A different view of the server market

STH has an interesting take on the server market for the next year or two.
https://www.servethehome.com/intel-accelerates-messaging-on-…

In summary, ARM, Intel, and AMD are taking very different approaches. ARM is “more cores, more better”. Intel is “more accelerators, more better”. AMD is taking a middle approach with a both more cores, and a few accelerators. The AMD Bergamo design is targeted directly at fending off the ARM direction. With the Pensando acquisition AMD is making progress on the accelerator direction.

They will all take different pieces of the market with their different strategies.
Alan

1 Like

Alan,

They will all take different pieces of the market with their different strategies.

Which pieces do you think that each will take, and why?

Norm.

Which pieces do you think that each will take, and why?

You can look through the article and watch the short videos and get a more complete idea of his expectations, but here are a few of the points I gleaned from it.

At the most fundamental level, they will use the solution with the lowest TCO that meets their needs. To some extent customers may stay on a higher TCO solution due to familiarity and experience with the legacy system.

The server market is huge and diverse, but here are some examples:

ARM will take applications that require large amounts of integer performance. They win at performance per watt and performance per dollar for integer heavy applications, but of course some applications are x86 only. Amazon has had their Graviton ARM CPU instances for a while. Microsoft just added Ampere ARM instances to the Azure cloud.

Intel will take applications with lots of “HTTPS” web transactions due to their quick assist accelerators.

Intel will also keep the “RAN” radio access network edge server market because of the software stack. I read that both AMD and ARM are working on this, so it could shift over time.

Software that has an expensive per core license fee will likely run on Intel as they have faster single core performance.

Applications with a light mixture of AI inferencing will probably run on Intel due to their built in AI accelerators. If they have a lot of AI inferencing they will likely use dedicated AI inferencing hardware like that available from NVIDIA. AMD is working on AI acceleration so this could change over time.

AMD owns the hyperscalers because of their best in class performance/watt and performance/$ on traditional X86 workloads.

Much of the AI training today is done on NVIDIA GPU instances. Amazon is pushing converting some of this to their Intel Gaudi AI instances:
https://aws.amazon.com/ec2/instance-types/dl1/

There are a lot of players in this market, and workloads are often a mixture of needs. As a result, companies that have resources deployed to help customers optimize workloads across the different platforms may get the business.

Alan

ARM will take applications that require large amounts of integer performance. They win at performance per watt and performance per dollar for integer heavy applications, but of course some applications are x86 only.

Decades ago, I wrote some packages which used double-precision floating-point to do >32-bit integer arithmetic. You don’t get 64-bit integer arithmetic but 52 to 55-bit integers with all the expected integer behavior. (Including truncation, modular arithmetic if needed, and most important, exact addition and subtraction.) The performance depends on the hardware of course, but a two to three times improvement over hardware integer support was common.

I needed it to solve a really nasty problem that is now gone. (Or at least it should be.) DEC VAXes had multiple floating-point formats none of them, at the time, IEEE. I needed to pass exact coordinates with about 50 bits of precision from VAXen to Sun Workstations. (The WG9 NumWG extended IEEE floating-point to trigonometric and exponential arithmetic with most functions a couple of lsbs. (The major exception is the exponential function.) Most important, the packages were as accurate as mathematically possible, and test libraries were included. I used it mostly for converting seismic data. You should be able to dig up an on-line version of the paper (really a book). If you need it, just find the command-line argument for strict floating point for your Ada compiler. I don’t know how many experts are left who understand it. I know how to use it, but my numerical expertise is in integer, modular and fixed-point. :wink:

Anyway, I expect anything that needs lots of integer arithmetic to use graphics cards. Crypto modular arithmetic and some FFT applications lean this way. (Not crypto as in using it, but in breaking it…Number field sieves are the algorithm of choice.)

I am not sure if it is possible to see the amount of capacity AWS or Azure has in different compute instances, but it sure would be interesting!
Alan