New consoles coming with low-clocked AMD x86. Can we now save moneys on our PC CPUs?

[Exophase]

Has 1.6GHz actually been confirmed? For Sony as well as MS?

The cost for PC ports isn't really because Windows uses CPU time, but a matter of API and abstraction, for instance making things like draw call count more sensitive. That probably varies a lot from game to game.

Non-APUs may disappear from the market, but it'll be a while before their GPU is as powerful as the SoC in PS4. Probably nothing on 32nm or even 28nm from GF. It could also be a while before they have access to memory as fast as Durango's, nevermind Orbis. Who knows if and when embedded fast RAMs will become standard for APUs, and how much it'll mitigate the difference.

It sounds like you're anticipating using the GPU part of an APU for compute and using a discrete GPU for graphics. I don't think this is going to be a very popular idea for PC ports..

 

[pjbliverpool]

You'll have to look to next gen consoles to see GPGPU physics take off. Perhaps someday, if Intel embraces GPU compute like AMD and make their integrated graphics more compute capable, then we'll see it make some headway in the PC space. But as long as it has to go over the PCIE bus, then it's going to be mostly a non-starter on PC.

But there's always the option of running the same GPGPU code that the consoles are running on the PC's CPU. Anything from Sandybridge upwards (and that includes AMD's Bulldozer FX) is significantly more powerful in SIMD terms than the Orbis CPU so should have a lot extra to spare on GPGPU. Whether thats enough to keep up with the consoles entirely depends on how much power they allocate to gameplay effecting GPGPU. Obviously none gameplay effecting GPGPU can be handed over to the discrete GPU.

In terms of Intels GPU compute ability, it sounds like it's going to be pretty capable from Haswell upwards so maybe leveraging that capability in systems that support it would be a faster alternative to running the code on the CPU. The only other option if one of those solutions isn't feasible is either to dial down the effects or skip the PC port altogether.

 

[pjbliverpool]

Non-APUs may disappear from the market, but it'll be a while before their GPU is as powerful as the SoC in PS4. Probably nothing on 32nm or even 28nm from GF. It could also be a while before they have access to memory as fast as Durango's, nevermind Orbis. Who knows if and when embedded fast RAMs will become standard for APUs, and how much it'll mitigate the difference.

I don't think this is really a desirable option for PC gamers. Even once parity is achieved it will be a massive step backwards in overall performance. What we need is a solution to keep up in GPGPU while retaining the power of discrete GPU's. So either go with very beefy SIMD CPU's or leverage onboard GPU's.

It sounds like you're anticipating using the GPU part of an APU for compute and using a discrete GPU for graphics. I don't think this is going to be a very popular idea for PC ports..

This is something I'm hoping will take off so I'd be interested on hearing your views as to why this won't be popular? AMD have suggested this useage model in one of their HSA slides but they are obviously not an objective observer!

 

[pjbliverpool]

Isn't this derailing a bit?

What I would like to know is what desktop CPU would be the equivalent to an 8-core Jaguar at 1.6GHz.

Based on the benchmarks I've seen comparing bobcat to modern architectures I'd guess something along the lines of an i3 around 3.2 Ghz?

Then

we could see how powerful the iGPU in the desktop APU ( HSA architecture) would need to be to keep up with the GPGPU operations in the PS4's APU. Of course, I don't think it has to be an APU with 18 GCN CUs @800MHz because the PS4 is never going to use them all at the same time for computing.

It depends completely on how much power the consoles allocate to gameplay effecting GPGPU but if you consider the top end Kaveri will have around 50% of the GPU flops of Orbis and around 75% of Durango then I'm sure Kaveri will be more than enough!

Haswell GT2 is much lower, I can't remember the exact figures but I think it's only about 1/6th of Durango's GPU. But then the CPU part of Haswell is around 1/3rd Durango's GPU so between them it should be more than enough.

