GPU Unlock mods boost RTX 30 and 40 series AI compute performance

Two new open-source Linux kernel modules are bypassing software restrictions on NVIDIA GeForce GPUs to enable professional-grade compute performance. Zagitta developed GPU Unlock for Ampere-based RTX 30-series cards, specifically targeting GA102, GA103, GA104, and GA106 GPUs. This tool increases instruction-issue rates for FP16 and BF16 operations with FP32 accumulation.

On an RTX 3090 Ti, Zagitta reported BF16 performance increasing from 87.5 TFLOPS to 157.41 TFLOPS. TF32 throughput improved by 94%, while FP16 saw a 35.4% increase. Real-world AI tests showed throughput gains of 60.2% for MiniMax H3 and 58.6% for MOSS-Image SFT. Other tests reported an 81.2% BF16 improvement on the RTX 3070 and roughly 70–73% on the RTX 3080 10GB. An RTX 3060 12GB showed an 84% BF16 increase at a 170W power limit.

JustAResearcher developed a separate project, GPU Unlock Ada, for the RTX 40-series. For the RTX 4090, the mod reportedly doubled FP16 and BF16 throughput, moving from approximately 165.6 TFLOPS to 331.6 TFLOPS. The tool also provided nearly identical gains for the RTX 4070 Ti SUPER.

These modifications do not require a VBIOS flash and can work alongside stock NVIDIA drivers. However, reports say using these mods will void warranties and they must be enabled at boot to function before the drivers load. While these tools improve LLM prefill and certain AI workloads, INT8 image generation and memory-bandwidth-limited decoding see little benefit. Gaming performance improvements remain unproven.