GPU-unlock project boosts RTX 30-series BF16 from 87.5 to 157.4 TFLOPS, 9.1k stars
StasBekman · x · 2026-10-27
The gpu-unlock GitHub project (9.1k stars) uses a Linux kernel module that runs the unlock and writes speed selectors (SS0 0x88888888, SS1 0x8) before the stock NVIDIA driver binds the card, restoring the full FP/BF16 issue rate of RTX 30-series consumer cards to their professional equivalents — no per-driver patching needed. Measured results: BF16 jumps from 87.5 to 157.41 TFLOPS, FP16 from 87.46 to 118.39, TF32 from 42.58 to 82.61. Stas Bekman notes that if NVIDIA sells the same chip far cheaper and still profits heavily, datacenter GPU margins must be enormous.
Related event: gpu-unlock Unlocks Hidden Tensor Core Performance on RTX 20/30/40 Cards(7 posts)→
More from Infra
- RTX 5080 price jumps from $1658 to $2318 in ten days on Amazon — Frizzy-MacDrizzle · 2026-10-27
- ComfyUI on 12GB VRAM: measured peaks show offloading turns 24s into 11 minutes — Impossible-Mail-4619 · 2026-10-27
- Neon rebuilds its docs search entirely on its own stack: Lakebase Search, Functions, AI Gateway — matei_zaharia · 2026-10-27
- PyTorch details 2.12x faster on-device streaming ASR on Arm SME2 via ExecuTorch — PyTorch · 2026-10-27
- NVIDIA shows speculative decoding speeding up LLM inference on Jetson with Qwen3.5 9B — NVIDIA Developer · 2026-10-27
- Revenue per GPU can rise as token prices fall, argues Gavin Baker — GavinSBaker · 2026-10-27