Is Local LLM Prefill Throughput Severely Underestimated?
GabryIta · reddit · 2026-07-07
A developer points out that when discussing the ROI of running local LLMs, the focus is almost entirely on decoding (output) speed, while prefill (input) speed is ignored. Using the example of running GLM 5.2 (with 4-bit quantization, speculative decoding, and other optimizations) on 4 NVIDIA DGX Spark units, one can achieve about 60 output tokens/second with 6 concurrent requests. Theoretically, running 24/7 yields about 5.18 million output tokens a day, costing roughly $22/day at $4.40 per million output tokens. However, the prefill throughput for the same configuration is about 3000 tokens/second—50 times faster than output. Even with a lower prefill unit price (about $1.40 per million input tokens), this massive throughput gap should significantly impact ROI. The author questions why the industry rarely includes prefill in its evaluations.
More from Infra
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- NVIDIA claims HGX Rubin NVL8 delivers 10x tokens per megawatt vs B200 — thursdai_pod · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Together lets multiple LoRA post-training experiments share the same GPU capacity — togethercompute · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27