AMD MI355X Beats Nvidia B300 on Performance-per-Dollar for Kimi K3
Running the 2.8‑trillion‑parameter Kimi K3 on AMD’s MI355X yields 952 tokens per second per node—over three times the throughput of a dual‑node Nvidia B200 setup, and 1.3× the single‑stream speed of a B300. At roughly 2.4× lower GPU cost, the MI355X delivers superior performance‑per‑dollar, reshaping inference hardware economics.