AMD achieves up to 2x better performance per dollar than B200 on agentic inference

AMD's engineering team achieved up to 2x better performance per dollar compared to the B200 on Kimi K3 Agentic Inference. The performance gains stem from software optimization efforts, including tuning AITER kernels and MonoKernel decode. Additionally, the MI355X's 2304GB of HBM per server allows it to avoid slow scale-out networking between servers. Open article →