’s unmodified driver binds.
Sources
3dmgame.com开源工具gpu-unlock解除RTX 30系限制，部分计算性能提升近80%_3DM单机
x.comDavid Y. Tan on X: "Tried gpu-unlock on an RTX 3090 (cold ...