What is PS4's 14+4 CU thing all about? *spawn

[almighty]

What you actually do with 4 CU's though? Certain PC Games take a dive bomb when running certain effects with compute....

Would you be able to do more with less on console because of more efficient coding?

 

[Rangers]

Good god not 14+4 again, the point of the original slide was just that there is a concept of diminishing returns, and some sort of knee in the performance curve given the rest of the system.

But the bolded part is still whats confusing and I guess what I wonder about. What exactly is the issue (s).

 

[Rangers]

AMD exec in a vrzone interview said the mantle sdk will go public in a year or so, so we'll be able to test 7850s and isolate individual CU and test whether there really is a drop off in benefit going from 14 to 18 additional CU for rendering.

I wish that somebody could do this test now. I also want somebody to test Microsoft's claim that a 6.6% clock increase is better than two more CU's.

I'm not sure how though. 7770 and 7790 are too dissimilar (bandwidth, 2 triangles versus 1).

But BTW the test you propose wouldn't tell you too much without it being in a actual PS4 testbed (impossible I assume).



What would be ideal would be a way to disable CU's on a Bonaire. Then test it with 12 CU's+6.6% clock and 14. No idea if that's possible, thinking no. But I guess all of that's off topic to this thread.

When would an 8 core Jaguar be available on desktop PC? That could also provide some interesting benchmarking.

Also in the end the results would be limited to the realm of PC software and only tell us so much. Still these would be tests I'd love to see.

 

[MrFox]

I hope we will eventually have an extensive post-mortem of Resogun, and see how much GPU they used for compute... it looks like the kind of game that has very different needs compared to a shooter or racing game. The brute processing must be heavily shifted to the GPU side.

I just realized that the consoles are coming out next month. I like saying "next month" now, it feels closer than 6 weeks.

 

[adev]

The PS4 can mask which CUs are used for certain jobs.

The PS4 also can reserve GPU memory (LDS/register space, etc.) on individual CUs for use by specific tasks.

The VSHELL has reserved memory on "several" CUs 100% of the time and I wouldn't be surprised if VSHELL tasks are masked to those CUs.

There's a graphics command queue dedicated to VSHELL and it can create high priority GPGPU tasks.

 

[BRiT]

What you actually do with 4 CU's though? Certain PC Games take a dive bomb when running certain effects with compute....

Would you be able to do more with less on console because of more efficient coding?

The PC doesn't have TTM technology. That's why the consoles can do more with less.

At least they didn't until AMD stepped up with Mantle. They're enabling lazy devs to do more with more.

 

[function]

I wish that somebody could do this test now. I also want somebody to test Microsoft's claim that a 6.6% clock increase is better than two more CU's.

MS didn't say that would be the case for every GCN based system, or even that it would be the case for every possible use case on the 1Bone, just that it was the case for the 1Bone software that they tested.

The none rop, fixed function stuff seemed to be the bigger isse based on today's DF article.

 

[patsu]

Heh heh, was waiting for dev leaks. Funny how my timing seems to be perrrfect every time.

The PS4 can mask which CUs are used for certain jobs.

The PS4 also can reserve GPU memory (LDS/register space, etc.) on individual CUs for use by specific tasks.

Are you a dev ?

Your claims...
Is the allocation/reservation of CUs and GPU memory for specific jobs done via the GNM API or something else ? How dynamic/flexible is it ?
I assume this is what Mark Cerny meant by a SPUR-like framework ?

How difficult is it to enforce/use ? How is it different from the regular wavefront scheduling ?

Are the launch games using this facility ?

The VSHELL has reserved memory on "several" CUs 100% of the time and I wouldn't be surprised if VSHELL tasks are masked to those CUs.

There's a graphics command queue dedicated to VSHELL and it can create high priority GPGPU tasks.

Who can issue commands to the VSHELL ? Only the OS ?

