GPU Unlock lifts RTX 30 tensor core throughput in Linux test
An open-source Linux kernel module raises tensor core throughput on Nvidia GeForce RTX 30 cards, with one RTX 3090 Ti test showing about 60% faster video generation.
GitHub developer Zagitta has released an open-source tool called GPU Unlock for Ampere-based Nvidia GeForce RTX 30 series graphics cards, according to aitimes.com. The tool raises the throughput of certain tensor core operations that are set lower on consumer products to the levels of professional products.
The targeted operations compute in FP16 or BF16 and accumulate results in FP32. According to the developer, instruction throughput for these operations is doubled.
GPU Unlock works as a Linux kernel module that changes hardware settings before the Nvidia driver initializes the GPU. It does not require modifying the driver or permanently changing the graphics card firmware.
Measured gains
In the developer's matrix multiplication test, RTX 3090 Ti BF16 performance rose from 87.5 TFLOPS to 157.4 TFLOPS, about 80%, with the power limit set to 450W. FP16 performance improved about 35% and TF32 about 94% in the same test. Improvements were also reported on RTX 3080, RTX 3070 and RTX 3060, with varying gains depending on the card and power settings.
An RTX 3090 Ti user's ComfyUI test showed MOSS-Image SFT in BF16 generating one image in about 45 seconds instead of about 72 seconds, a roughly 37% reduction in time and about 59% more images over the same period. Throughput rose about 57% for Qwen-Image 2.1 Turbo and about 56% for Z-Image Turbo, while MiniMax H3 video generation time fell from about 253 seconds to 158 seconds. MOSS-Image SFT run in INT8 showed only a 0.3% throughput increase, and a separate vLLM test on RTX 3090 showed prompt processing about 33% faster with almost no change in token-by-token generation.
The current tool targets Linux, and results are not guaranteed on all RTX 30 cards; some cards showed no expected gain after unlocking. The tool does not increase graphics memory capacity or memory bandwidth.
Quick answers
What does GPU Unlock do?
It raises the throughput of certain FP16/BF16 tensor core operations on Nvidia GeForce RTX 30 series cards to professional product levels, according to the developer.
Does GPU Unlock work on Windows?
No. The current tool targets Linux and runs as a kernel module that changes hardware settings before the Nvidia driver initializes the GPU.
Does GPU Unlock add graphics memory?
No. It does not increase graphics memory capacity or memory bandwidth.