PlayStation 4 (codename Orbis) technical hardware investigation (news and rumours)

[3dilettante]

The job scheduling capability for GCN can provide a soft reservation without reserving a CU. It can either create some long-running work items that sit idling, or some arrangement can be made where the system permits it to assign itself a high priority status for job allocation.

The presence of a separate high-priority VSHELL graphics front end might be the way Sony wants to maintain system responsiveness for transitions to the non-game portion of the UI. The context never leaves the GPU, but sits there in a holding pattern in the secondary front end.
When the user pulls up the background UI, the high-priority graphics queue can force itself to the front of the line.

 

[Boxercide]

The fact that PS4 will have 4,5 Gb of direct memory + 0,5 of flexible (?)

It is 1 GB flexible, so up to 5.5 currently.

 

[Aeoniss]

The job scheduling capability for GCN can provide a soft reservation without reserving a CU. It can either create some long-running work items that sit idling, or some arrangement can be made where the system permits it to assign itself a high priority status for job allocation.

The presence of a separate high-priority VSHELL graphics front end might be the way Sony wants to maintain system responsiveness for transitions to the non-game portion of the UI. The context never leaves the GPU, but sits there in a holding pattern in the secondary front end.
When the user pulls up the background UI, the high-priority graphics queue can force itself to the front of the line.

Would this result in less GPU resources allocated to developers? I'm assuming it does.

 

[3dilettante]

The question would be "less than what?"
Less than if there were no background OS at all and the game was the only thing on the system, yes.
Less than if there were a static unit reservation or percentage cycles? Maybe not.

I can't definitively state what the exact scheme is, but a dedicated front end can be used so that the background context can wait off to the side until it is needed.
It can then inject itself into the job queues in an operation without the typical context switching or initialization steps that take so much time.

That would keep its burden pretty low until the user pulls up the background functionality, in which case the user has indicated the game is no longer the most important thing running.

 

[bkilian]

Did PS3 and 360 have such a GPU reservation? (I honestly don't know..)

And what would be a reasonable allottment of resources for the OS GPU wise? Or is it even necessary?

Yes for 360, in a way. I don't recall the actual numbers. They were a little nebulous since there was no way to carve off a chunk of GPU power in a neat way. Might have been something like 1ms every 16. PS3 I am not familiar with.

 

[MrFox]

It is 1 GB flexible, so up to 5.5 currently.

The 1GB is split into .5 real physical memory and .5 virtual memory (which we should basically ignore).
In practical terms it's 5GB guaranteed for games.

That is, if that rumor is true, and up to date.

 

[firelink]

I don't want to bring back the GDDR5 latency debate, but I have a thought on it and just wanted a Yes/No if I am on to something.

A lot of people refute latency issues due to documents from Hynix. However, there is something else in that documentation that I think might lead somewhere.

If you go here, http://www.anandtech.com/show/2841/12, you see this:

To combat this, GDDR5 memory controllers can perform basic error detection on both reads and writes by implementing a CRC-8 hash function.

If the PS4 memory controller is anything like normal ones, it most likely also requires a CRC check. On the Hynix documentation, they show the relative latency of these checks:

Programmable CRC WRITE latency = 7 to 14 tCK

So then, shouldn't this CRC check latency be added to their other latency, changing

Programmable CAS latency: 5 to 20 tCK

to

Programmable CAS latency: 12 to 34 tCK

?

Meaning a 100% to 300% increase compared to DDR3?

http://www.skhynix.com/products/graphics/view.jsp?info.ramKind=26&info.serialNo=H5GQ2H24AFR

http://www.hynix.com/products/computing/computing_sub.jsp

Am I on to something, or just blowing smoke where there is no fire?

 

[Gitaroo]

found this on magicbox

http://www.the-magicbox.com/

"Vblank Entertainment's PS4 developer head Brian Provinciano has told the media that the report froM Digital Foundry about PS4 only has 3.5GB of memory available for games, is "absolutely false". While not going into details, he mentioned that games can access a combination of both "Direct Memory" and "Flexible Memory", which makes up the 8GB RAM in the system. There are games in development that uses 6GB of RAM."

edit: they got the 3.5 gig backward, it was 3.5 for OS, not games. I googled it up and also found, some articles about this. for example

http://playstationgang.com/update-ps4-5gb-of-ram-is-absolutely-false/

not sure how reliable.

 

[warb]

The claim was 5GB, not 3.5GB.

 

[Gitaroo]

The claim was 5GB, not 3.5GB.

The article quote the wrong number, they meant 3.5 gb for os. VBlank dev stated everything is wrong from DF, there are ps4 games in developmentthat uses 6 gb already, indicate 2gb ram for os.

 

[zupallinere]

... there are ps4 games in development that uses 6 gb already, indicate 2gb ram for os.

Those 6gb games could be a year or 2 out so I'm fairly sure that when the ps4 launches it will share similar reserves to the xb1 which make sense at this point.

