In addition to our new DeepSeek V4.1 Flash results, we’re adding AMD’s MI355X to AA-AgentPerf for Kimi K3, GLM-5.2/5.3 and DeepSeek V4 Pro - comparing it to our latest results for NVIDIA systems on coding agent workloads.

AA-AgentPerf is our agentic inference performance benchmark. It replays real agentic coding sessions, including prompt caching and tool-call delays, to give a realistic measure of inference performance in day-to-day agentic workloads.

MI355X is AMD's flagship accelerator in the MI350 series, launched in June 2025 on its CDNA 4 architecture. It has the same 288GB of HBM3E at 8TB/s as NVIDIA's B300, around 10 PFLOPS of dense FP4 compute (vs. B300 at ~13.5 and B200 at ~9) and board power of up to 1,400W. All our MI355X results use ROCm nightly and dev builds of vLLM and SGLang.

We are debuting results for AMD's MI355X on a single 8-GPU node across all four featured AA-AgentPerf models. Results on our leaderboard are live and will evolve over time - we look forward to tracking software/serving improvements on AMD hardware.

Initial results on MI355X:

➤ Its strongest result is on DeepSeek V4 Pro 0813, where MI355X serves ~15% to 40% more throughput per MW than B200 between 35 and 75 tok/s per user, likely helped by its 288GB of HBM per GPU (B200 has 180GB) on a 1.6T parameter model

➤ On DeepSeek V4.1 Flash, MI355X is in line with NVIDIA Blackwell for a single agent at 647 tok/s (B300 at 652, B200 at 669), but falls behind as load rises, reaching ~128M tokens/s per MW at 100 tok/s per user against ~193M for B200 and ~220M for B300

➤ On GLM-5.2/5.3, MI355X trails B200 and B300 at 100 tok/s per user, with ~26M tokens/s per MW against ~42M and ~44M respectively

➤ On Kimi K3 (2.8T parameters), MI355X serves ~8M tokens/s per MW at 50 tok/s per user against ~24M for B300 using NVIDIA's multi-node configs, and 152 tok/s for a single agent against 235