New dev tools ? PlayStation 3 Edge

[patsu]

The graph at the top in the movie shows CPU usage, but I don't know how to read it! Were the 5 darker threads 5 SPUs? Or are the lines each aspects of Edge distributed over Cell?

Can any devs comment on what advantages/disadvantages this is bringing to graphics? It's unclear how this compares to a standard GPU-centric solution. eg. Are those 1.4 million triangles all rendered, versus perhaps 2 million triangles per frame of which half are not seen on a conventional GPU?

Ah... so many questions, so little time.

PPU % usage would be interesting to track too.

I saw those graphs... am inclined to interpret it as 60-70% utililization but I really want someone to confirm. My eyes are not too good with colors.

EDIT: Actually I think the PPU usage (both hardware threads) are displayed in the chart as well. They are like what... 70 - 80% ?

 

[Jesus2006]

Well considering in theory Xenos can have 48 vertex processors why is this such a suprise?

What does RSX have? 8?

However, 1 RSX Vertex Shader > 1 Xenos "Vertex Shader" unit afaik...

 

[Shifty Geezer]

I saw those graphs... am inclined to interpret it as 60-70% utililization but I really want someone to confirm. My eyes are not too good with colors.

My eyes have no problems with colours, and I can't make head nor tail of 'em! The actual colours they describe are hard to pin down due to off-screen capture. It's more 'pale' and 'dark' then cyan and blue.

 

[Rangers]

I think we can all retrospectively agree that Joker's poositions were one born of a newness to the industry and a lack of experience. His positions were fairly obliterated halfway into that thing. That was a good thread though, and I do wonder where he's gone to, as more developers is always nice to have around.

He basically said all the stuff that is now being proven by this thread (that you need to use Cell to help to reach Xenos performance, etc). And it was the first time I'd heard of it, at least. I'd say he's vindicated..

 

[patsu]

Well... he said a few things and generalized from them. I remember:
* RSX needs Cell to help in culling while Xenos can do it by itself
* PS3 has a broken scaler (so only software scaling is possible).
* Xenos can always outperform RSX because of larger unified and unused memory, techniques that benefits RSX can also benefits Xenos, etc...

I think only the first one is on the mark. The others were disputed in this thread: http://forum.beyond3d.com/showthread.php?t=37033


At the time of writing, joker454's team only used 1 SPU while first and second parties were embracing Cell, and suggested that the official PS3 developer forum might be a better platform for further discussion (http://forum.beyond3d.com/showpost.php?p=899962&postcount=141).

Did anyone attend the GDC RSX session ?

 

[Laa-Yosh]

If that is the same technology being used in the game too, then that's a pretty cool insight. They mention that the scene features 768 people and 50 cars. They say it's 720p, with anti-aliasing, with 1.4m polygons/frame (after geometry processing on the SPUs) at 60fps.

I wonder if the game can devote similar resources to graphics only... although it can probably benefit a lot more from lowlevel optimizations then this quick techdemo.

 

[Titanio]

I wonder if the game can devote similar resources to graphics only... although it can probably benefit a lot more from lowlevel optimizations then this quick techdemo.

Apparently someone asked how many SPUs that was using, and it turns out it was all of them. Although, of course, there's the issue of optimisation/utilisation to balance that out if the rest of the game were to need more than the PPU.

I think, however, there are games in the pipeline matching and exceeding this demo anyway. Think of Heavenly Sword - also 720p with AA, 2-3m polys per frame at 30fps (so a little less geometry than this, but..), 2,000+ enemies in some armies (versus less than 1,000 people here, who don't have any physics, AI etc. applied as they do in HS). So I think there are games out there that prove that you can do this and more, and without spending all your SPUs on geometry processing, but I think the idea was to show how quickly you could put something like that together using the tools.

 

[nAo]

HS does not use any SPUs to do vertex shading and/or to pre cull primitives though..

 

[patsu]

This is news to me. I always assumed Ninja Theory use SPUs to do culling.

Now I understand this statement better:
http://forum.beyond3d.com/showpost.php?p=948227&postcount=42

EDIT:

Apparently someone asked how many SPUs that was using, and it turns out it was all of them. Although, of course, there's the issue of optimisation/utilisation to balance that out if the rest of the game were to need more than the PPU.

If you look at the dynamic workload chart in the video, there are 2 long bars (PPU threads ?) and 6 shorter bars (6 SPUs ?). One of the short ones does not have any activities. So MrWibble could be correct... might be 5 SPUs.

 

[nAo]

This is news to me. I always assumed Ninja Theory use SPUs to do culling.
http://forum.beyond3d.com/showpost.php?p=948227&postcount=42

We use SPUs to cull a coarse representation of objects in the scene, while EDGE culls at triangle level. Yep..they check every single triangle

 

[AbbA]

Why "only" 5 SPU, 2 SPU are reserved?