 

[AlexV]

...AMD's Bulldozer FX) is significantly more powerful in SIMD terms than the Orbis CPU

Umm, how did you arrive at that conclusion? Is the assessment per cycle? Because per-cycle that's not really the case, unless you start looking at BD's FMA...but the world is not exactly a FMA fest, as is easy to see in practice...

 

[Exophase]

This is something I'm hoping will take off so I'd be interested on hearing your views as to why this won't be popular? AMD have suggested this useage model in one of their HSA slides but they are obviously not an objective observer!

For one thing, the API support isn't really there, and of course it has to be something Intel gets on board with in addition to MS (w/DX). Intel's APUs will need to get more compute friendly in general. And IGPs are still designed to display graphics out of dedicated display port. As the baseline for systems with APUs and without discrete GPUs will rise I don't think game developers will want to require discrete GPUs. It's not too hard to scale down graphics to lower resolutions and quality to support IGPs that are below console quality. But requiring that they also give up part of their power in compute is another story. This sounds like a pretty serious load balancing scenario.

Relying on low latency compute may also require different optimization patterns for different APUs..

If it's there it'll probably be in addition to using the CPU, since a lot of systems will have more spare CPU power than IGP power.

 

[pjbliverpool]

Umm, how did you arrive at that conclusion? Is the assessment per cycle? Because per-cycle that's not really the case, unless you start looking at BD's FMA...but the world is not exactly a FMA fest, as is easy to see in practice...

Not per cycle, but overall it obviously is if you're talking about 4 module parts. Bulldozer FX runs at a much higher clockspeed than Jaguar so at 4Ghz it's peaking at 256GFLOPs with FMA. Without taking advantage of FMA whatsoever it's still peaking at 128GFLOPS which is both higher and a lot more flexible than 8 Jaguars at 1.6ghz. The reality will be somewhere inbetween depending on how FMA heavy the code is.

 

[Silent_Buddha]

What about the fact that GPUs are limited in what they are fast at? Branchy code is bad, among other things. For example some physics simulation has not been feasible because of this. Or H.264 encoders that only do the minimum because some things wouldn't work well. Now obviously the latest GPUs are better but still ....

Plus whenever I've used Physx it's not as if there's no performance impact. Lol. It's more like a performance implosion.

True, but each generation of GPU has improved upon those things. Assuming you don't have to deal with the PCIE bus that is. Which unfortunately is currently a huge impediment to GPGPU based physics.

And we're talking about Physics simulations. Many of which are particularly well suited to GPGPU as long as you can remove the penalty of having to communicate over PCIE to both main memory as well as the CPU. Those that aren't useful on the PC side are generally not useful due to PCIE bandwidth and more importantly latency.

But there's always the option of running the same GPGPU code that the consoles are running on the PC's CPU. Anything from Sandybridge upwards (and that includes AMD's Bulldozer FX) is significantly more powerful in SIMD terms than the Orbis CPU so should have a lot extra to spare on GPGPU. Whether thats enough to keep up with the consoles entirely depends on how much power they allocate to gameplay effecting GPGPU. Obviously none gameplay effecting GPGPU can be handed over to the discrete GPU.

In terms of Intels GPU compute ability, it sounds like it's going to be pretty capable from Haswell upwards so maybe leveraging that capability in systems that support it would be a faster alternative to running the code on the CPU. The only other option if one of those solutions isn't feasible is either to dial down the effects or skip the PC port altogether.

Possibly. I'm not up on exactly what CPU you would need to be able to match 500+ GFLOPs (PS4 CPU + 4 CUs). And that's assuming a developer doesn't decide to use even more CUs for physics calculations if they want to go balls to the walls with physics calcs. Also the PS4 CPU isn't having to deal with much related to the OS as Windows does. Even their video encode for live video streaming appears to be using a separate dedicated chip rather than using any CPU or GPU resources.

I think it's much more reasonable to expect that a non-APU on the PC side would instead get reduced physics simulations.

And yes, hopefully, Haswell and up will be competitive with AMD's APUs with regards to GPGPU.

