Hacker News
Kimi K3 on 8× AMD MI355X Beats B300 on Throughput Per Dollar — $48 vs $33 tok/s per Dollar
Wafer.ai's July 31 writeup runs the 2.8T-parameter Kimi K3 at TP8 on a single MI355X node, hitting 952 tok/s aggregate, 118 tok/s single-stream decode and roughly 13k tok/s steady-state prefill after optimization work on missing PyTorch sampling functions and AITER MLA prefill kernels via SGLang and ROCm. Against a two-node TP16 B200 deployment that's 3.8× the aggregate throughput per node and over 1.3× single-stream decode; B300 delivers 1.65× higher aggregate throughput but costs 2.4× more per GPU, landing MI355X at $48 tok/s/$ versus B300's $33 at $2.50/GPU-hr. The caveat worth flagging: the numbers come from a vendor-adjacent blog, not an independent benchmark house.
↳ Follow the thread