Report
AMD's MI355X reportedly reaches up to 2x better performance per dollar than B200 on Kimi K3 agentic inference
AMD's Anush Elangovan and his team reportedly got there through software optimizations including AITER kernel tuning and MonoKernel decode.
TLDR
MI355X servers reportedly have 2,304GB of HBM each, letting them avoid the slower inter-server networking that B200 needs with less than 1.5TB of HBM. The reported performance-per-dollar gain also reflects AITER kernel tuning and MonoKernel decode.
Combined views
—
2 Sources, first seen ago
