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

	Running Kimi K3 on MI355X at Better Performance per Dollar Than B300(wafer.ai)
	208 points by ilreb 1 day ago \| 102 comments
	tl;dr: Wafer ran the 2.8T-parameter Kimi K3 model on AMD MI355X (TP8), hitting 952 tok/s/node and 118 tok/s single-stream — worse absolute throughput than a B300 but ~1.5× better performance per dollar given the MI355X's ~2.4× lower price. Getting there required two fixes: implementing a missing top-k renorm function in sglang's ROCm sampling path to enable speculative decoding, and zero-padding attention heads from 12 to 16 to unlock AITER's fast MLA prefill kernel (~2–3× faster TTFT). No custom kernels were needed.
	HN Discussion: ↓The price comparison is unfair or cherry-picked, using selective cloud rental rates while ignoring power and other costs ↓The article reads as AMD/Wafer marketing slop rather than legitimate technical analysis ↓B300 wins on every raw performance metric; the comparison is misleading since the model barely fits the hardware being compared ↓The capex on MI355X hardware isn't justified by the quoted hourly rental rates ↓The writing quality and AI-assisted engineering choices raise credibility concerns