GPU Unlock Ada: Repository, Supported Evidence and Limits

The project is JustAResearcher/gpu-unlock-ada. Its README describes a Linux kernel module that applies an SM speed-select override before the stock NVIDIA driver starts. The current project description names RTX 4090 and experimental RTX 4070 Ti SUPER work; that is not blanket support for every Ada card.

What does the reported improvement measure?

The README reports roughly doubled FP16/BF16 tensor instruction throughput with FP32 accumulation in particular matched runs. It separately reports workload-dependent GEMM changes and a quantized model-serving example with no gain. The author explicitly distinguishes instruction rates from whole-application speedups and extended stability tests.

A “2×” number therefore needs its workload and measurement context. It does not establish twice the frame rate, twice the model-serving speed or improved scalar FP32 performance on your machine.

Read before trying a change

Use the repository's current Quick Start, compatibility information and evidence links. Match the exact device, driver, operating system and project revision. Check the recovery/rollback route before changing a machine you depend on; do not copy an isolated command from a benchmark screenshot.

For your own comparison, retain an unchanged baseline, the workload, correctness checks, power conditions and full results. A speed change without a correctness or stability check is an incomplete result.

What this page verifies

Repository documentation and its stated limits were reviewed October 30, 2026. The repository is marked GPL-2.0; firmware and other components have their own documented provenance and terms to inspect.

Varwo has not loaded the module, modified a GPU, generated firmware or reproduced a benchmark. This guide provides the source and an explanation of the evidence, not a tested installation procedure or a compatibility promise.