Maynard Handley (name99.delete@this.name99.org) on February 4, 2015 2:59 pm wrote:
> David Kanter (dkanter.delete@this.realworldtech.com) on February 4, 2015 11:51 am wrote:
> > Eric Bron (eric.bron.delete@this.zvisuel.privatefortest.com) on February 4, 2015 3:42 am wrote:
> > > > Go to this page:
> > > >
> > > > http://www.arm.com/products/processors/cortex-a/cortex-a72-processor.php
> > > >
> > > > and click on the performance tab. 3.5/1.9 ~= 1.84.
> > > >
> > > > And don't forget to note the title "Increase in sustained performance in smartphone power budget.
> > > >
> > > > That makes it hard to even guess how much IPC increase there is, or what top performance
> > > > is compared to unconstrained A15/A57.
> > >
> > > indeed, and thank you for the link
> > >
> > > btw reasonable people seem to dismiss insane IPC gains:
> > >
> > > http://forums.anandtech.com/showpost.php?p=37140914&postcount=20
> > >
> > > https://twitter.com/TheKanter/status/562788528137187329
> > >
> >
> > At the risk of offending people from ARM...I find that ARM's marketing to be absolutely ridiculous
> > and deceptive. Comparing IP blocks at different process nodes is just bullshit designed to obfuscate
> > the value that the microarchitecture is adding. That is just as terrible as when Intel tries
> > to talk about advances in server performance by comparing against Nehalem-EP.
> >
> > While a comparison against an old core on a old process MIGHT be indicative of what users MAY experience,
> > it simply isn't even remotely apples-to-apples. Moreover, USERS DON'T BUY FROM ARM. It's just rubbish.
> >
> > At least Nvidia's marketing does proper generation to generation comparisons.
> >
> > ARM has plenty of smart architects, and frankly the marketing
> > group does a disservice to them by not talking
> > openly and honestly about the value they are adding for
> > customers. Intel will honestly tell you that improving
> > IPC by 5% is hard - and that's the truth for their design point. ARM has much lower performance cores that
> > have way more low hanging fruit, so 10-15% should be feasible. It's better to be honest than try and dress
> > up someone elses technical innovations (e.g., TSMC's FinFETs) as your own magnificence.
> >
> > Put another way, I can travel from point A to point B at roughly 120mph. Does that mean I am
> > an amazing speedy person? No, it means I have a Honda Civic and can find the highway...
> >
> > David
>
> Hmm. When I was talking about the fact that Cyclone could match Intel in IPC, I was told in
> no uncertain terms that an interest purely in IPC, decoupled from the actual frequencies that
> a chip could reach and the TOTAL overall performance, was the sign of idiotic fanboy'ism.

Yes it is.

> Just saying...

These are entirely consistent. I'm saying it's important to normalize for what matters. In your case, you were trying to argue that IPC of a 1GHz chip phone chip is comparable to a 4GHz notebook/desktop chip, which is clearly idiotic fanboyism :)

Now I'm saying that comparing the performance of two IP cores with the same design target (e.g., smart phones) on totally different process nodes is bullshit. Which it is.

David