By: Anon (no.delete@this.email.com), January 10, 2015 1:00 am
Room: Moderated Discussions
Mark Roulo (nothanks.delete@this.xxx.com) on January 9, 2015 9:13 am wrote:
> coppice (coppice.delete@this.dis.org) on January 8, 2015 11:38 pm wrote:
> > Mark Roulo (nothanks.delete@this.xxx.com) on January 8, 2015 6:37 pm wrote:
> > > juanrga (nospam.delete@this.juanrga.com) on January 7, 2015 6:03 am wrote:
> > > > I recall Aaron asking me why I believed that mixing VLIW with
> > > > RA was a good idea. Funny, because we know now that
> > > > Denver is an in-order VLIW+RA core that outperforms wide OoO cores such as Cyclone or Haswell.
> > >
> > > Have you not noticed that nVidia did not select the fastest 15W Haswell CPU they could?
> > >
> > > The 2955U Celeron is a 15W TDP part with no hyperthreading, a clockspeed
> > > of 1.4 GHz and 2 MB of L3 cache. Turbo-boost is disabled.
> > >
> > > The 4650U is *also* a 15W TDP Haswell part, but it has a base frequency
> > > of 1.7 GHz, 4 MB of L3 cache and will turbo to 3.3 GHz.
> > >
> > > Eyeballing the nVidia chart, the Denver (@ 2.5GHz?) beat
> > > the 2955U by about 10% on SpecInt2000. My guess is
> > > that a 4650U would beat the Denver chip by close to 2:1. When you lose by 2:1, you aren't "outperforming".
> >
> > The 4650U is a very expensive chip, while dual core denvers are going into tablets. Is that
> > a fair comparison? I guess winning, rather than fairness, was uppermost when the comparison
> > with the 2955U was made, but you have to look beyond one parameter, like watts.
>
> Since we are addressing the *performance* of the CPU, I think that the cost of the 4650U isn't terribly
> relevant. For price/performance ... oh, yes! And if I thought that the manufacturing cost for the
> CPU portion of the 4650U was substantially higher than that of the 2955U, then, okay.
>
> But I think that the CPU portion between these two parts is pretty much the same. Intel charges much
> more for the faster chip (which also has a better IGPU, I think), but only "because they can."
>
> The 2955U isn't cheaper to manufacture because it has hyper-threading
> disabled. Or because turbo boosting is disabled.
>
> juanrgo claimed very specifically: "Denver is an in-order VLIW+RA
> core that outperforms wide OoO cores such as Cyclone or Haswell."
>
> The only way to get to this claim is with nVidia's choice of a slow/cheap Haswell that Intel sells cheaply
> for market segmentation purposes. Which is bogus if the question is "How well can Haswell perform?"
>
>
I would beg to differ, rather strongly.
IF what we are trying to do is compare performance of different TECHNOLOGIES then we should take
effort to level the playing field where possible.
Now, if it is the execution approach of the core we wish to discuss, that would mean considering:
cache, transistor budget, power budget, IO performance (mainly memory bandwidth/latency),
implementation tricks (such as integrated memory controllers, hyperthreading, clock boost, etc)
which are not related directly to core function, and possibly development effort.
juanrgo (as much as I consider him a bit of a nutjob..) was, as you say, claiming things about CORE
performance, which I do believe matters.
If we are trying to make a purchasing decision comparison, of course, it is quite different.
However, to my mind RWT is a forum for discussion of technical merit, not one for procurement
decisions, no?
Intel, as we should all be well aware, has a depth of capability in x86 implementation and process
technology that is pretty hard to even approach.
I would say that it is reasonably fair to compare Denver with an intel cpu without hyperthreading,
clock boost, maximised cache, and the ability to select from a massive production base IF we are
discussion merit of core design, and IF we are comparing with a much smaller design house.
The reason being that when considering core technologies (rather than a specific implementation) we
should, to get any real meaning in comparison, consider what the cores would look like if the two
players were reversed for implementation.
ie: what would Denver look like if built by intels developers and fab tech?
What would Haswell look like from NVIDIA/TSMC?
I would suggest that Denver on Intels process, with hyperthreading, clock boost, intel caches, etc
WOULD in fact be a much more performant beast, and Haswell would take quite a hit from NVIDIA/TSMC.
Of course none of that impacts the procurement process, where you consider the products as available.
However it matters a lot when comparing the technologies at play in the core logical design.
