Support NeoGAF

[BlueTsunami]

~Not the Traditional benchmark Some of you may be thinking~

Its more or less benchmarking how Cell handles certain code and other intensive processing tasks.

For all who care. In all honesty, alot of it went over my head but its good to Wiki alot of the terms and see if you can fathom just a fraction of some of the more "Intense" wordings in the writeup.

http://www-128.ibm.com/developerworks/power/library/pa-cellperf/

Props to macabre at B3D to bringing it to the communities attention

EDIT: DAMN wrong forum :I . Meant to post this in the Gaming Forum.....this is what I get for creating a thread at 6:30am Eastern...

 

[Mupepe]

~Not the Traditional benchmark Some of you may be thinking~

Its more or less benchmarking how Cell handles certain code and other intensive processing tasks.

For all who care. In all honesty, alot of it went over my head but its good to Wiki alot of the terms and see if you can fathom just a fraction of some of the more "Intense" wordings in the writeup.

http://www-128.ibm.com/developerworks/power/library/pa-cellperf/

Props to macabre at B3D to bringing it to the communities attention

in short, gives us the good and bad please.

 

[BlueTsunami]

in short, gives us the good and bad please.

Well, there are many that are more suited than me to comment on this but it looks to be extremley well at decoding mpeg-2 files (SDTV and HDTV) and is also good at Video imaging. All these things where pretty much known already but its cool to see physical numbers like that.

Also, the things Cell specializes at, it seems that its "order of magnitude" higher at those tasks when compared to general CPUs.

 

[Kroole]

Well yeah. Those things (encoding, compression etc) really benefits more from a multicore design.

 

[gofreak]

Nice. Good to have some absolute frames per second figures for that Terrain Rendering demo too - 30fps on Cell, 0.85fps on 2.7 Ghz G5. If you scaled the G5 up to 3.2Ghz, it'd get 1fps - so a 30x speed up, clock for clock, over the G5 is pretty damn good!

Also interesting to note the "test" vertex TnL rates - with 7 SPUs, it'd be ~1.5bn verts/sec.

The EIB looks really good. The trend of much larger gains with optimised code is predictable, but interesting to note.

 

[Myoclonic Jerk]

Nice. Good to have some absolute frames per second figures for that Terrain Rendering demo too - 30fps on Cell, 0.85fps on 2.7 Ghz G5. If you scaled the G5 up to 3.2Ghz, it'd get 1fps - so a 30x speed up, clock for clock, over the G5 is pretty damn good!

Also interesting to note the "test" vertex TnL rates - with 7 SPUs, it'd be ~1.5bn verts/sec.

The EIB looks really good. The trend of much larger gains with optimised code is predictable, but interesting to note.

indeed

 

[gofreak]

Since I know GAF likes summaries:

 

[Wollan]

Spoiler

OMG

 

[seanoff]

2 SPEs = 400M verts at full capacity.

Even if they hit 25% efficiency it can still hit 100M verts. oooohh. and 5 spes + PPE to do other stuff with.

Any educated guesses on the RSX T&L and can it shade an extra 100m verts anyway??

Holy cow these modern consoles pack some power. YIKES

 

[TTP]

So Killzone was real time after all.

 

[Wax Free Vanilla]

So Killzone was real time after all.

yes. the final game will look even better.

 

[Razoric]

Can anyone compare this to the 360s tri-core?

 

[TTP]

Can anyone compare this to the 360s tri-core?

Yes please. And can someone cook some pop-corn?

 

[TheJesusFactor]

I agree that the Cell will be powerful, but I rather wait till the games come out to make my decision. With all the rumors that Games will be difficult and expensive to produce on PS3. It is very possible that we may never see its full potential until the end of its life cycle, which is sad given its horse power. Hopefully, the system will be cheap enough for me to grab one.

 

[DarienA]

I agree that the PS2 will be powerful, but I rather wait till the games come out to make my decision. With all the rumors that Games will be difficult and expensive to produce on PS2. It is very possible that we may never see its full potential until the end of its life cycle, which is sad given its horse power. Hopefully, the system will be cheap enough for me to grab one.

I knew this sounded familiar....

 

[gofreak]

2 SPEs = 400M verts at full capacity.

Even if they hit 25% efficiency it can still hit 100M verts. oooohh. and 5 spes + PPE to do other stuff with.