 

[(((interference)))]

That's exactly what the VGLeaks doc stated months ago with regards to 14+4 (the +4 provide a "minor boost if used for rendering"). So there's something limiting about the design of the PS4 or the resolution this platform is targeting (1080p) that makes using more than 14 CUs 'wasteful.' The rendering performance must have a significant dip in efficiency for this to surface in the leaked docs and mildly confirmed by Cerny in his interview.

Basically, as Cerny was hinting at the PS4 as a design is a bit ALU heavy.
So there's a point where you get diminishing returns from using additional CUs for rendering and get more bang for your buck using them for things like compute.

 

[expletive]

Basically, as Cerny was hinting at the PS4 as a design is a bit ALU heavy.
So there's a point where you get diminishing returns from using additional CUs for rendering and get more bang for your buck using them for things like compute.

Thanks, any thoughts on what other part of the system are being outstripped by the high number of CUs? Or is it just the relatively low resolution of HDTV not pushing it hard enough?

 

[(((interference)))]

Thanks, any thoughts on what other part of the system are being outstripped by the high number of CUs? Or is it just the relatively low resolution of HDTV not pushing it hard enough?

No, unfortunately I don't know, but my source was guessing it could be related to cache sizes, register file sizes as well as things like bandwidth, ROPs etc.

The 14+4 numbers are probably based on how similar PC hardware performs with modern game engines (like Crysis) as that's the only source of data AMD would have had to draw their numbers from.

So, I guess if your workloads/rendering techniques/engine significantly differ from what's done on current AAA PC games it might not apply

 

[temesgen]

No, unfortunately I don't know, but my source was guessing it could be related to cache sizes, register file sizes as well as things like bandwidth etc.

The 14+4 numbers are probably based on how similar PC hardware performs with modern game engines (like Crysis) as that's the only source of data AMD would have had to draw their numbers from.

So, I guess if your workloads/rendering techniques/engine significantly differ from what's done on current AAA PC games it might not apply

Serious question - do you think cloud computing or GPU compute on PS4 will have more of an impact in this upcoming generation? Personally I'd guess cloud computing but its more a hunch....

 

[(((interference)))]

Serious question - do you think cloud computing or GPU compute on PS4 will have more of an impact in this upcoming generation? Personally I'd guess cloud computing but its more a hunch....

I honestly have no idea what will happen but would think GPU compute since that is what Sony has been talking up and have customised the system in that direction.

 

[temesgen]

I honestly have no idea what will happen but would think GPU compute since that is what Sony has been talking up and have customised the system in that direction.

Interesting, I've been under the impression that GPU compute is somewhat difficult and honestly I was thinking the design decision WRT to number of ROPs and CUs was due to the nature of hardware (they were either going to have 16 or 32) and by going with 32 by default they knew they had some additional processing available based off what we know about the rest of the design (CPU and Bandwidth).

If I'm right it seems to me that the team looked at the additional processing power and wanted to find ways to use it since it was already there. I'm confident the first party studios will use it not sure if 3rd parties will touch it. Perhaps it would be a good exercise to look at how GPU compute could be useful.

Edit: Now I'm wondering if the number of CUs has to follow the number of ROPs? Is this accurate?

 

[dumbo11]

If I'm right it seems to me that the team looked at the additional processing power and wanted to find ways to use it since it was already there. I'm confident the first party studios will use it not sure if 3rd parties will touch it. Perhaps it would be a good exercise to look at how GPU compute could be useful.

I think Cerny mentioned during one of the early interviews that Sony primarily expect middleware to use GPU compute.

 

[forumaccount]

Serious question - do you think cloud computing or GPU compute on PS4 will have more of an impact in this upcoming generation? Personally I'd guess cloud computing but its more a hunch....

Someone will have to do something interesting with cloud computing before it becomes worth talking about. Microsoft have turned it into a running joke.

There are plenty of things on a server end that could potentially benefit from cloud computing (at a cost, apparently, if I understood all of the Titanfall stuff correctly), and especially the flexible dedicated server clouds will be very nice. But someone has to invent something that's both latency-tolerant and "interesting" to do on the client side before any of this cloud computing stuff gets taken seriously or is at all comparable to something like GPU compute.

If I'm right it seems to me that the team looked at the additional processing power and wanted to find ways to use it since it was already there. I'm confident the first party studios will use it not sure if 3rd parties will touch it. Perhaps it would be a good exercise to look at how GPU compute could be useful.

Of course 3rd parties will use it. Don't buy into this prevailing game development mythos where only first parties have any clue what they're doing. It's not at all grounded in reality.

 

[mozmo]

I think Cerny mentioned during one of the early interviews that Sony primarily expect middleware to use GPU compute.

I still fail to see how GPGPU is going to work with the coherent system memory. It looks like any GPU access to cached system pages needs to go through the onion bus, that's only 10GB/s. I don't see how you'll be able to run jobs on GPGPU and share cpu data structures at the same time without hitting bandwidth limits eventually.

GPGPU is going to involve a lot of tricks and copies to work imo.

Also don't these jaguar cores have AVX support, so they're pretty wide 8 floats, might be easier to just use AVX for vector heavy calcs. You'll enjoy more coherent memory bandwidth and you have a lot more L2 space vs the GPU.