David Kanter (dkanter.delete@this.realworldtech.com) on August 15, 2014 9:41 am wrote:
> That's right. I'm very very skeptical. For a number of reasons:
>
> 1. The only time Intel has really lost share in servers was when two things happened:
> A) Intel started screwing up with IA64 and the P4 AND
> B) AMD really executed well with the K8
>
> Intel isn't screwing up anymore. So I am somewhat skeptical that good execution by a competitor is
> sufficient to make a dent in Intel's server presence. I could be wrong, but it seems unlikely.
>
I think ARM's licensing model is a key difference this time. It significantly lowers the barriers for entry in the server space, which allows vendors to compete more cost effectively with Intel. Also a lot of common infrastructural things and key standards are already been taken care of, such as Linux, compilers, SBSA, etc. Intel has a tendency to come up with surprisingly effective strategies to deal with any competition, so it definitely won't be an easy battle for anyone involved. Frankly I expect most upcoming ARM-based vendors to fail to establish a foothold in this market. And only one or two companies may be mean and lean enought to survice or even thrive in their corner of the server space.

> 2. Everyone underestimates the time to market for servers.
> It really takes a while to get into general availability.
>
Agree. Any ARM-based company first need to come up with a first generation product that is competive against Intel Xeon starting from day 1. In my opinion that would mean a TCO that is at least 30% lower including amortized cost of migration. Then, equally important, as a company you need to prove you have the stamina and engineering skills to remain competitive for a number of years.



> 3. I think a lot of people underestimate the complexity of a first generation server part. It's
> not just a good CPU core, it's the on-chip fabric, the cache, memory controllers, IO, networking,
> etc. And it's important to get nearly everything right...and then do something to differnetiate
> from Intel (right now, highly integrated networking seems like a good call).
>
TCO and risk are the only two factors that determines if any enterprise product is successful or not. 64-bit ARM will likely be perceived as a higher risk compared to Intel for the years to come. So ARM-based vendor need to bridge that disadvantage with significant TCO advantages. This can be done by either lowering cost of purchase / energy usage, allowing higher density, or by providing features that are offered by Intel in their high-end Xeons only.

The balance between TCO advantage and risk disadvantage is different per market. I expect first movers will be the likes of Google, Facebook, and possibly the HPC market. These parties are less senstive to risk (if you develop self driving cars I believe that is a fair statement to say), and are very very sensitive to TCO. One of these million server companies removed even the logos from their servers to increase airflow a tiny bit (that was before they start designing and building their own server hardware from scratch).

I assume for now there is going to be a TCO advantage for ARM-based servers. But if none of the these companies is able to achieve that consistently, then there is simply never going to be a business case anyway. ISA may help reducing TCO (by how much we can debate endlessly), but it never has been and never will be a key differentiating factor. Otherwise x86 would have been dead and buried a long time ago. Sorry geeks.

> 4. Marketing people talk about Xeon-class performance, but
> then typically end up delivering Xeon E3 performance :)
>
> >You are rejecting such claims with arguments like vaporware, unreleased
> > products, NDA, not freely available, etc. etc. At the other hand you have no problem to be very
> > critical on Broadcom Vulcan which is two years to be released down the road.
>
> I am skeptical about the first generation of products from companies that have little
> to no experience with servers. I believe Cavium can make a great networking chip and
> sell it - I'm far less confident about markets where they have no presence.
>
> I also am skeptical that anyone can match Intel in performance while trailing behind by a node. To do that,
> it is necessary to offer a very different product with a different system architecture. E.g., target workloads
> where cache does not help and simply slap down more memory controllers and cores (hint: that's a GPU!).
>
> > I think we could either restrict our opinions on officially released products and verifiable benchmark
> > figures, or allow some free (and sometimes wild) thinking about what may be available in the future.
> > With the former approach this forum will quickly become very boring. So suggest to go with the latter
> > approach. Of cause critical thinking is always welcome, but then a level play field.
>
> I just haven't heard any reasons why Broadcom's claims are credible.
>
> David

Broadcom has an interesting architecture and Jim Keller has a fantastic track record for designing brilliant architectures, but both products are still far out in the future. Also Vulcan and possibly K12 depend on 16 nm FINFET which is in my opinion bares some risk for additional delays. Agree with David about Cavium.

So for now AMD and Applied Micro would be my pick for the top dogs in the ARM server space. First generation seems to be good enough to generate some revenue and to start the ball rolling. But 2nd or 3rd generation is going to be critical for real longer term success.

May you live in interesting times