Science & Tech · Gadgets & EVs · 5 hrs ago
Free Linux Tool Unlocks More AI Compute on Some Nvidia RTX 30 Cards
GPU Unlock removes a software limit on some Nvidia graphics cards, allowing them to do more AI-related calculations.
The open-source project targets Nvidia's Ampere-based GeForce cards, which include many models in the RTX 30 series.
On an RTX 3090 Ti, the developer measured BF16 performance rising from 87.5 to 157.4 trillion operations per second, while TF32 performance nearly doubled.
In practical tests, AI workloads improved by varying amounts: image generation rose about 47% to 59%, and some language-model processing tasks gained 14% to 36%.
The tool runs as a Linux kernel module before the graphics driver loads, so it does not require a modified driver or a permanent firmware change.
Gaming and token decoding saw little benefit, and results vary by card and workload; RTX 40 and RTX 50 cards are not supported, while a Windows version is still experimental.
GPU Unlock is a free open-source project that removes a software throttle on some Nvidia Ampere-based GeForce cards.
On an RTX 3090 Ti, its published BF16 matrix-multiply result rose from 87.5 to 157.4 TFLOPS.
Reported real-world gains range from 14% to 36% in vLLM prefill and about 47% to 59% in ComfyUI image generation.
The tool runs as a Linux kernel module before the graphics driver loads and does not require a VBIOS flash or driver patch.
RTX 40 and RTX 50 cards are not supported, and compatibility and performance gains vary by card and workload.
- Who
- GPU Unlock, an open-source project published by a developer using the name Zagitta.
- What
- The tool removes a software throttle on FP16 and BF16 Tensor Core operations in some Nvidia GeForce cards.
- When
- The reports were published on October 24, 2026.
- Where
- On Linux, for supported Nvidia graphics cards.
- Why
- To increase AI compute performance by lifting the throttle on certain Tensor Core operations.
This story does not have two clearly opposing sides.
This story does not have a timeline yet.
- Project
- GPU Unlock
- Target cards
- Some Ampere-based GeForce cards, including RTX 30 models
- RTX 3090 Ti BF16 result
- 87.5 to 157.4 TFLOPS
- vLLM prefill gains
- 14% to 36%
- ComfyUI image-generation gains
- About 47% to 59%
- Unsupported generations
- RTX 40 and RTX 50