GPU Unlock for RTX 30: Linux AI Benchmarks and Limits

GPU unlock for RTX 30 is a Linux kernel module from Zagitta that changes the throughput of specific floating-point operations before NVIDIA’s driver initializes the GPU. Community tests show substantial matrix-compute gains, but those gains do not translate into a universal doubling of AI performance.

Checked: 25 October 2026. This is a source-based analysis, not a hands-on installation review. The Signal has not reproduced these hardware measurements. Compatibility and installation details may change as the project develops.

What does gpu-unlock change?

The project documentation describes an early-boot module that adjusts speed selectors and resets the card before the unmodified NVIDIA driver binds it. The targeted path is FP16/BF16 matrix arithmetic accumulated into FP32. It does not require flashing the card’s vBIOS.

The developer says this avoids manually patching every NVIDIA driver release. That is a maintenance advantage, not a guarantee of compatibility with every future kernel or driver.

“Professional-equivalent” throughput in a particular instruction path should not be read as a complete workstation-card conversion. The useful question is whether your workload spends enough time in the affected operations. For the wider hardware and runtime context, see our local AI guide.

RTX 3090 Ti benchmarks: a large compute gain

Community report #1 gives the following results for dense 8192³ GEMM at a 450 W power limit, using FP32 accumulation and the median of five samples. The percentage changes below are calculated from the reported values.

These are matrix-multiplication measurements under a defined power budget. They are not image-generation speeds, gaming frame rates or LLM output-token rates. The different gains across precisions also show why one headline percentage cannot describe the whole card.

Supported chips versus tested cards

The repository lists GA102, GA103, GA104 and GA106 support. Its tested-card summary distinguishes successful reports from untested configurations and exceptions:

Check the exact device ID and current compatibility reports rather than relying only on the RTX product name. This article covers Zagitta’s Linux module; it does not establish RTX 40 compatibility.

LLM prefill and decode need separate measurements

Prefill processes the input prompt; decode generates the answer token by token. A faster prompt-processing phase can make a long-document request feel more responsive without increasing the speed at which the answer streams.

In an RTX 3090 report on Debian, prompt throughput at 16K context rose from 2,594 to 3,009 tokens/s. Time to first token fell from 6.35 to 5.48 seconds, while decode at that context stayed near 97 tokens/s. The same report’s 4K prompt test was essentially unchanged. These are results for that configuration, not a general promise.

Quantization and kernel selection matter too. A separate RTX 3080 report found essentially unchanged prompt and generation rates in its quantized llama.cpp tests. A gain in floating-point GEMM therefore does not establish a gain for a particular GGUF model.

Does it help ComfyUI image and video generation?

There is already a useful real-workflow data point. In report #14, two INT4 diffusion workloads in ComfyUI had warm mean generation times fall from 9.30 to 7.98 seconds and from 9.34 to 8.17 seconds: reductions of 14.2% and 12.5%. The report describes three warm seeds, 864 × 1536 output, eight steps and PyTorch cross-attention.

Those modest end-to-end improvements are more informative than assuming the matrix benchmark applies to every stage. Model loading, text encoding, offloading, sampling and output decoding can occupy different proportions of a workflow.

Another RTX 3090 report measured roughly 53–54% higher FlashAttention 2 throughput in its tested shapes. It also found that a previously useful custom attention optimization became slower than FlashAttention 2 after unlocking. Backend choices may need to be measured again.

There is no Wan or MiniMax H3 speedup claim in this article. We have not verified a controlled before-and-after test for those workflows. Start with our ComfyUI installation guide for the application setup, then measure your actual graph separately.

The 381 tokens/s Qwen result is a different project

The widely shared Qwen3.8-27B result should not be attributed to gpu-unlock. In the developer’s original post, the optimized inference stack combines DFlash2 drafting and lookup-based proposals. It reports around 133 tokens/s on real chat prompts and around 382 when reproducing text from the supplied context.

This is a different acceleration mechanism and a highly specific workload. Reusing prompt text can make speculative proposals unusually easy to accept. Neither that result nor the GEMM table predicts ordinary chat speed on an arbitrary model.

Installation scope and operational limits

This is a system-level Linux modification, not a ComfyUI custom node. The documented process builds and installs a kernel module, includes it in the early boot environment and explicitly selects the target GPU. The module must run before NVIDIA initializes the card; loading it after the driver has bound the device is insufficient.

Read the current upstream instructions for your distribution and bootloader. Before experimenting, keep a baseline, record the project commit, and retain a working boot entry. Community reports document unsigned-module and early-boot loading issues; the RTX 3080 discussion includes a fix for a systemd/mkinitcpio loading problem. Do not apply an old workaround without checking the version you are installing.

The documented rollback is to remove the target option from the kernel command line; package removal also removes the module and its initramfs additions. We have not tested that procedure on a physical RTX system.

How to evaluate the benefit on your own workload

1. Record the environment: card, power limit, driver, kernel, runtime, model precision and attention backend.
2. Keep the workload fixed: same model files, prompt, seed, resolution, steps and context length.
3. Separate cold and warm runs: loading and compilation can hide changes in steady-state execution.
4. Measure what users wait for: seconds per image or clip; for LLMs, prefill throughput, time to first token and decode throughput separately.
5. Check correctness and stability: compare outputs where deterministic execution permits it, and inspect temperatures and driver errors during repeated runs.

Remember that throughput and elapsed time use different denominators: an 80% throughput increase would reduce time by about 44% for the same isolated work, not by 80%. A complete pipeline usually gains less when other stages remain unchanged.

Should this change a local-versus-cloud decision?

For an existing Linux RTX 30 owner, the evidence makes this project worth evaluating on a spare or recoverable setup. For a hardware purchase, the sensible comparison remains usable VRAM, measured workload performance, power, condition, price and maintenance effort.

We have not independently verified completed-sale market prices or established that gpu-unlock caused any price increase. Do not build a purchase decision around viral asking prices. Our Comfy Cloud versus local GPU comparison explains the broader trade-offs.

Sources and verification

Primary evidence: gpu-unlock repository and compatibility summary; RTX 3090 Ti measurements; RTX 3090 LLM report; RTX 3080 compute, ComfyUI and llama.cpp report; RTX 3090 attention report; and the separate Qwen inference developer report. These are community measurements, with different methods and environments, rather than a standardized independent review.