RTX 20, 30 and 40: modders unlock hidden AI power

The gpu-unlock, gpu-unlock-ada and Turing-unlock projects target a throughput limitation affecting certain Tensor instructions on GeForce cards. The best-documented cases involve FP16 or BF16 calculations with FP32 accumulation, as well as TF32 on compatible architectures. Accumulation refers to the precision used to add intermediate results: two calculations using FP16 data therefore do not necessarily follow the same hardware path.

The unlock takes place before the NVIDIA driver loads and changes internal GPU settings. It adds neither cores nor video memory and does not require flashing the VBIOS. Its effect is volatile: cutting power to the GPU can erase it. The projects build on work carried out on CMP mining cards, adapted notably by Zagitta and, for Ada, drawing on TOTORONG's work.

The results below come from the developers and their contributors. Hardware Reference cross-checked the publications and test methods; we have not reproduced these tests on our own test bench.

The repository gpu-unlock by Zagitta collects the following dense matrix multiplication measurements with FP32 accumulation. The figures are rounded in its summary; the percentages are those stated by the project from the original measurements.

These reports use different configurations: they show the before-and-after gain on each card and are not a ranking produced under a shared test protocol. TFLOPS here express compute throughput, not frames per second in a game.

The matched before-and-after report from gpu-unlock-ada is particularly informative. The same RTX 4090 was compared before and after activation, using NVIDIA driver 615.78.08, CUDA 13.4.92 and a 450 W limit. Instruction throughput rises from 160.4 to 320.4 TFLOPS in FP16/FP32, and from 159.7 to 320.2 TFLOPS in BF16/FP32.

Enlarge

The cuBLAS matrix multiplications in the same comparison show more modest changes. All use FP32 accumulation; N denotes the size of square N × N matrices. The values are the published medians, and the changes below are calculated from those values.

The BF16 case at 4 096³ therefore remains virtually unchanged. The report covers just one pair of measurements, with the desktop active, and does not establish long-term stability. FP16 with FP16 accumulation and the INT8 control tests also remain close to their original levels.

The repository Turing-unlock documents a test on an ASUS RTX 2080 Ti with 11 GiB, carried out on 22 October. FP16 instruction throughput with FP32 accumulation rises from 46.9361 to 93.6649 TFLOPS, almost doubling. In cuBLAS matrix multiplication at 8 192³, throughput increases from 48.1488 to 85.1705 TFLOPS, or about 77%. FP16 with FP16 accumulation remains unchanged.

This demonstration concerns one precisely identified card and platform. Its experimental loader is not a universal installer. It also shows that unlocking compute throughput does not require increasing memory capacity: the card's 11 GiB remained 11 GiB.

The test by Cryptolab provides a useful counterpoint. It compares 18 requests in total using Qwen3.8 27B Q4_K_M and MTP in Ollama 0.33.2, on a 24 GiB RTX 3090 limited to 350 W. Each profile includes three runs per state, with an output of 256 tokens. The throughput figures below are medians, in tokens per second.

Interpret these results with care: clock speeds were not locked and thermal conditions differed between the test series. In particular, the measured median SM clock rose from 1 470 to 1 545 MHz. These differences describe the observed performance without isolating the effect of the unlock alone. The paired responses nevertheless have identical SHA-256 hashes.

Diffusion workloads can benefit more from the calculations affected. A contributor to nv-uefi-unlocker reports the first two results below on an RTX 3080 LHR running Windows, with three runs per state. A separate report on an RTX 4090 measures LTX-2.5 in ComfyUI. Each row compares the same workload before and after the unlock; the machines and models differ between rows.

The percentage indicates the throughput increase for the same amount of work, calculated as time before / time after − 1. A 28.8% throughput increase therefore corresponds to about 22.3% less time. On the RTX 4090, the benefit changes with the power limit; unlocking alone does not remove that constraint.

An LLM first processes the prompt, then generates its response token by token. The first phase can involve many matrix multiplications. The second is often limited by memory transfers, especially for a single request. Speeding up Tensor Cores does not resolve that second bottleneck.

A report published on 31 October using an RTX 3070 Ti illustrates this difference with Qwen3-1.7B BF16 in llama.cpp: processing 4 096 tokens rises from 9 382 to 12 895 tokens/s, about +37%, while generation stays at 141.8 and then 141.7 tokens/s. The same model in Q8_0 gains only around 2% in prompt processing.

In the LocalLLaMA discussion linked around the project, a user with a memory-modified RTX 3080 also reports no gain with their Q6_K model. This account does not replace a full test protocol, but highlights an essential point: the weight format alone cannot predict the gain. You need to know which operations the inference engine actually uses.

INT8 and conventional CUDA FP32 compute do not double in the published control tests. FP8 needs a more nuanced view: the Ada repository describes gains in certain FP8 paths with FP32 accumulation, but also tests that remain virtually unchanged. It would therefore be misleading to present all INT8 or FP8 models as automatically accelerated. These results do not establish any gain in gaming FPS.

The new development covered by VideoCardz is nv-uefi-unlocker, developed by physicsdolphin. This tool runs from a USB drive before the operating system, applies the procedure and then hands control back to the Windows or Linux bootloader. It avoids having to repeat a long sequence manually for every test, but still requires preparation and verification on the target machine.

By default, its startup screen asks for confirmation. The documentation provides an explicit automated mode and instructs users to repeat the operation after a cold boot. A test on an RTX 4060 Laptop measures an increase from 29.89 to 49.02–49.33 TFLOPS in FP16 GEMM at 4 096, or around +65%, from this unlock alone.

Source: detailed compatibility matrix, accessed on 31 October 2026.

The standard UEFI procedure requires Secure Boot to be disabled, or an EFI application signed with a trusted key. Laptops may lose the unlock when the dedicated GPU is powered off. The authors also report boot failures on some platforms: automation does not turn an experimental modification into a feature officially supported by NVIDIA.

Mia's post of 24 October, shown below, shared the first Ampere results, then mainly focused on Linux. The projects have since evolved, notably with the UEFI tool.