DeepSeek-V4.1-Flash hits 1,129 output tokens/sec on four AMD MI355X GPUs
Scobleizer · x · 2026-11-18
Scobleizer amplifies an article from Hassan's team: Sciforium's DeepSeek-V4.1-Flash deployment on four AMD Instinct MI355X GPUs reaches 1,129 output tokens per second at P90 using traffic-tuned speculative decoding drafters — roughly 10,000 tokens in about nine seconds.
Key points:
- Achieved under real production traffic, not a toy project; the team notes they didn't rewrite vLLM
- Not long ago people doubted AMD GPUs could keep up; now records are being set on them
- Scoble's observation: response speed changes how you work with AI — minute-long answers pull you out of the loop, second-level ones keep you in flow
More from Infra
- CoreWeave faces $19.05B principal-plus-interest bill by end of 2027 — SumitGup · 2026-11-18
- Sarvam's token factory vs OpenRouter: vertical integration vs aggregation explained — itsOmSarraf_ · 2026-11-18
- x402 opens four public working groups: domain discovery, identity, tax, card acceptance — kleffew94 · 2026-11-18
- Chutes partners with Alibaba Cloud to offer GLM-5.3 and DeepSeek V4.1 at 35% off — markjeffrey · 2026-11-18
- Bloomberg: New US Chip Rule Freezes $480M Nvidia GPU Loan for ByteDance in Malaysia — rohanpaul_ai · 2026-11-18
- One RTX 4090 runs Qwen3.8-Flash-Next: a hand-built inference engine, fully gated — Distinct-Pie2389 · 2026-11-18