EDIT:
I vaguely recall 3dilettante suggested that the PS4 OS may reserve some GPU resources 100% of the time, perhaps for serving game/app requests ? What might those (high priority jobs) be ? ^_^

 

[astrograd]

How? By possibly being CPU bound? From what I read/heard from other developers, the Jaguar CPU is quite capable of handling large amounts of throughput. So that leaves system memory and bandwidth allocation, which seems PS4 has enough of - in spades.

Evidently Cerny and his team at Sony very much disagree. There would seem to be bottlenecks somewhere along the line to prevent effective utilization of those 4 CU's for purely graphical rendering tasks.

 

[patsu]

It's negative to console warriors. The XBox contingent would say PS4 only has 14 CUs for graphics versus 12 in XB1, and could go as far to say the other 4 CUs are for audio and stuff.

As a technical discussion, the most relevant question is, "what are the limiting factors in the rest of the GPU where 14 CUs would be optimal and 18 CUs a little too much for optimal graphics?"

It is unclear what that slides meant by balance. 14+4 suggests a relatively static arrangement, but during run-time, it could in fact be dynamic.

Even on PS3, the SPUs run assorted stuff at the same time to prevent stalling. Sometimes, they work on different frame from the RSX too.


As for possible bottlenecks, it would depend on the software behavior and what they hope to achieve. I don't know if we can arrive at a definite conclusion just by staring at the h/w specs here.

The PS3 is considered to be "unbalanced", but some developers still manage to surprise us.

 

[astrograd]

But the bolded part is still whats confusing and I guess what I wonder about. What exactly is the issue (s).

My guess is the best way to think about it is that the PS4 was built around 14 CU's for graphics rendering. That's what the system was balanced at and designed for in terms of graphics. Sony sees great potential in GPGPU compute and they accordingly gave devs 4 more CU's specifically to use for that purpose along with some tweaks to help accommodate them.

 

[patsu]

My guess is the best way to think about it is that the PS4 was built around 14 CU's for graphics rendering. That's what the system was balanced at and designed for in terms of graphics. Sony sees great potential in GPGPU compute and they accordingly gave devs 4 more CU's specifically to use for that purpose along with some tweaks to help accommodate them.

They will need to allocate not just additional CUs for GPGPU jobs. They will also need to plan for memory and bandwidth usage for these additional tasks. I doubt experienced developers in Sony's team will just sprinkle 4 more CUs just to hope for better GPGPU usage.

Hardware folks are very stingy people. They will do something to assure themselves that the additional h/w investments are worthwhile.

 

[astrograd]

They will need to allocate not just additional CUs for GPGPU jobs. They will also need to plan for memory and bandwidth usage for these additional tasks. I doubt experienced developers in Sony's team will just sprinkle 4 more CUs just to hope for better GPGPU usage.

Hardware folks are very stingy people. They will do something to assure themselves that the additional h/w investments are worthwhile.

Not sure what point you are trying to make here. Are you suggesting that they didn't do what I said? Because...I got that interpretation pretty much verbatim from what Cerny said on the matter. I'm sure they made various tweaks and modifications to their design to accommodate their GPGPU ambitions.

 

[Betanumerical]

My guess is the best way to think about it is that the PS4 was built around 14 CU's for graphics rendering. That's what the system was balanced at and designed for in terms of graphics. Sony sees great potential in GPGPU compute and they accordingly gave devs 4 more CU's specifically to use for that purpose along with some tweaks to help accommodate them.

But you can use GPGPU for graphics, not to mention as been previously mentioned where the cliff is and how it hits you is dependent on your engine and most likely exists for all consoles. This isn't a hard and fast rule that it will suddenly have 0 use for any CU's past 14. We should also point out that for metrics that aren't tied per CU (Bandwidth and ROP's, etc) that the PS4 has more per CU then desktop cards. So I can't really see how its 'balanced for 14 CU's' anywhere, can you show me?.

Not sure what point you are trying to make here. Are you suggesting that they didn't do what I said? Because...I got that interpretation pretty much verbatim from what Cerny said on the matter. I'm sure they made various tweaks and modifications to their design to accommodate their GPGPU ambitions.

