AMD says MI355X delivers up to 2x B200 performance per dollar on Kimi K3
Overview
SemiAnalysis reports that AMD says its MI355X delivers up to 2x better performance per dollar than Nvidia's B200 on Kimi K3 agentic inference.
AMD attributes the gain to software work, including tuned AITER kernels and MonoKernel decode.
AMD also says the MI355X's 2304GB of HBM per server lets it avoid the 8x400Gbit/s scale-out networking between servers that B200 must use, because B200 has less than 1.5TB of HBM. The figures are AMD's claims as relayed by SemiAnalysis, and the report does not include independent benchmark results.
Written by AI from the articles below · updated Oct 31, 3:30 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
SemiAnalysis@SemiAnalysis_XAMD MI355X claims up to 2x B200 performance per dollar on Kimi K3AIAMD says its MI355X delivers up to 2x better performance per dollar than B200 on Kimi K3 agentic inference. The gain comes from software optimizations, including tuned AITER kernels and MonoKernel decode. AMD also says the MI355X's 2304GB of HBM per server avoids the slow 8x400Gbit/s scale-out networking that B200 needs, since B200 has less than 1.5TB of HBM.

Heat Trend
Not enough continuous observations to show a trend yet.