Megakernel optimization speeds DeepSeek V4.1 Flash decode up to 2.27x on AMD MI325X with no accuracy loss
Electronic-Ebb9771 · reddit · 2026-11-05
Sciforium engineers combined many small GPU operations into a single megakernel, speeding up DeepSeek V4.1 Flash 1017 decode by up to 2.27x per user on AMD MI325X with no accuracy loss, across code generation, math, writing, and general Q&A workloads — same weights, only the decode kernel changed.
The takeaway for agent workloads: model choice is only part of the performance equation; kernel and serving-stack optimization can materially improve latency, throughput, and infra efficiency without touching the model. The company owns its GPUs and prices its serverless API 50% below the standard rates of major inference providers it benchmarked against.
More from Infra
- Nvidia down, Amazon up: traders may be betting existing datacenters gain value as construction slows — Darpinian · 2026-11-05
- RL agent reportedly makes $85,771 trading a real Texas battery, 4x the human operator's $20,647 — xeophon · 2026-11-05
- Qualcomm Q4 revenue falls 8% y/y, non-GAAP earnings down 28% — firstadopter · 2026-11-05
- Cybersecurity could eat 20% of token compute, analysts model an AI arms race — BenBajarin · 2026-11-05
- Apple cuts engineers' internal AI spending caps from $300 to $200 per day, plans deeper cuts — rohanpaul_ai · 2026-11-05
- SkyPilot Tokens lists Kimi K3.1 Pro day-0: 2.8T params, 1M context, 40% cheaper — skypilot_org · 2026-11-05