A whole corpus queued for offline AI scoring — which pulls harder on the 8558, the 48 cores or AMX, and where does the run stall?

  AMX does the pulling; memory and I/O decide the pace. Scoring is short matrix batches: with AMX engaged, bf16/int8 throughput across 48 cores and 96 threads climbs visibly and the 4GHz boost keeps each text slice brief, but the corpus still streams in from storage and results stream out, 260MB of L3 holds only a corner of the weights, and disks plus bandwidth usually hit the ceiling first.

  Shops with heavy offline scoring volume should buy by throughput; interactive single-request work never needs all 48 cores awake. Switch the batch stack to bf16/int8 and confirm the AMX path is live before the run — miss that instruction set and the core count earns nothing.