NVIDIA consumer GPUs run FP32-accumulation GEMMs at half speed — it's just a firmware flag
panchovix · reddit · 2026-10-24
A detailed Reddit breakdown of NVIDIA's consumer GPU 'unlock' saga: RTX 20/30/40/50 series deliberately run FP16/BF16 (and FP8 on 40+) GEMMs with FP32 accumulation at half speed, while pro cards (A6000, L40, RTX 6000 Ada/PRO) are full speed. Affected workloads: LoRA/fine-tuning, diffusion, LLM prefill and large-batch inference — the author measured an RTX 6000 PRO 70% faster than a 5090 despite only 10-15% more cores. Batch-1 textgen (bandwidth-bound), INT4/8 and FP4 are largely unaffected. Digging through code shows the difference is literally a "HALFSPEED"/"FULLSPEED" flag; 3090 Ti and A6000 share the same die. Unlock feasibility: high for 20/30 series, moderate for 40 series (UEFI patch + custom VBIOS), low for 50 series due to added fuses. Unlocking poses no hardware risk.
More from Infra
- US electricity prices up 34% at investor-owned utilities since 2020 vs 19% public — randal_olson · 2026-10-24
- ExecuTorch now builds on Windows Snapdragon X Elite, demoing on-device Llama 3.2 1B — carrycooldude · 2026-10-24
- PyTorch models now run directly on Cortex-M microcontrollers, under 90 KiB RAM — PyTorch · 2026-10-24
- Strata adds Qwen 3.6 35B support, but decode speed loses to llama.cpp on MI50 — JsThiago5 · 2026-10-24
- Google Cloud launches CLI Remote MCP Server for agents to run gcloud & bq without local installs — rseroter · 2026-10-24
- Baseten launches Environment Network, pairing vetted RL environment firms with post-training customers — baseten · 2026-10-24