Reply to post: Re: How many companies have to fail at server-side ARM64 ...

Re: How many companies have to fail at server-side ARM64 ...

What has happened is the major buyers of sophisticated CPUs -- the cloud companies -- want performance per watt as well as performance per rack unit.

Compare the resulting pricing for AWS: Intel US$4.08ph, AMD $3.70ph, ARM $2.18ph. Graviton2 is about 20% slower than the equivalent Intel server, but about half the cost. Remember this is the second release of Amazon's ARM design up against the decades of tuning of Intel's design, and the difference is only 20%. Obviously that difference has further to shrink.

Other cloud providers will be facing similar pricing, but with the advantage that they use more of their compute cycles for their own services. That is, they can more readily re-target their internal services from AMD64 to ARM64 than Amazon's clients can.

I can't see that any company will take the risk of developing a server ARM chip. As you point out, plenty of startups have been burned. So the market will leave that development to the cloud providers themselves, who have abundant engineering resources to turn ARM IP into silicon.

The major difference between now and the past is Intel's years-long failure to deliver process improvements compared with its competitor TSMC. There is little reason to expect that to change. That failure alters the economics for cloud companies. In the past chips with better architecture would have their performance blitzed by Intel's process improvements. DEC's Alpha being an excellent historical example. So there was no incentive for cloud companies to explore CPU architecture. That blitzing-by-process-improvement is no longer in Intel's power to do. An architectural improvement over Intel's microarchitecture is now a long-run win. So CPU architecture is now worth cloud companies' efforts.

None of this is likely to be reflected in the "enterprise server" market. But that market is becoming increasingly odd and continually smaller. In many ways very much like the IBM mainframe business of the pre-PC 1980s. And just as likely to have a nasty surprise.