FULL STORY

RTX 30 Driver Lock Broken, Gains Debated

Open-source project gpu-unlock cracked NVIDIA's consumer GPU driver lock, nearly doubling RTX 3090 compute; community benchmarks later showed gains mostly in prefill and large-batch inference.

2026-10-24 ~ 2026-10-25 · 2 episodes · 10 posts

Episode 1 · GPU Unlock Cracks NVIDIA's Driver-Level Limiting, Nearly Doubling RTX 3090 AI Throughput (2026-10-24, 8 posts)

The open-source project Zagitta/gpu-unlock exploits a vulnerability in the GPU boot firmware to gain register write access before the NVIDIA driver loads, removing consumer RTX cards' instruction issue rate caps on FP16/BF16-input, FP32-accumulate GEMM operations — the "product segmentation" NVIDIA uses to separate consumer cards from professional ones. In real testing, a 3090 under the same 360W power limit saw BF16 throughput jump from about 75 TFLOPS to 135, with real workloads like ComfyUI benefiting noticeably.

Confirmed

  • The unlock applies to the RTX 3000 series (GA102/103/104/106); KohakuLab and ComfyUI both confirmed successful unlocks
  • 3090 PyTorch benchmarks show compute rising from 75 to 135 TFLOPS (relayed by @lxfater and @CeFurkan)
  • Reddit user @under--null's system testing: a 3090 Ti gained 47%-59% throughput on multiple BF16 models in ComfyUI
  • Reddit user @panchovix's in-depth write-up notes that consumer RTX 20/30/40/50 series FP16/BF16 (and FP8 from the 40 series onward) GEMM with FP32 accumulation is deliberately halved in performance, and that in testing this limit turns out to be just a flag in the firmware

Why it matters

  • The performance gap between consumer and professional cards (such as A6000, L40, RTX 6000 series) is largely an artificial software restriction, not a hardware difference — a single firmware flag removes it
  • @TheZachMueller commented that consumers previously had to either buy castrated cards or pay up for workstation-class products; now the card you buy is the full card — "this has never been achieved before." He added that the remaining shortcoming after the unlock is at the instruction-set level, arguing that if the instruction set were fully opened too, AMD would be unstoppable

Episode 2 · Nvidia Unlock Debated: Prefill and Large-Batch Gains, Not a Breakthrough (2026-10-24, 2 posts)

LottoLabs argues Nvidia's hardware unlock mainly speeds up prefill and large-batch inference rather than being a breakthrough, while TheZachMueller pushed back, leaving the community divided on its real impact.