NVIDIA deliberately halved FP32 accumulation on consumer RTX GPUs — unlocks possible for 20/30 series, not 50
panchovix · reddit · 2026-10-23
Reddit user panchovix breaks down NVIDIA's market segmentation: consumer RTX GPUs run FP32-accumulation GEMMs at half speed — controlled by a literal "HALFSPEED" flag in code — while prosumer/datacenter cards (A6000, A100, L40, RTX 6000 PRO, B200, Hopper) are unlocked.
Impact by generation: Turing (RTX 20) halves FP16; Ampere (RTX 30) halves FP16 and BF16; Ada (RTX 40) halves FP16, BF16, and FP8; Blackwell (RTX 50) halves FP16/BF16/FP8 but oddly leaves FP4 untouched. Affected workloads: LoRA/finetuning, diffusion generation, LLM prefill, and large-batch generation. The author's RTX 6000 PRO is 70% faster than a 5090 at prefill despite only 10-15% more cores. Batch-size-1 generation stays bandwidth-bound and largely unaffected.
Unlock outlook: RTX 20/30 series have a high chance via reverse engineering; RTX 40 is feasible (UEFI patching + custom VBIOS); RTX 50 is likely locked for good since NVIDIA added fuses that read the flag.
Related event: gpu-unlock kernel module unlocks RTX 3090's half-speed tensor performance(6 posts)→
More from Infra
- 2-bit quantized DeepSeek 4.1 Flash tested locally: KDL degrades but still works — bakawolf123 · 2026-10-23
- Developers now deploying 1T-parameter models on their own infrastructure — dejavucoder · 2026-10-23
- Liquid AI launches RIPL inference engine generating full LLM responses in 30ms — JosephJacks_ · 2026-10-23
- Tech infra manias historically last 4-6 years — next year is AI's fifth — oran_ge · 2026-10-23
- Cambricon revenue doubles: $1.34B in first three quarters, net profit up 131% — teortaxesTex · 2026-10-23
- Which US State Will Rebrand as 'The Data Center State' and Build a Sovereign Wealth Fund? — matt_slotnick · 2026-10-23