Performance per dollar has been discussed to death, but what seems to be missed is the bit about 'the workloads'. There wouldn't be much of an impetus to move to ARM even if it was slightly better on performance per dollar. So it isn't general purpose processor performance that would win out. It may be possible to add specialized support to a slower processor and get much better value for money.
Right, that's similar to what I said, except I suggested that maybe it was a discrete component that does not replace the Xeon processor at all. After all, why get rid of the fastest scalar engine in the system and replace it with a slower one, if all you need is the additional accelerator? A stand-alone IP is cheaper and lower power, and good software can hide the latency of a discrete IP (just look at what has been accomplished via discrete graphics).
Market Data 
Markets 
AI