Any educated guesses on the RSX T&L and can it shade an extra 100m verts anyway??

Holy cow these modern consoles pack some power. YIKES

The vertex rate is a realworld benchmark, not a paper figure - though of course, as a benchmark it's only doing the minimum amount of processing possible. Of course, with heavier processing, such as that in a game and so forth, the number would go down (if that's what you mean by "efficiency").

On a related note, there's an interview with Alex Chow of IBM up also, in which he seems to suggest there's a Cg compiler for Cell - as in, NVidia Cg:

http://www-128.ibm.com/developerworks/library/pa-expert8/

dW: Do the programming models require some pre-knowledge by the compiler?

Chow: The currently released compiler and toolchain do not have pre-knowledge of any programming model. Some other compilers, for example, openMP and Cg, define their own programming model through their languages.

Interesting, no?

 

[TTP]

I knew this sounded familiar....

Yeah he must be posting through www.randomobviousthoughtsgenerator.com

 

[quetz67]

I agree that the Cell will be powerful, but I rather wait till the games come out to make my decision. With all the rumors that Games will be difficult and expensive to produce on PS3. It is very possible that we may never see its full potential until the end of its life cycle, which is sad given its horse power. Hopefully, the system will be cheap enough for me to grab one.

I think we will see less games, but the games from companies able to afford the production value needed will be breathtaking.

 

[TheJesusFactor]

Yeah he must be posting through www.randomobviousthoughtsgenerator.com

1st: It is not random, it is related to the topic.

2nd: You'll be surprise how this is not obvious for some people.

 

[Naked Snake]

In before panajev, Marconelly, and dark10x

 

[Deg]

Can it do at least 2xAA?

 

[teiresias]

Can it do at least 2xAA?

Not as well as the Dreamcast.

 

[xexex]

not surprising that Cell can massively outperform a conventional CPU in certain graphics-related areas.

Killzone in realtime depends largely on the GPU rendering it - I don't think even Nvidia's next-gen DX10 NV50 GPU will be able to manage Killzone CG in realtime, much less RSX which is basicly NV47+

 

[Deg]

not surprising that Cell can massively outperform a conventional CPU in certain graphics-related areas.

Killzone in realtime depends largely on the GPU rendering it - I don't think even Nvidia's next-gen DX10 NV50 GPU will be able to manage Killzone CG in realtime, much less RSX which is basicly NV47+

But ATI will AM I RITE?

 

[gofreak]

not surprising that Cell can massively outperform a conventional CPU in certain graphics-related areas.

Killzone in realtime depends largely on the GPU rendering it - I don't think even Nvidia's next-gen DX10 NV50 GPU will be able to manage Killzone CG in realtime, much less RSX which is basicly NV47+

What about Cell+RSX

Sorry, I'm still just contemplating Cg shader compilation across both chips

Without comment on KZ, to be honest I think we'd be surprised by what you'll get out of SM3.0+ chips. PC hardware moves so fast that we never get to see a part used fully (or arguably, properly). SM3.0 is a relatively complete model too - performance is the main issue.

 

[Arsynic]

Spoiler

OMG

I thought the Cell in PS3 was using 7 SPE's. If so, those numbers need to be modified just a bit.

 

[gofreak]

I thought the Cell in PS3 was using 7 SPE's. If so, those numbers need to be modified just a bit.

The only ones that would really require modification would be the first 3. The Terrain Renderer, in the demo IBM showed, uses a SPU to compress the framebuffer to send over the network to a client, so it's only using 7 SPUs for the raycasting/rendering anyway. The others show performance for 1 SPU only.

 

[Arsynic]

Yeah, like the T&L is fine cuz it only uses 2 SPE's, but the rest need to be changed unless...unless...unless...

YIELDS ARE SO GOOD THAT SONY IS ENABLING THE 8TH SPE...THE SLEEPING GIANT! CELL FTW!

 

[gofreak]

Yeah, like the T&L is fine cuz it only uses 2 SPE's, but the rest need to be changed unless...unless...unless...

YIELDS ARE SO GOOD THAT SONY IS ENABLING THE 8TH SPE

Hehe, no. But I'm not sure I follow. The T&L example, as with all except the first 4 in the chart, give performance for 1 SPU (217 million vertices per sec).

