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

AMD's MI355X GPU delivers superior cost-efficiency compared to NVIDIA's B300 and B200 when running the 2.8 trillion-parameter Kimi K3 model, achieving 48 tokens/second per dollar versus B300's 33 tok/s/$ at 2.4× lower cost per GPU. While B300 nodes still achieve higher absolute throughput, the MI355X's 288GB VRAM capacity matches B300's specs and provides a practical advantage for extremely large models like Kimi K3 that don't fit on single B200 nodes, marking a significant opportunity for AMD in frontier model serving.