Additional tweaks does not mean allocating 1/3 of your graphics for GPGPU permanently.

 

[patsu]

Cerny's comment is pretty vague. When he adds more ALUs, he may also beef up other parts to make sure GPGPU tasks can be fed properly. The CUs may only be part of his formula.

It is built with 18 CUs for both graphics and compute jobs. The developers decide how to juggle jobs between them via software. I don't know what performance targets he has in mind (1080p/60fps ? VR ? 3D ? etc.)

 

[taisui]

Cerny's comment is pretty vague. When he adds more ALUs, he may also beef up other parts to make sure GPGPU tasks can be fed properly. The CUs may only be part of his formula.

It is built with 18 CUs for both graphics and compute jobs. The developers decide how to juggle jobs between them via software. I don't know what performance targets he has in mind (1080p/60fps ? VR ? 3D ? etc.)

You can communicate to a new cybercity. Did you see the movie The Matrix? Same interface. Same concept. Starting from next year, you can jack into The Matrix! - Ken Kutaragi

 

[patsu]

Additional tweaks does not mean allocating 1/3 of your graphics for GPGPU permanently.

Yes, task allocation can be done via software anyway. No need to specify/limit it in hardware.

It should be possible for the devs to use all 18 CUs for graphics jobs anytime he feels like it. As for the pockets of inefficiency while running all 18 CUs, can we still stuff minor jobs in between to up the utility ? May be, may be not. Case by case. If a job's on the critical path before some parallel work, the dev may want to finish it first asap.

 

[Bigus Dickus]

I have a feeling that the reason people are concerned about the 14+4 is the implication that somehow the PS4 is no longer 50% more powerful graphically (but the reality is that it never was at any point). ;-)

Bingo. It seems very few actually misunderstood the intent of the 14+4 slide or Sony's own followup comments. Rather, the accusation that anyone daring to mention the 14+4 dodo bird has terribly misunderstood the intent seems meant to divert attention from that very intent. No one wants to hear, not even from Sony, that 50% more CU isn't going to make the PS4 a relative graphics monstour.

 

[itsmydamnation]

Bingo. It seems very few actually misunderstood the intent of the 14+4 slide or Sony's own followup comments. Rather, the accusation that anyone daring to mention the 14+4 dodo bird has terribly misunderstood the intent seems meant to divert attention from that very intent. No one wants to hear, not even from Sony, that 50% more CU isn't going to make the PS4 a relative graphics monstour.

this is complete crap. GPU resources are there to be used, the entire idea of "balance" changes when you change the context. They are execution resources there to do stuff, what ever that might be. Thats the beginning and the end of it, over time dev's will find the optimal use of execution resources for certain tasks with certain compromises.

i see the people trying to down play the difference is execution resources as the ones with the agenda.

 

[astrograd]

But you can use GPGPU for graphics, not to mention as been previously mentioned where the cliff is and how it hits you is dependent on your engine and most likely exists for all consoles.

If it was a vague as you suggest here Sony would never have bothered making it a bullet point for their dev presentations. Again, just because you lack the imagination and/or understanding of how their platform works in detail and as such can't fathom other bottlenecks for rendering doesn't somehow suggest Sony is wrong. I trust their dev docs and Cerny over forumites who suggest they know better.

This isn't a hard and fast rule that it will suddenly have 0 use for any CU's past 14.

Nobody suggested as much.

We should also point out that for metrics that aren't tied per CU (Bandwidth and ROP's, etc) that the PS4 has more per CU then desktop cards. So I can't really see how its 'balanced for 14 CU's' anywhere, can you show me?.

No. But I'm willing to bet Sony can.

Additional tweaks does not mean allocating 1/3 of your graphics for GPGPU permanently.

Read my post again. I said they seem to have designed a machine around 14 CU's for rendering and then added extra CU's specifically for their GPGPU ambitions...with additional tweaks to help accommodate that.