By: Michael S (already5chosen.delete@this.yahoo.com), May 13, 2013 12:30 am
Room: Moderated Discussions
Doug S (foo.delete@this.bar.bar) on May 12, 2013 8:22 pm wrote:
> mpx (mpx.delete@this.nomail.pl) on May 12, 2013 12:04 pm wrote:
> > Doug S (foo.delete@this.bar.bar) on May 11, 2013 10:49 am wrote:
> > > What the hell would people be doing on their phone that they need
> > > 8 threads?
> >
> > A general characteristics of CPU - memory is high-latency - is the main driver for
> > SMT. Smaprthones of the past were less bothered by it because of sub-1GHz frequencies.
> > Now, with frequencies approaching 2GHz they would benefit from SMT more.
>
>
> You still need to be actually doing something that maxes out your four real cores while still needing
> more resources. On a phone. I'm talking a real world task that would be done more than a fraction of
> the userbase, so no talk about a couple games that can theoretically be a tiny bit faster with 4C/8T
> versus 4C, or people running a browser with >4 tabs in a process-per-tab model where even the inactive
> ones are maxing out the CPU (Flash could do this, but luckily that abomination is rapidly dying)
>
May be, in your world flash is dying.
In mine it is alive and well.
And I prefer it this way, if for nothing else them because it gives me single point of control over useless animations.
> I agree that as frequencies increase, you are more cycles away from main memory.
> But so what? Until you're doing more things that are maxing out all cores on
> human timescales SMT won't make any difference for the user experience.
>
> Your argument works equally well to reason that 4 cores on a phone isn't enough, they should have 6, or
> 8, or 10. Sure SMT is less expensive than adding another real core, but it isn't free either because each
> core needs more execution resources than the it otherwise would have had, and more cache than it otherwise
> would have had, and another full register file.
No, it does not have to had "another full register file".
Haswell integer register file has 168 entries. Hypothetical Haswell-without-SMT for the same single-thread performance will need 152 entries. That's 10% smaller, not twice.
> Adding a second core made a real difference to many people's
> workloads, both on desktops to a limited extent, on phones. Going to quad core is hardly noticeable except
> for niche uses, so there's no point in going beyond that except for marketing purposes.
What are you arguing for? That in phones 2C/4T had no advantage over 2C/2T?
> mpx (mpx.delete@this.nomail.pl) on May 12, 2013 12:04 pm wrote:
> > Doug S (foo.delete@this.bar.bar) on May 11, 2013 10:49 am wrote:
> > > What the hell would people be doing on their phone that they need
> > > 8 threads?
> >
> > A general characteristics of CPU - memory is high-latency - is the main driver for
> > SMT. Smaprthones of the past were less bothered by it because of sub-1GHz frequencies.
> > Now, with frequencies approaching 2GHz they would benefit from SMT more.
>
>
> You still need to be actually doing something that maxes out your four real cores while still needing
> more resources. On a phone. I'm talking a real world task that would be done more than a fraction of
> the userbase, so no talk about a couple games that can theoretically be a tiny bit faster with 4C/8T
> versus 4C, or people running a browser with >4 tabs in a process-per-tab model where even the inactive
> ones are maxing out the CPU (Flash could do this, but luckily that abomination is rapidly dying)
>
May be, in your world flash is dying.
In mine it is alive and well.
And I prefer it this way, if for nothing else them because it gives me single point of control over useless animations.
> I agree that as frequencies increase, you are more cycles away from main memory.
> But so what? Until you're doing more things that are maxing out all cores on
> human timescales SMT won't make any difference for the user experience.
>
> Your argument works equally well to reason that 4 cores on a phone isn't enough, they should have 6, or
> 8, or 10. Sure SMT is less expensive than adding another real core, but it isn't free either because each
> core needs more execution resources than the it otherwise would have had, and more cache than it otherwise
> would have had, and another full register file.
No, it does not have to had "another full register file".
Haswell integer register file has 168 entries. Hypothetical Haswell-without-SMT for the same single-thread performance will need 152 entries. That's 10% smaller, not twice.
> Adding a second core made a real difference to many people's
> workloads, both on desktops to a limited extent, on phones. Going to quad core is hardly noticeable except
> for niche uses, so there's no point in going beyond that except for marketing purposes.
What are you arguing for? That in phones 2C/4T had no advantage over 2C/2T?