Report
Optimized Kimi K3 inference on AMD MI355X is claimed to outpace a public Nvidia B300 result
Lorenz Research reports 38% more Kimi K3 throughput than B300's public result, rising to 86% with speculative decoding.
TLDR
Lorenz Research says its optimized Kimi K3 setup on one eight-GPU AMD MI355X node served 38% more tokens per GPU than Nvidia B300's public SemiAnalysis InferenceX result without speculative decoding, and 86% more with it. Its runs used 80 concurrent agents, compared with 70 in the B300 result. Lorenz says Kimi Verifier testing showed nearly unchanged quality.
Combined views
—
3 Sources, first seen ago
