GPU FP32-accumulation unlocks explained: 4090 hits 331 TFLOPS, but Blackwell stays locked
panchovix · reddit · 2026-10-27
A detailed Reddit writeup explains the recent GPU "FP32 accumulation unlocks": NVIDIA fuses consumer cards to half-speed tensor ops (FP16/BF16/FP8/TF32 with FP32 accumulation) for market segmentation, while an override register locked to NVIDIA's most-privileged firmware can restore full speed. A buffer overflow in the signed Booter firmware (Jon Pry's "A Canary in the Crypto Mine") lets unlock tools open that lock on Turing, Ampere and Ada: 4090 FP16/BF16 goes 165→331 TFLOPS, 3090 Ti BF16 GEMM 87→157, 2080 Ti 47→94 (real GEMM gains under 2x due to power limits). Blackwell cards carry the same fuses but can't yet be unlocked due to changed firmware security.
Related event: How GPU FP32 Accumulation Unlocking Works and Why RTX 50 Can't(2 posts)→
More from Infra
- Unsloth details running DeepSeek-V4.1-Flash locally: 1-bit GGUF fits 128GB RAM with 1M context — gnukeith · 2026-10-27
- SK hynix down 47% since July while DDR5 spot prices rose 31%: a striking disconnect — tengyanAI · 2026-10-27
- JustGOAL paper: general frontier model beats specialized GPU kernel agents — azed_ai · 2026-10-27
- NVIDIA's Nemotron-3.5 Bolt 74B-A9B matches Super within 1% with 9B active params — ryanshrout · 2026-10-27
- Nvidia CEO says Kazakhstan's Data Center Valley could generate $1 trillion in economic activity — pstAsiatech · 2026-10-27
- Agentic AI Costs Are Scaling Faster Than Value—FinOps to the Rescue — DavidLinthicum · 2026-10-27