Now think what happens if absolute tip top performance is not required, just cost effectiveness as for Google.
Performance is ALL that matters, since Google has a throughput requirement to keep serving webpages to their billions of viewers. Not to mention that the content per viewer is and has been trending upward since the Internet first went online. And yet, Google has to meet those requirements with a budget that is not exponential. The very fact that they are investing in accelerators should tell you that performance matters a great deal!
Now, you're point might be that maybe performance matters, but they'll get that performance from an architecture other than Xeon. That's where I'm telling you that you're naive to how the model works. Accelerators take a great deal of the workload, but not 100% of the workload - and it will NEVER be 100% of the workload. In order to make the workload go faster, you need more accelerators, as well as more XEONs.
>> as for Google. Why should they have separate Xeon and FPGA cards?
Google has two distinct needs for CPU's.
1- to abstract and categorize content as they sift data from web-sites.
2- serve client requests for search.
Possibly #1 can benefit from an integrated solution. At the cost of complexity. Does the benefit outweigh ? maybe, and they might well experiment with that.
But #2 will stay with traditional servers. And if ARM cores did the job better, they'll switch. If not, not. So far, and for the forseeable future, not.
Market Data 
Markets 
AI