I would suggest that no one has enough information on Denver just now to tell if it is a good idea
or not - perhaps we will never know due to business considerations. But it does at least look
interesting at a technological level.
> coppice (coppice.delete@this.dis.org) on January 8, 2015 11:38 pm wrote:
> > Mark Roulo (nothanks.delete@this.xxx.com) on January 8, 2015 6:37 pm wrote:
> > > juanrga (nospam.delete@this.juanrga.com) on January 7, 2015 6:03 am wrote:
> > > > I recall Aaron asking me why I believed that mixing VLIW with
> > > > RA was a good idea. Funny, because we know now that
> > > > Denver is an in-order VLIW+RA core that outperforms wide OoO cores such as Cyclone or Haswell.
> > >
> > > Have you not noticed that nVidia did not select the fastest 15W Haswell CPU they could?
> > >
> > > The 2955U Celeron is a 15W TDP part with no hyperthreading, a clockspeed
> > > of 1.4 GHz and 2 MB of L3 cache. Turbo-boost is disabled.
> > >
> > > The 4650U is *also* a 15W TDP Haswell part, but it has a base frequency
> > > of 1.7 GHz, 4 MB of L3 cache and will turbo to 3.3 GHz.
> > >
> > > Eyeballing the nVidia chart, the Denver (@ 2.5GHz?) beat
> > > the 2955U by about 10% on SpecInt2000. My guess is
> > > that a 4650U would beat the Denver chip by close to 2:1. When you lose by 2:1, you aren't "outperforming".
> >
> > The 4650U is a very expensive chip, while dual core denvers are going into tablets. Is that
> > a fair comparison? I guess winning, rather than fairness, was uppermost when the comparison
> > with the 2955U was made, but you have to look beyond one parameter, like watts.
>
> Since we are addressing the *performance* of the CPU, I think that the cost of the 4650U isn't terribly
> relevant. For price/performance ... oh, yes! And if I thought that the manufacturing cost for the
> CPU portion of the 4650U was substantially higher than that of the 2955U, then, okay.
>
> But I think that the CPU portion between these two parts is pretty much the same. Intel charges much
> more for the faster chip (which also has a better IGPU, I think), but only "because they can."
>
> The 2955U isn't cheaper to manufacture because it has hyper-threading
> disabled. Or because turbo boosting is disabled.
>
> juanrgo claimed very specifically: "Denver is an in-order VLIW+RA
> core that outperforms wide OoO cores such as Cyclone or Haswell."
>
> The only way to get to this claim is with nVidia's choice of a slow/cheap Haswell that Intel sells cheaply
> for market segmentation purposes. Which is bogus if the question is "How well can Haswell perform?"
>
>
I would beg to differ, rather strongly.
IF what we are trying to do is compare performance of different TECHNOLOGIES then we should take
effort to level the playing field where possible.
Now, if it is the execution approach of the core we wish to discuss, that would mean considering:
cache, transistor budget, power budget, IO performance (mainly memory bandwidth/latency),
implementation tricks (such as integrated memory controllers, hyperthreading, clock boost, etc)
which are not related directly to core function, and possibly development effort.
juanrgo (as much as I consider him a bit of a nutjob..) was, as you say, claiming things about CORE
performance, which I do believe matters.
If we are trying to make a purchasing decision comparison, of course, it is quite different.
However, to my mind RWT is a forum for discussion of technical merit, not one for procurement
decisions, no?
Intel, as we should all be well aware, has a depth of capability in x86 implementation and process
technology that is pretty hard to even approach.
I would say that it is reasonably fair to compare Denver with an intel cpu without hyperthreading,
clock boost, maximised cache, and the ability to select from a massive production base IF we are
discussion merit of core design, and IF we are comparing with a much smaller design house.
The reason being that when considering core technologies (rather than a specific implementation) we
should, to get any real meaning in comparison, consider what the cores would look like if the two
players were reversed for implementation.
ie: what would Denver look like if built by intels developers and fab tech?
What would Haswell look like from NVIDIA/TSMC?
I would suggest that Denver on Intels process, with hyperthreading, clock boost, intel caches, etc
WOULD in fact be a much more performant beast, and Haswell would take quite a hit from NVIDIA/TSMC.
Of course none of that impacts the procurement process, where you consider the products as available.
However it matters a lot when comparing the technologies at play in the core logical design.
I would suggest that no one has enough information on Denver just now to tell if it is a good idea
or not - perhaps we will never know due to business considerations. But it does at least look
interesting at a technological level.