gpu-unlock doubles RTX 3090 Tensor throughput to 157 BF16 TFLOPS via kernel module
tokenbender · x · 2026-10-25
Open-source gpu-unlock (3.7k stars) uses a Linux kernel module to rewrite registers before the NVIDIA driver binds, unlocking the FP16/BF16 instruction issue rate NVIDIA caps on RTX 30-series cards—no BIOS flash or driver patch needed.
- Results: 3090 Ti BF16 87.5→157 TFLOPS, TF32 42.6→82.6; RTX 3080 FP16/BF16 65→110+ TFLOPS.
- Real-world gains: vLLM long-context prefill 33–36% faster, 25% lower time-to-first-token, bit-identical outputs, reproduced on 3070/3080/3090.
- Usage (Linux only): clone, make install, add gpuunlock.targets=all kernel param, disable Secure Boot, verify via journalctl -k | grep gpuunlock.
- Caveat: warm reboots fail with -ETIMEDOUT and can leave GPUs unusable until a full cold reboot.
More from coding & agent
- pgrust's rigorous testing has already surfaced multiple real Postgres bugs — cedric_chee · 2026-10-25
- penberg: Foundation models find C++ security bugs en masse, yet people still call Rust 'slow to compile' — mgill25 · 2026-10-25
- Thorsten Ball: how he 10x's with Amp, and agents leaking secrets via CPU cache — Register Spill (Thorsten Ball) · 2026-10-25
- Personal AI assistant space ships fast: Genspark Claw and Lindy roll out new features in 4 days — FinanceYF5 · 2026-10-25
- Reddit cost table: cheap-first-then-escalate beats Sonnet-only for agent runs — lucasbennett_1 · 2026-10-25
- Infographic rounds up 20 AI agent tools from model choice to deployment — Aiden_Tech_Ai · 2026-10-25