The first 3, adjusted linearly for 7 SPUs as opposed to 8, would look like this:

Matrix Multiplication (S.P.) 25.6 Gflops* (w/SIMD) 175 GFlops (7SPEs) 7x (7SPEs)
Linpack (S.P.) 4kx4k 25.6 GFlops* (w/SIMD) 136.5 GFlops (7SPEs) 5.25x (7SPEs)
Linpack (D.P.) 1kx1k 7.2 GFlops (3.6GHz IA32/SSE3) 8.46 GFLops (8SPEs) 1.14x (8SPEs)

 

[snoopers]

Nobody to notice that this article has been written by IBM, also creator (with Toshiba and Sony) of the Cell ?

 

[BlueTsunami]

Nobody to notice that this article has been written by IBM, also creator (with Toshiba and Sony) of the Cell ?

The link has IBM in it and the font used in the article has an IBM'esque look to it. I believe that everyone that has clicked on the link knows that its an article written up by IBM.

 

[Lord Error]

On paper, Cell definitely impresses, but I really want to see how it performs in real life apps first.

Nobody to notice that this article has been written by IBM, also creator (with Toshiba and Sony) of the Cell ?

Yeah, I think that was difficult to notice, the logo on the page kinda gives it away though

They compared Cell with their own other CPU though.

 

[Wax Free Vanilla]

Nobody to notice that this article has been written by IBM, also creator (with Toshiba and Sony) of the Cell ?

Yes, the same company that makes the Xbox 360 CPU, whats your point?

In Cell We Trust

 

[iapetus]

The first 3, adjusted linearly for 7 SPUs as opposed to 8

Which, of course, you can't do because performance doesn't scale linearly as you add processors...

 

[mrklaw]

Which, of course, you can't do because performance doesn't scale linearly as you add processors...

but most of the benchmarks are listed based on 1 SPE. And some of the code - eg matrix transforms - would (roughly) scale linearly if the SPEs can be left to handle discreet, non-overlapping models.

 

[ToxicAdam]

Where's my 120 FPS?!

 

[gofreak]

Which, of course, you can't do because performance doesn't scale linearly as you add processors...

In some of the instances shown, it does scale pretty linearly. They highlight the matrix multiplication explicitly as doing so, for example. The Linpack stuff is probably very similar (and the TRE seems to be too, comparing their figures for a Cell vs a 2Ghz G5 with Cell vs a 2.7Ghz Cell).

And anyway, unless you think performance will scale better than linearly with more SPUs, my downscaling of the figures linearly from 8 to 7 SPUs is only going to understate the performance with the latter

 

[HomerSimpson-Man]

Wow, those are impressive numbers, the verts rate on one SPU is insane though! With 7 SPU and RSX capable of hundreds of millions itself...crazy.

 

[DJ Sl4m]

Another words, Sony and 3rd party devs for the PS3 dont have any new in game movies to showoff to counter some of the 360 buzz at the moment so we get this untill then ?

Impressive numbers or not, I can't wait to see some real in game footage of PS3 games, surely they have something available at this point.

 

[teiresias]

Another words, Sony and 3rd party devs for the PS3 dont have any new in game movies to showoff to counter some of the 360 buzz at the moment so we get this untill then ?

Impressive numbers or not, I can't wait to see some real in game footage of PS3 games, surely they have something available at this point.

The release of this has less to do with PS3 and Sony and more to do with the promotion of the CBE architecture in general by the others involved in its development - there is a difference.

 

[gofreak]

Another words, Sony and 3rd party devs for the PS3 dont have any new in game movies to showoff to counter some of the 360 buzz at the moment so we get this untill then ?

This has nothing to do with PS3 PR countering X360 PR. This is a IBM doc.

Impressive numbers or not, I can't wait to see some real in game footage of PS3 games, surely they have something available at this point.

Everyone's definition is different (what do you mean by "in game" footage? We have seen someone play a game, if that's what you mean). You have certainly seen what PS3 games will look like however, from a graphical standpoint at least.

 

[dorio]

So can cell and RSX be used synergistically to create scenes of billions and billions of highly shaded polys? How additive is it etc?

 

[dorio]

The release of this has less to do with PS3 and Sony and more to do with the promotion of the CBE architecture in general by the others involved in its development - there is a difference.

Ibm makes the cpu inside the 360 also so that reasoning doesn't make sense.

