MFU Is a Meaningless Metric — tok/s/gpu Is the Only One That Matters
mike64_t · x · 2026-10-30
mike64t continues the efficiency debate: commercially, consumers drive large-scale revenue (SpaceX earns more from Starlink than launches), and with depreciating hardware rent-seeking loses in the limit. Technically, he argues MFU isn't comparable across model sizes, precisions and sharding strategies, and most people compute it wrong enough to exceed 100% — tok/s/gpu is the only metric worth optimizing.
More from Infra
- You can fingerprint your exact GPU with a single torch.index_add_ call — CatAstro_Piyush · 2026-10-31
- SGLang boosts DeepSeek V4.1 Flash throughput 67% to 1.09M tok/s per GPU in 4 days — BanghuaZ · 2026-10-31
- llama.cpp ROCm regression: prompt processing drops from ~1200 to ~460 t/s on RX 9070XT — uber-linny · 2026-10-31
- Sub-scale neocloud warning: VC-backed GPU cloud startups reportedly struggling — AccBalanced · 2026-10-31
- Qwen4 with HC architecture could widen the Q4 vs Q6 quantization gap significantly — Ok-Shower7286 · 2026-10-31
- Too many inference providers, too little demand: a developer's take — AccBalanced · 2026-10-31