AMD MI355X beats NVIDIA B300 by 24% throughput serving Kimi K3, 87% more tokens per dollar
AccBalanced · x · 2026-10-17
SemiAnalysis benchmarking shows AMD's MI355X is now competitive with NVIDIA's B300 for serving Kimi Moonshot's K3: 8× MI355X delivers 24.4% more throughput than HGX B300 (7,567 vs 6,083 tok/s/GPU) at 50 tok/s/user on the AgentX agentic coding benchmark, and 87.4% more tokens per dollar. The gains come from vLLM optimizations, AMD kernels, and hardware-level improvements closing the gap.
Related event: AMD MI355X beats NVIDIA B300 by 24% in Kimi K3 inference throughput(2 posts)→
More from Infra
- Unsloth adds native ARM64 Windows CUDA support for RTX Spark, 2.7x faster training — danielhanchen · 2026-10-17
- SGLang Summit sponsors jump from 11 to 36 in a week; Lip-Bu Tan, Lilian Weng to speak — BanghuaZ · 2026-10-17
- AMD Helios switch tray spotted at OCP 2026 with dual Broadcom Tomahawk 6 chips — wkmyrhang · 2026-10-17
- Reddit thread compares open-source LLM inference providers as newcomer Mango appears — No-Candidate8594 · 2026-10-17
- VaeZip: open-source codec stores VAE latent datasets 5-10x smaller with near-zero quality loss — TensorForger · 2026-10-17
- One plain-text file runs time on every device; UCLA prof maintains tzdata as volunteer since 2005 — PMinervini · 2026-10-17