By: Joel (joel.hruska.delete@this.gmail.com), April 29, 2015 3:12 pm
Room: Moderated Discussions
Klimax (danklima.delete@this.gmail.com) on April 29, 2015 12:58 pm wrote:
> Joel (joel.hruska.delete@this.gmail.com) on April 29, 2015 12:26 pm wrote:
> > David Hess (davidwhess.delete@this.gmail.com) on April 29, 2015 8:38 am wrote:
> > > Pierre (Boutoukoat.delete@this.yahoo.fr) on April 29, 2015 4:38 am wrote:
> > > >
> > > > Dissipated power is O(frequency) and O(squared voltage). The article mention the voltage could be reduced
> > > > by 200 millivolts, e.g. go from 0.9 to 0.7 volts. This mean a reduction of power of 40%, and in layman
> > > > terms, 40% more battery life on a processor 4 times cheaper. Reasoning at constant power and ignoring
> > > > o a lot of details, frequency could increase by 40% if voltage goes from 0.9V to 0.7V. However, most of
> > > > the perceived performance out of a processor comes from memory speed (unchanged ...) , parallelism and
> > > > integration of co-processors like GPU and NICs (more transistors), and software benchmarks.
> > >
> > > Power density is also becoming a limiting factor. Heat pipes helped here however they have their own
> > > hard power density limits and now require heat spreaders between the CPU and evaporator. If you reduce
> > > the power by 40% but have the same number of transistors in 1/4 of the area, then the power density
> > > increases by 160%. Raising the frequency by 40% and keeping the power constant would increase the power
> > > density by 300%. Either the junction to case thermal resistance has to be lowered (diamond?) or a higher
> > > junction temperature is needed which will adversely affect reliability and leakage.
> > >
> > > > For the last 10 years, frequency of desktops for gamers has not increased (peak around 4 Ghz on
> > > > the most expensive Intel CPUs). It looks unlikely to change in the near future, and the trend
> > > > to add more cores will continue. I hope this helps to answer your obsolete question "how faster
> > > > ?" typical from the 90's. Right questions now are "how cheaper ?" , "how longer ?" ...
> > >
> > > AMDs current CPUs top out in the 4 GHz range as well. Adding more cores to take advantage of higher
> > > density processes is going to run up against power density limits and most applications including
> > > games cannot take advantage of more cores anyway. Do any games make use of more than 2 cores?
> > >
> >
> > Yes. There are many modern titles that can make some use of up to four cores, though
> > not much scales past that. DX12 is expected to vastly improve this scenario; it
> > allows for effective multi-threading in ways that DX11 never supported.
> >
> > Early DX12 games and applications can take full advantage of all eight
> > cores on a Haswell-E system and see performance gains from doing so.
>
> There won't be likely any real jump. Those early examples tend to be badly written benchmarks
> at best intended to present severely contrived case, which in practice won't exist.
> (Not to mention that current engines don't make use of half of DX11 anyway) Also DX12
> and co makes very bad trade offs regarding future changes in GPUs. (IMO)
Yeah, every developer I've spoken to -- and I've spoken to more than a few -- disagrees with that.
Mantle provides real evidence as well. Average performance uptick for AMD APUs using Mantle is about 13% with another 13% improvement in performance per watt. DX12 allows for asynchronous compute (sharing compute workloads between two vastly different GPUs) as well as a far more diverse set of multi-GPU configurations. For the first time, workloads don't *have* to be duplicated across VRAM.
No, I don't expect the first DX12 titles to be 50% faster than DX11 games, but low-overhead APIs allow a lot of options that previous APIs didn't.
If this wasn't the case, why would *everyone* -- Apple, Khronos, and Microsoft -- have cumulatively leapt for low-overhead API designs? If AMD didn't have the core of a good idea with Mantle, nobody would've paid attention to it. Instead we get Vulkhan based on Mantle and DX12 offering essentially identical features alongside Apple's Metal.
> Joel (joel.hruska.delete@this.gmail.com) on April 29, 2015 12:26 pm wrote:
> > David Hess (davidwhess.delete@this.gmail.com) on April 29, 2015 8:38 am wrote:
> > > Pierre (Boutoukoat.delete@this.yahoo.fr) on April 29, 2015 4:38 am wrote:
> > > >
> > > > Dissipated power is O(frequency) and O(squared voltage). The article mention the voltage could be reduced
> > > > by 200 millivolts, e.g. go from 0.9 to 0.7 volts. This mean a reduction of power of 40%, and in layman
> > > > terms, 40% more battery life on a processor 4 times cheaper. Reasoning at constant power and ignoring
> > > > o a lot of details, frequency could increase by 40% if voltage goes from 0.9V to 0.7V. However, most of
> > > > the perceived performance out of a processor comes from memory speed (unchanged ...) , parallelism and
> > > > integration of co-processors like GPU and NICs (more transistors), and software benchmarks.
> > >
> > > Power density is also becoming a limiting factor. Heat pipes helped here however they have their own
> > > hard power density limits and now require heat spreaders between the CPU and evaporator. If you reduce
> > > the power by 40% but have the same number of transistors in 1/4 of the area, then the power density
> > > increases by 160%. Raising the frequency by 40% and keeping the power constant would increase the power
> > > density by 300%. Either the junction to case thermal resistance has to be lowered (diamond?) or a higher
> > > junction temperature is needed which will adversely affect reliability and leakage.
> > >
> > > > For the last 10 years, frequency of desktops for gamers has not increased (peak around 4 Ghz on
> > > > the most expensive Intel CPUs). It looks unlikely to change in the near future, and the trend
> > > > to add more cores will continue. I hope this helps to answer your obsolete question "how faster
> > > > ?" typical from the 90's. Right questions now are "how cheaper ?" , "how longer ?" ...
> > >
> > > AMDs current CPUs top out in the 4 GHz range as well. Adding more cores to take advantage of higher
> > > density processes is going to run up against power density limits and most applications including
> > > games cannot take advantage of more cores anyway. Do any games make use of more than 2 cores?
> > >
> >
> > Yes. There are many modern titles that can make some use of up to four cores, though
> > not much scales past that. DX12 is expected to vastly improve this scenario; it
> > allows for effective multi-threading in ways that DX11 never supported.
> >
> > Early DX12 games and applications can take full advantage of all eight
> > cores on a Haswell-E system and see performance gains from doing so.
>
> There won't be likely any real jump. Those early examples tend to be badly written benchmarks
> at best intended to present severely contrived case, which in practice won't exist.
> (Not to mention that current engines don't make use of half of DX11 anyway) Also DX12
> and co makes very bad trade offs regarding future changes in GPUs. (IMO)
Yeah, every developer I've spoken to -- and I've spoken to more than a few -- disagrees with that.
Mantle provides real evidence as well. Average performance uptick for AMD APUs using Mantle is about 13% with another 13% improvement in performance per watt. DX12 allows for asynchronous compute (sharing compute workloads between two vastly different GPUs) as well as a far more diverse set of multi-GPU configurations. For the first time, workloads don't *have* to be duplicated across VRAM.
No, I don't expect the first DX12 titles to be 50% faster than DX11 games, but low-overhead APIs allow a lot of options that previous APIs didn't.
If this wasn't the case, why would *everyone* -- Apple, Khronos, and Microsoft -- have cumulatively leapt for low-overhead API designs? If AMD didn't have the core of a good idea with Mantle, nobody would've paid attention to it. Instead we get Vulkhan based on Mantle and DX12 offering essentially identical features alongside Apple's Metal.