Regards,
SB

 

[pjbliverpool]

For one thing, the API support isn't really there, and of course it has to be something Intel gets on board with in addition to MS (w/DX). Intel's APUs will need to get more compute friendly in general. And IGPs are still designed to display graphics out of dedicated display port. As the baseline for systems with APUs and without discrete GPUs will rise I don't think game developers will want to require discrete GPUs. It's not too hard to scale down graphics to lower resolutions and quality to support IGPs that are below console quality. But requiring that they also give up part of their power in compute is another story. This sounds like a pretty serious load balancing scenario.

Relying on low latency compute may also require different optimization patterns for different APUs..

If it's there it'll probably be in addition to using the CPU, since a lot of systems will have more spare CPU power than IGP power.

Thanks for the unput, I guess we'll have to see how things turn out then!

 

[Blazkowicz]

Has 1.6GHz actually been confirmed? For Sony as well as MS?

The architecture puts a ceiling on clocks at maybe 2GHz.
With AMD saying the changes allows +10% clock and the fastest Bobcat at 1.7GHz (E-450) I'd say a clock like 1.8GHz or 1.83GHz is likely on consoles.

 

[Alexko]

The architecture puts a ceiling on clocks at maybe 2GHz.
With AMD saying the changes allows +10% clock and the fastest Bobcat at 1.7GHz (E-450) I'd say a clock like 1.8GHz or 1.83GHz is likely on consoles.

To be precise, that's +10% on the same process. So if 28nm brings an extra 10%, that's enough to reach 2GHz.

 

[homerdog]

With both consoles having 8 threads, will i7 CPUs finally show a real benefit over i5 CPUs in games?

 

[pjbliverpool]

With both consoles having 8 threads, will i7 CPUs finally show a real benefit over i5 CPUs in games?

Current consoles have 6+ threads so I doubt it.

 

[DJ12]

They wont both have 8 threads. One has 3 cores reserved the other is said to have none reserved but it's probably wise to assume it will have at least 1

 

[Pete]

4 @ 3GHz ~= 8 @ 1.6GHz, no (ignoring reserved threads, IPC, micro-stutter, etc)? Dunno how big a role L2 cache would play in the comparison (ie, each Jaguar thread has its own L2 vs. two i5 threads "sharing" one that's twice as big).

 

[Exophase]

They wont both have 8 threads. One has 3 cores reserved the other is said to have none reserved but it's probably wise to assume it will have at least 1

Now that consoles have a ton of cache coherent, general purpose CPUs running on a competent OS this core reserving nonsense makes zero sense. Even reserving one core. What possible justification could there be for reserving three?

 

[DJ12]

If it has windows 8 running at all times or a stripped version of it I guess it makes sense?

 

[Exophase]

If it has windows 8 running at all times or a stripped version of it I guess it makes sense?

Does it? Does Windows 8 reserve one of your cores on your home PC? Does it reserve nearly half of them?

 

[Entropy]

With both consoles having 8 threads, will i7 CPUs finally show a real benefit over i5 CPUs in games?

No.
The Core i7 offers hyper threading, which, to put it in simple terms, is a way of increasing utilization of the 4 physical cores in highly threaded situations. (Or put another way, a means to reduce the impact of latency.) It is not equivalent to having 8 physical processors. At best it will be some small fraction faster than its non-hyperthreading siblings. Typically, as now, the differences will be negligeable.

Not that having 8 processors implies twice the performance of having four, in the overwhelming majority of cases you will not see an overall linear increase.

Anyone with a PC CPU from the latest generation is sitting pretty for the rest of this decade, as far as console game ports go. A single Ivy bridge core is roughly three times as fast or more (at typical clocks) as a Bobcat core at 1.6 GHz. Where the consoles may show a clear advantage is in CPU to GPU communication. Changing PC CPU won't help much with that though, unless it is to a future APU.

 

[DJ12]

Does it? Does Windows 8 reserve one of your cores on your home PC? Does it reserve nearly half of them?

No, but if I am doing intensive CPU tasks, with high priority or real time set I have no luck doing anything else with it either.