GitHub user Zagitta has developed the open-source tool gpu-unlock, which removes the instruction issue rate limit for certain Tensor Core computations on Ampere-architecture GeForce graphics cards. Community tests show that the computational throughput of the RTX 3090 increased by nearly 80% in some cases, although the actual gain varies depending on the task.
The developer states that FP16 and BF16 operations using FP32 accumulation run at about half the speed on relevant consumer-grade graphics cards compared to Nvidia's professional GPUs. gpu-unlock removes this limitation by changing the hardware speed selection register.
This solution follows previous research on unlocking the capabilities of Nvidia CMP mining cards. The Linux kernel module modifies registers and resets the GPU before the Nvidia graphics driver loads, then hands control over to the unmodified official driver, thus eliminating the need to permanently flash the graphics card's VBIOS or rely on modified firmware or drivers.
Compatibility reports show that an RTX 3090's BF16 matrix multiplication performance rose from 77.5 to 129.6 TFLOPS, and TF32 performance increased from 40.2 to 69.8 TFLOPS; the RTX 3090 Ti's BF16 throughput jumped from 87.5 to 157.4 TFLOPS.
Results are inconsistent across different workloads. Another RTX 3090 user testing local large language model inference saw a 16% increase in long-context prompt processing speed, with virtually no change in token generation speed. When running ComfyUI on the RTX 3090 Ti, throughput for several BF16-heavy image generation workloads improved by approximately 47% to 59%.
DLSS also utilizes Tensor Cores, but there are currently no independently verified game benchmarks proving that this tool can increase frame rates. The original project currently only supports Linux, and the experimental Windows port has only successfully launched on a specific system configuration. This unlock method remains an experimental solution and poses risks when used on a primary gaming PC.
Detailed report by foreign media:
Nvidia's RTX 3090 graphics card, released six years ago, seems to have some untapped computational potential. It is claimed that a new open-source project called gpu-unlock, developed by GitHub user Zagitta, can achieve double-digit performance improvements in certain Tensor Core workloads without requiring any hardware modifications or installing patched Nvidia drivers.
This project aims to address the instruction issue rate limitation for FP16 and BF16 Tensor Core operations (with FP32 accumulation) on Ampere-based GeForce graphics cards. According to the developer, affected operations run at approximately half the speed of professional-grade Nvidia GPUs. This unlock solution removes the restriction by modifying the hardware speed selection register.
Interestingly, this technology builds upon previous research used to unlock the performance of Nvidia CMP mining GPUs. Zagitta's implementation uses a Linux kernel module that executes before the Nvidia graphics driver loads, modifies relevant registers and resets the GPU, then hands control back to the unmodified driver. Therefore, it does not require permanently flashing the graphics card's VBIOS or modifying firmware.
In a community compatibility report, the RTX 3090 achieved 129.6 TFLOPS in BF16 matrix multiplication, an improvement over the previous 77.5 TFLOPS. TF32 performance also increased from 40.2 TFLOPS to 69.8 TFLOPS. Its sibling product, the RTX 3090 Ti, saw its BF16 throughput jump from 87.5 TFLOPS to 157.4 TFLOPS.
In other words, performance gains largely depend on the workload. Another RTX 3090 user reported that during local LLM inference, processing speed for long-context prompts improved by 16%, while token generation speed showed almost no change. Additionally, tests of ComfyUI on the RTX 3090 Ti indicated that throughput increased by approximately 47% to 59% in several BF16-intensive image generation workloads.
So what about gaming? Nvidia's DLSS technology also uses Tensor Cores, which means certain AI-assisted rendering operations might benefit from this. Unfortunately, there are currently no independently verified game benchmarks confirming any increase in frame rates. Currently, the project is limited to Linux systems, although an independent experimental Windows port has demonstrated that an unlocked RTX 3090 graphics card can successfully boot Windows on specific systems.
This unlocking technique is currently still in an experimental stage, so installing it on a primary gaming PC carries risks. However, if you enjoy taking risks, the possibility of significantly boosting computational performance from older RTX 30 series hardware without buying a new graphics card is indeed appealing.