juanrga (nospam.delete@this.juanrga.com) on August 13, 2014 3:16 pm wrote:
> David Kanter (dkanter.delete@this.realworldtech.com) on August 13, 2014 1:07 pm wrote:
> > juanrga (nospam.delete@this.juanrga.com) on August 11, 2014 7:00 pm wrote:
> > > David Kanter (dkanter.delete@this.realworldtech.com) on August 10, 2014 9:12 pm wrote:
> > > > > 20--30% sounds as the right efficiency numbers at that level. Precisely 90W
> > > > > ARM SoCs are providing around 80--90% of performance of 140W Haswell Xeons.
> > > >
> > > > Seriously dude. What 90W ARM server chip is providing performance to customers?
> > > >
> > > > Let me remind you that server chips must be implemented in
> > > > silicon, not power point to provide value to customers.
> > > >
> > > > And I'm quite willing to bet that no ARM server design in
> > > > the next 2-3 years will provide 80% of the performance
> > > > of the highest bin Xeon. If they are lucky, they might get to the low-end territory. They probably also
> > > > won't have as much memory capacity and generally be inferior across a number of dimensions.
> > > >
> > > > DK
> > >
> > > First, you pretended that you didn't hear of any high-performance
> > > ARM design, despite many being announced in many places.
> > >
> > > Then you ignored the specs given to you for one of those designs.
> >
> > There are a number of companies attempting to design high performance ARM server
> > chips. The specs even sound interesting on paper. But in reality, that doesn't
> > matter until the products are shipping to customers in general availability.
> >
> > Design targets are just that - targets. There is no guarantee Broadcom, Applied,
> > Cavium, etc. will hit their frequency or power targets. We don't know until the
> > chips are generally available in high volume, which won't be for a long time.
> >
> > Even Applied Micro systems are only available for pre-order right now, and they taped out a long time ago.
> >
> > In contrast to these 'design targets', Intel is shipping real products in high
> > volume today with widely available benchmark data. The two are entirely different
> > in terms of credibility, and it's deceptive to think about it otherwise.
> >
> > Moreover, remember that Intel is a moving target. They are shipping IVB-EP
> > right now, and will have HSW-EP next year, and BDW-EP the year after.
> >
> > David
>
> I agree with all of this David. But this isn't the point.
>
> The point is that when I began mentioning the server-class ARM cores/SoCs, some people here said me
> that no company had announced high-performance cores; that they never heard about 90W ARM SoC; that
> ARM couldn't scale up; that I was mentioning waporware because nothing was shipping, and so on. Some
> people even go further with ad-hominem and fallacies whereas ignoring my arguments, info, and links.

I am one of these people, and I violently disagree with you because:

1. Most ARM-based server chips target I agree entirely on that Broadcom can claim 90% single thread of Haswell Xeon performance and hit only
> 75% on final silicon, but then all of us have to wait and see final silicon benchmarked.

Broadcom's claims just aren't plausible. On paper, they have a nice looking CPU core...but it takes a lot more than picking the right ROB sizes, etc. The Broadcom cache hierarchy has historically been goofy (exclusive caches are really not a good idea!). Moreover, I'm not sure the fabric really has enough bandwidth (it's about 750GB/s).

At this moment
> I can say that Broadcom claims 90% of single thread performance and nobody would attack me by saying
> that. At best skeptics would share its skepticism on targets but they wouldn't attack me.

Except any reasonable person would be skeptical of such claims from broadcom. Moreover, you've been advocating the magic of ISA and making ridiculous technical claims that absolutely demonstrate a lack of critical thinking skills.

> I would add that there is no warranty that Intel, AMD, IBM, or any other company will hit targets >either.

Except Intel has been delivering the goods for the last 7 years quite consistently. Broadcom hasn't even delivered a single server chip. It's exceptionally likely that Broadcom won't be in high volume till late 2015 or 2016 - there's an awful lot of potential hangups out there.

David