Xiaomi AI Cube and Xring O100: 1.22 TB/S, 330 Tokens/S and 120B Local AI (aicybr.com)

TOPS seems to be low though, so fill rate is probably ~10x slower than a 4090.