Running Kimi K3 on MI355X at Better Performance per Dollar Than B300

AMD's MI355X GPU runs the large 2.8 trillion parameter Kimi K3 model at 952 tokens per second per node, outperforming NVIDIA's B200 nodes in throughput per dollar despite some performance trade-offs. This demonstrates AMD's cost-effective hardware viability for serving frontier open-source AI models, overcoming previous software support challenges.