Open source project GPU Unlock can release the hidden AI computing power of Nvidia RTX 30 series graphics cards, and the actual performance can be improved by up to 60%.开源项目 GPU Unlock 可释放英伟达 RTX 30 系列显卡隐藏 AI 算力，实际性能提升可达 60%

IT Home October 25th news, Nvidia has been on the market for some time GeForce RTX 30 series (code name Ampere, Ampere) graphics cards, obviously there is a lot of computing potential yet to be unlocked. Zagitta, a GitHub user, has developed an open-source new project called GPU Unlock, which can lift the performance limitations of some artificial settings on ampere-based graphics cards. Under certain loads, the computational throughput of the tensor core can almost double, and there is no need to modify the Nvidia driver, and there is no need to permanently flash the VBIOS of the graphics card. IT House notes that this project is specifically aimed at an instruction transmission rate limitation: in FP32 cumulative FP16/BF16 tensor core operations, consumer-grade ampere graphics cards run at a lower rate than professional cards. By removing this limitation, the tool can release the extra computing power that the graphics hardware itself provides. Interestingly, GPU Unlock builds on earlier research on Nvidia CMP cryptocurrency mining graphics cards. It combines the firmware unlocking technology of the CMP project with a set of early initialization schemes, relying on a Linux kernel module to perform operations before the Nvidia graphics card driver loads. The module modifies the corresponding hardware speed selection register, performs a functional level reset (FLR), and then the original driver can initialize the graphics card normally. This also means that users no longer need to patch the driver every time they update the new version of the driver. Developers currently indicate that they support GA102, GA103, GA104, and GA106 GPU models, covering most of the models in the RTX 30 series. However, not every graphics card is necessarily compatible, and performance improvement is not guaranteed. The increase in performance can be considerable. According to the benchmark test data released by the project GitHub repository, one power consumption limit is set at 450W RTX 3090 Ti, and the results in the dense matrix multiplication (GEMM) test using FP32 accumulation are as follows: This kind of improvement does not only appear on RTX 3090 Ti. A RTX 3080 Ti user reported that BF16 throughput increased from 71.59 TFLOPS to 113.98 TFLOPS and TF32 performance increased from 35.65 TFLOPS to 65.28 TFLOPS with a graphics card power consumption limit of 315W. There is also a RTX 3080 10GB compatibility report showing that FP16/BF16 performance increased from approximately 65 TFLOPS to 110–113 TFLOPS under the original 320W power limit for graphics cards. These test results are also widely discussed on Platform X. And how does it perform in practice? Fortunately, community users have begun to test this unlocking function under real AI load, and the preliminary results are relatively optimistic. The last RTX 3090 Ti user on Reddit ran a score test in ComfyUI, increasing MOSS-Image SFT BF16 throughput by 58.6%, Qwen-Image 2.1 Turbo BF16 by 57.2%, and Z-Image Turbo BF16 by 56.1%. The increase in INT8 precision mode is very small, and the increase in MOSS-Image SFT INT8 is only 0.3%. The tester also said that the default state and the decoding result of the image generated after unlocking are exactly the same. Large language model (LLM) reasoning is also a potential benefit scenario, but the height of the boost depends on the type of load. For example, a compatibility report of RTX 3090 showed that the prompt processing and inference throughput of vLLM increased by 33%, but the generation speed of word elements (tokens) remained basically unchanged. It is also well understood that the bottleneck in the decoding of large language models is often the graphics memory bandwidth, rather than the raw computational throughput of the tensor core. Another noteworthy development is the experimental Windows implementation, which no longer uses the original Linux kernel modules, but is based on the UEFI approach. Developers documented the success of the RTX 3090, which improved BF16 computational throughput by approximately 82.5%. However, this version is still experimental, only suitable for specific hardware, and is not the official version of Windows that is generally available. There are also significant limitations to this option. GPU Unlock will not increase the bandwidth of the graphics memory, will not unlock more CUDA cores, and will not automatically speed up all loads. At least one RTX 3080 Ti 20GB feedback showed that the unlocking process was successful, but the BF16 performance did not improve as expected, indicating that the effect is not universal. In terms of games, Nvidia's DLSS technology will also use tensor cores, but there is no reliable evidence that GPU Unlock can increase game frame rate or reduce DLSS operation time. Benefits are only possible if the task itself invokes these affected commands and accumulation patterns. Therefore, these dazzling calculations cannot be equated with the simultaneous increase in the number of frames in the game. Even so, GPU Unlock is an attractive new endeavor for RTX 30 series graphics card holders, especially those who use graphics cards for local AI inference, image generation, and other high computing load scenarios. This project is still in the experimental stage, but it is really attractive to be able to extract considerable additional performance from existing graphics cards without permanently changing the hardware.