GPU Unlock Doubles RTX 30 AI Throughput
This open-source Linux kernel module bypasses the FP16/BF16 instruction-rate cap on supported NVIDIA RTX 30-series GPUs before the unmodified driver loads. Reported RTX 3090 Ti BF16 throughput rises from 87.5 to 157.4 TFLOPS, though real application gains vary by workload.
gpu-unlock could extend the useful life of aging RTX hardware, but its headline benchmark gains should not be mistaken for universal inference speedups.
- –Runs from the initramfs, modifies GPU speed selectors, and resets the card before NVIDIA’s driver binds
- –Requires no VBIOS flashing or recurring driver patches
- –Community reports show roughly 10–20% gains in some LLM workloads and larger gains in BF16-heavy image generation
- –INT8 kernels, memory transfers, decoding, and other non-FP/BF16 stages may see little or no improvement
- –Experimental Linux-only kernel-level hardware manipulation carries compatibility and stability risks
DISCOVERED
2h ago
2026-10-24
PUBLISHED
2h ago
2026-10-24
RELEVANCE
AUTHOR
0x0SojalSec