AMD MI355X beats B200 on DeepSeek V4 Pro per-MW throughput but trails badly on Kimi K3
ArtificialAnlys · x · 2026-10-15
Artificial Analysis added AMD's flagship MI355X to its AA-AgentPerf agentic inference benchmark, replaying real coding-agent sessions on a single 8-GPU node (ROCm nightly, vLLM/SGLang dev builds) against NVIDIA B200/B300.
Key findings:
- DeepSeek V4 Pro 0813 (1.6T params): MI355X's best showing — 15-40% more throughput per MW than B200 at 35-75 tok/s per user, peaking at 94M tokens/s/MW vs B300's 125M in disaggregated multi-node configs; single-agent 240 tok/s vs B200's 367. Its 288GB HBM3E (vs B200's 180GB) helps on large models.
- DeepSeek V4.1 Flash: single-agent 647 tok/s, on par with B300 (652) and B200 (669), but falls behind under load (128M vs 193M/220M tokens/s/MW at 100 tok/s).
- GLM-5.2/5.3: between Hopper and Blackwell — 26M tokens/s/MW at 100 tok/s vs H200's 16M, B200's 42M, B300's 44M; single-agent 228 tok/s matches H200.
- Kimi K3 (2.8T/104B active): largest gap — 8M tokens/s/MW at 50 tok/s, about a third of B300's 24M with multi-node disaggregation; 152 vs 235 tok/s single-agent.
Results are live on the leaderboard and will track AMD software/serving improvements.
Related event: AMD MI355X Benchmarked on AA-AgentPerf: Mixed Efficiency Results(2 posts)→
More from Infra
- Mine Bittensor Subnet 92 by pointing Claude Code at one link — airesearch12 · 2026-10-15
- OpenAI infra lead leaves to start up after locking in double-digit GWs of compute — DesireeCachette · 2026-10-15
- Who Sells Data directory maps 159 AI training data vendors, compares Scale vs Mercor — cneuralnetwork · 2026-10-15
- Compute is the next commodity: building financial infrastructure for GPU markets — AnneliesGamble · 2026-10-15
- OCP slide reveals NVIDIA Rubin Ultra NVL576 will use CPO for scale-up networking — BenBajarin · 2026-10-15
- TensorFold ships macOS app for local LLM inference, reports 874 tok/s on RTX PRO 6000 — HankYeomans · 2026-10-15