IT Home On October 25, NVIDIA has launched the GeForce RTX 30 series (codename Ampere) graphics cards after a while, and there is clearly still a lot of computational potential that hasn’t been released yet. GitHub user Zagitta developed an open-source project called GPU Unlock, which can remove some artificially set performance restrictions on Ampere architecture graphics cards. Under specific loads, the tensor core computing throughput can almost double, and neither NVIDIA drivers need to be modified nor the VBIOS of the graphics card permanently rewritten.
IT Home noticed that this project specifically addresses a command transmission rate limit: in FP16/BF16 tensor core operations using FP32 accumulation, consumer-grade ampere graphics cards have a lower operating speed compared to professional cards. After removing this limit, the tool can leverage the additional computing performance that the graphics card hardware inherently has.
Interestingly, GPU Unlock is based on earlier research related to NVIDIA CMP cryptocurrency mining graphics cards. It combines the firmware unlocking technology of the CMP project with a pre-initialization scheme, and relies on a Linux kernel module to perform operations before the NVIDIA graphics card driver loads. This module modifies the corresponding hardware speed selection registers and executes a functional level reset (FLR), after which the original driver can initialize the graphics card properly.
This also means that when users update the driver to a new version, they no longer need to patch the driver. The developers currently indicate support for the GA102, GA103, GA104, and GA106 model GPUs, which cover most of the RTX 30 series models. However, not every graphics card will be compatible, and it is not guaranteed that performance improvements will be achieved.
The improvement in performance can be quite significant. According to the benchmark test data released on the project’s GitHub repository, for an RTX 3090 Ti with a maximum power consumption of 450W, the results in the dense matrix multiplication (GEMM) test using FP32 accumulation are as follows:
These improvements are not limited to the RTX 3090 Ti. A user of the RTX 3080 Ti reported that, under a graphics card power limit of 315W, the BF16 throughput increased from 71.59 TFLOPS to 113.98 TFLOPS, and the TF32 performance increased from 35.65 TFLOPS to 65.28 TFLOPS. Another compatibility report for the RTX 3080 10GB shows that, under the original 320W power limit of the graphics card, the FP16/BF16 performance increased from approximately 65 TFLOPS to 110–113 TFLOPS. These test results have also been widely discussed on the X platform.
How does it perform in actual applications? Fortunately, community users have already started testing this unlocking feature under real AI workloads, and the initial results are quite optimistic.
The last Reddit user with an RTX 3090 Ti conducted benchmark tests in ComfyUI. The MOSS-Image SFT BF16 throughput increased by 58.6%, Qwen-Image 2.1 Turbo BF16 increased by 57.2%, and Z-Image Turbo BF16 increased by 56.1%. However, the improvement in INT8 precision mode was very small—MOSS-Image SFT INT8 only improved by 0.3%. The tester also said that the image decoding results are exactly the same in both the default state and after unlocking.
Large language model (LLM) inference is also a potential beneficiary scenario, but the improvement depends heavily on the type of load. For example, a compatibility report for an RTX 3090 shows that vLLM improves prompt processing and inference throughput by 33%, but the token generation speed remains essentially unchanged. This makes sense; the bottleneck in the decoding stage of large language models is often the VRAM bandwidth, rather than the raw computational throughput of the tensor core.
Another noteworthy advancement is the implementation scheme for the beta version of Windows, which no longer uses the original Linux kernel modules but is based on UEFI. Developers have recorded a successful case with the RTX 3090, and the BF16 computing throughput increased by approximately 82.5%. However, this version remains experimental in nature and is only compatible with specific hardware; it is not a general-use official Windows version.
This solution also has obvious limitations. GPU Unlock does not increase memory bandwidth, does not unlock more CUDA cores, and cannot automatically speed up all workloads. At least some feedback for the RTX 3080 Ti 20GB indicates that the unlock process was successful, but BF16 performance did not improve as expected, suggesting that the effect is not universal.
In terms of games, NVIDIA’s DLSS technology also utilizes the Tensor Core, but there is currently no reliable evidence that GPU Unlock can increase game frame rates or reduce the runtime of DLSS. Benefits may only be achieved when the task itself calls these affected instructions and accumulation modes. Therefore, these impressive computing scores cannot be equated with a simultaneous increase in game frames.
Even so, for those who own the RTX 30 series graphics cards, GPU Unlock remains a very attractive new initiative, especially for users who use the cards for local AI inference, image generation, and other high-computing-load scenarios. This project is still in the experimental stage, but it offers significant additional performance from existing graphics cards without any permanent hardware changes, which is indeed very attractive.