 

[DJ Sl4m]

Ok, my bad guys, I see your point and didn't realise it was a promo from the IBM perspective for other ventures other than the PS3 at this point.
I'm a bit behind in all online news (esp gaming, since I've been in Gulfport miss the last few weeks)

I suppose what I mean by in game movies is, someone playing a game (controlling the charachter) from a PS3 dev kit or hardware not running off a Mac or PC at this point.

Is there a particular site ( similar to Blim's) which vids can be watched without having to register or get a membership ?

 

[gofreak]

Ibm makes the cpu inside the 360 also so that reasoning doesn't make sense.

Xenon isn't being pushed outside of X360, though.

So can cell and RSX be used synergistically to create scenes of billions and billions of highly shaded polys? How additive is it etc?

You can use Cell to do vertex processing, as you can with any CPU. However traditionally in PCs, there isn't a lot of power spare to work with (particularly when you have to scale aswell across many different CPUs of varying power), and there isn't a lot of bandwidth to the GPU. Assuming those things are covered, as I think they are in Cell, the question is then how it'd be facilitated. Cg on Cell would be a good start here, as it would allow you to compile vertex shaders to run on both the GPU and CPU (versus writing seperate vertex processing code for the GPU and CPU). Beyond that, then there'd be the question of how work would be split between the two - dynamically at run-time by some provided library, or at compile time by the dev?

Also, you'll ultimately be limited by the triangle setup rate too, of course.

But yes, vertex shading across both can be done. It's more a question of how, exactly.

 

[DCharlie]

Interesting stuff - although i thought it was 1 redundant SPU and 1 OS locked SPU?
Not that that makes a massive difference.

 

[gofreak]

Interesting stuff - although i thought it was 1 redundant SPU and 1 OS locked SPU?
Not that that makes a massive difference.

It doesn't when most of the benches are on 1 SPU

AFAIK, it's unconfirmed if a SPU will be reserved for the OS. I'd expect it though (Sony seemed to hint at this themselves, as two of the SPUs were faded out on their Cell diagrams at E3, for no apparent reason).

 

[mrklaw]

Also, you'll ultimately be limited by the triangle setup rate too, of course.
.

But then you can always fall back to software rendering. If you really need more triangles than RSX can deliver.


1SPU reserved for OS seems overkill if it can handle 25GLOPS. Maybe you can't rely on 7 full ones (i.e the last one might use some for OS stuff). And what OS stuff will be happening in the background anyway? Video decompression/save games/guide stuff would be foreground tasks so thats more a memory reserved thing than FLOPS thing.

 

[gofreak]

But then you can always fall back to software rendering. If you really need more triangles than RSX can deliver.

Synching that with what the GPU would be rendering could be tricky. And you'd be adding pixel shading and rasterisation into Cell's bucket then too

If you wanted to render all transparencies on Cell, though, that could be done independently and blended with the GPU's buffer when finished.

I think an obvious use would be if you had the power spare on the CPU side, to simply split your geometry between the vertex shaders and Cell, and have them both apply the same shaders to their respective allocations, and basically just wolf through the vertex work. It'd be a quick and presumably easy way to soak up any excess power on Cell too (assuming a Cell Cg compiler). You could have more sophisticated setups too, maybe applying a "pre-processing" shader on Cell to all or some pf the geometry before passing it to the VS - the "pre-processing" shader basically reducing the workload of the vertex shader (i.e. taking a shader and splitting it into two independent parts, the first "half" done on Cell, the second on the GPU).

1SPU reserved for OS seems overkill if it can handle 25GLOPS. Maybe you can't rely on 7 full ones (i.e the last one might use some for OS stuff). And what OS stuff will be happening in the background anyway? Video decompression/save games/guide stuff would be foreground tasks so thats more a memory reserved thing than FLOPS thing.

What if someone wants to watch a HD movie on a second screen while you're playing a game? Or your PSP-owning brother or friend comes a long and wants to stream a video to his PSP from the PS3 HD (which your SPU would format appropriately for the PSP etc.). It could provide other services for games too, like maybe eyetoy processing etc.

 

[Diablos]

Wow, and I once thought EE, Vector Units and GS discussion was complicated.
I don't even know what the fuck I just read. :lol

Anyone here able to break it down for some of us?