 

[patsu]

No... only 1 is reserved permanently by the OS. That leaves 1 SPU for other activities.

 

[joker454]

It makes me wonder if ports will fare a little better given these tools. The 2D crowds in Fight Night 3 were really hurting! Well if Spurs turns out to be as good as it sounds, perhaps Joker454 can stop bitching about PS3 being vertex limited... I wonder what happened to that guy, it seemed as if he had just begun working on a PS3 title. Wonder what his opinion is now given these new tools?

Au contraire! Remember, my original beef was not that the PS3 was vertex limited, it was that the RSX was vertex limited, which I still stand by 100% today. In fact, I think the article quoted in this thread basically proves it. If you want a real world example, one of my code systems here at work can render ~1.2 million verticies on the 360's gpu faster than the rsx can render the same system with reduced lod models at ~700k verticies. This is all gpu, no cpu help. Yes, the difference is that lopsided in our game, even after I managed to bring that particular system down to two vertex inputs and optimize the heck out of the vertex shader and geometry layout ;( I don't wanna go through that whole thread again like last time, but needless to say to me, rsx still sucks. Your mileage may vary.

I never had a beef with cell though. It has a learning curve, and you need to pretty much rewrite all that code you've been saving over the years, but I think its worth it long term. The new tools and libraries that Sony has been putting out are awesome! I'll wait and see what real world performance we can get before I jump up and down for joy though. We only get 5 spu's full time, and the 6th is "most of the time" since Sony can use it if they need to (the 7th is theirs full time), and I assume with Home and all this other cool stuff they have coming out that they will indeed make use of the 6th.

We're currently fully using 2 spu's, and will most likely use all six by the time we are done since our goal is to hit 60fps on both platforms no matter what. As for how many we'll have free to handle geometry culling, I don't know just yet. I've been having fun with them so far though. I've moved all skinning off the gpu side to cpu side, since we get our biggest gains here on the PS3 by babying the RSX, ie, giving it as little to do as possible to get around its inadequacies. I actually tried it on the 360 as well and it gave back some fps, so the vmx units on the 360 will be fully worked also.

Doing geometry culling on the spu's is something I will do for sure, but I will also try it to a certain extent on the 360 as well, just to see what happens.

 

[inefficient]

If you want a real world example, one of my code systems here at work can render ~1.2 million verticies on the 360's gpu faster than the rsx can render the same system with reduced lod models at ~700k verticies. This is all gpu, no cpu help.

Without knowing what kind of fragment shaders you are running at the same time isn't that comparison useless?

If you created more work for the ALUs that was non vertex related, the amount of resources available for your VS code per time frame would drop on the 360's GPU - while on RSX it is a constant that you can always depend on to be there.

 

[patsu]

joker454, are you guys using SPURS ? How do you partition your work amongst the SPUs ? (assuming you can talk about it).

 

[joker454]

Without knowing what kind of fragment shaders you are running at the same time isn't that comparison useless?

If you created more work for the ALUs that was non vertex related, the amount of resources available for your VS code per time frame would drop on the 360's GPU - while on RSX it is a constant that you can always depend on to be there.

It's a real world example from an actual game running shipping shaders. I could have changed the pixel shader to just return the color red in theory to make it just be a vertex test. But then it would no longer be a valid test since, as you say, the 360's gpu can can then allocate more of itself to vertex processing, when in reality in the actual game it would not be able to do so. That's why I stress that the only valid tests are real world examples, assuming you've made optimizations that each platform likes.

 

[joker454]

joker454, are you guys using SPURS ? How do you partition your work amongst the SPUs ? (assuming you can talk about it).

Yup we're using spurs which is great. I can't get into the details of how divide work though, partly because of nda's, and partly because it's still a work in progress

 

[rounin]

Without knowing what kind of fragment shaders you are running at the same time isn't that comparison useless?

If you created more work for the ALUs that was non vertex related, the amount of resources available for your VS code per time frame would drop on the 360's GPU - while on RSX it is a constant that you can always depend on to be there.

As a further question, in the case where you create more work for ALUs that was vertex related, wouldn't performance suffer on the pixel-shading end in comparison to the PS3's RSX (I am under the impression that RSX has more power in the PS shading department)? In the end, wouldn't it come down a question of which platform you designed your game around to in the beginning and whether you've designed the PS/VS load for your game depending on the primary platform or not?

 

[patsu]

Yup we're using spurs which is great. I can't get into the details of how divide work though, partly because of nda's, and partly because it's still a work in progress

Ok... ok... [grumbles about NDA]

What are your goals though. Using up 6 SPUs eventually is fine and all, but are users going to notice it ? Are you guys going to try something mind-blowing ?

 

[joker454]

What are your goals though. Using up 6 SPUs eventually is fine and all, but are users going to notice it ? Are you guys going to try something mind-blowing ?

At this point, they will be used purely for performance reasons. Our title is a sports title, so 60fps will definitely help with playability compared to 30fps.