Artificial segmentation? That’s the only thing I can think of. They’ve literally doubled/quadruped the surface area for 1/2 the power? Somehow also with less ports.
Either that or somehow without the Mellanox port they’ve reduced idle further from 25-35W? But that seems unlikely imo.
Or maybe it’s the other way around, the Spark guzzles power for minimal performance benefit (like 600W RTX PRO 6000 Workstations vs Max-Q) - I guess one possibility is power limiting Spark to ~65-70W which is the average I’ve seen for LLM usage.