I got my boss to get me the most powerful server we could find, $15000 or so. In benchmarks there was minimal benefit and sometimes a loss going with more than 40 cores even though it has 56. (52? - I can't check now) sometimes using more cores slows the build down. We have concluded that memory bandwidth is the limit, but are not sure how to prove it.

If that's true than have you looked at the threadripper or the new Ryzen AI+ 395? I think it has north of 200gbps

i have not (above machine was a intel), someone else did get a threaeripper though I don't know which. He reborted similar numbers though I think he was able to use more cores still not all.

The larger point is the fastest may not be faster for your workload so benchmark before spending money. Your workload may be different.