Open-Source 'GPU Unlock' Removes RTX 30-Series AI Compute Throttling

IT Home, October 25 news — NVIDIA's GeForce RTX 30 series (codenamed Ampere) graphics cards, which have been on the market for some time, clearly still have a significant amount of untapped computing potential. GitHub user Zagitta has developed a new open-source project called GPU Unlock, which removes certain artificial performance limitations on Ampere architecture GPUs. Under specific workloads, it can nearly double Tensor Core compute throughput without requiring modifications to NVIDIA drivers or permanently flashing the card's VBIOS.

IT Home noted that this project specifically targets an instruction issue rate limitation: in FP16/BF16 Tensor Core operations using FP32 accumulation, consumer-grade Ampere graphics cards operate at a lower rate compared to professional cards. By removing this restriction, the tool unlocks the extra compute performance inherent to the GPU hardware itself.

Interestingly, GPU Unlock builds upon earlier research into NVIDIA's CMP cryptocurrency mining GPUs. It combines firmware unlocking techniques from CMP projects with an early initialization scheme, relying on a Linux kernel module that executes before the NVIDIA graphics driver loads. This module modifies the corresponding hardware speed selection registers and performs a Function Level Reset (FLR), after which the stock driver can initialize the graphics card normally.

This also means users do not need to re-patch the driver every time a new version is updated. The developer currently lists support for GA102, GA103, GA104, and GA106 GPUs, covering most models in the RTX 30 series; however, compatibility is not guaranteed for every single card, nor is a performance boost assured in all cases.

The magnitude of the performance gain can be quite substantial. According to benchmark data published in the project's GitHub repository, an RTX 3090 Ti with a power limit set to 450W showed the following results in dense general matrix multiplication (GEMM) tests using FP32 accumulation:

Such improvements are not limited to the RTX 3090 Ti. An RTX 3080 Ti user reported that under a 315W power limit, BF16 throughput increased from 71.59 TFLOPS to 113.98 TFLOPS, and TF32 performance rose from 35.65 TFLOPS to 65.28 TFLOPS. Another compatibility report for an RTX 3080 10GB showed that under the factory 320W power limit, FP16/BF16 performance increased from roughly 65 TFLOPS to 110–113 TFLOPS. These test results have also been widely discussed on the X platform.

So how does it perform in real-world applications? Fortunately, community members have already begun testing this unlocking feature under real AI workloads, and the preliminary results are relatively promising.

A Reddit user with an RTX 3090 Ti conducted benchmark tests in ComfyUI, observing a 58.6% throughput increase in MOSS-Image SFT BF16, a 57.2% increase in Qwen-Image 2.1 Turbo BF16, and a 56.1% increase in Z-Image Turbo BF16. In contrast, gains in INT8 precision mode were minimal, with MOSS-Image SFT INT8 improving by only 0.3%. The tester also stated that the decoded image outputs between the default and unlocked states were completely identical.

Large language model (LLM) inference is another scenario that could potentially benefit, though the extent of improvement is highly dependent on the workload type. For instance, a compatibility report for an RTX 3090 showed a 33% increase in vLLM prompt processing and inference throughput, but token generation speed remained essentially unchanged. This is understandable, as the bottleneck in the LLM decoding phase is typically VRAM bandwidth rather than the raw compute throughput of the Tensor Cores.

Another noteworthy development is an experimental Windows implementation that bypasses the original Linux kernel module in favor of a UEFI-based approach. The developer documented a successful case with an RTX 3090, achieving an approximately 82.5% increase in BF16 compute throughput. However, this version remains experimental, is tailored only to specific hardware, and is not a universally usable official Windows release.

The approach also comes with distinct limitations. GPU Unlock does not increase VRAM bandwidth, unlock more CUDA cores, or automatically speed up all workloads. At least one report regarding an RTX 3080 Ti 20GB indicated that while the unlock process executed successfully, the expected BF16 performance boost did not materialize, demonstrating that the effect is not universal.

In terms of gaming, NVIDIA's DLSS technology also utilizes Tensor Cores, but there is currently no reliable evidence showing that GPU Unlock can improve gaming frame rates or reduce DLSS execution time. Gains are only possible when the task itself invokes the affected instructions and accumulation modes. Therefore, these impressive synthetic compute benchmarks do not equate to a simultaneous increase in gaming frame rates.

Even so, for RTX 30 series owners, GPU Unlock remains an attractive new endeavor—especially for those who use their graphics cards for local AI inference, image generation, and other compute-heavy workloads. While the project is still in its experimental stages, being able to squeeze substantial extra performance out of an existing GPU without permanently modifying the hardware is undeniably appealing.