Announcement
AMD MI355X reportedly tops Nvidia HGX B300 in per-GPU throughput serving Kimi K3
The vLLM project reports 24.4% higher per-GPU throughput at 50 tokens per second per user.
TLDR
The vLLM project says eight AMD MI355X GPUs delivered 7,567 tokens per second per GPU, versus 6,083 for HGX B300, while serving Kimi K3 at 50 tokens per second per user on SemiAnalysis’ AgentX agentic coding benchmark. It also reports 87.4% more tokens per dollar and says its writeup covers vLLM optimizations, AMD kernels, hardware details and ways to spot bottlenecks.
Combined views
—
1 Source, first seen ago
