[Techno+]

hello guys

I have two questions about the cell processor, which i hope you would answer.

1) How many GFLOPS is CEll capable of?
2)How is gfx processing dine on Cell?

[Razor1]

Depends on the cell processor, the ones used in the PS3 supossedly can do up to 100 gigflops

the cell processor isn't that good for graphics processing, its good for vertex processing but thats about it

[Arun]

The vast majority of the PS3 Graphics Processing is done on RSX, not CELL. CELL is not good enough at hiding latency, and doesn't have all the fixed-function hardware that is traditionally required to make graphics processing really fast. I'm just simplifying here, but that's the very basic gist of it.

Uttar

[nonamer]

Originally Posted by Razor1

Depends on the cell processor, the ones used in the PS3 supossedly can do up to 100 gigflops

the cell processor isn't that good for graphics processing, its good for vertex processing but thats about it

200 GFLOPS actually. But you're right, it's not something primarily for graphics.

[Bobbler]

Originally Posted by Razor1

Depends on the cell processor, the ones used in the PS3 supossedly can do up to 100 gigflops

the cell processor isn't that good for graphics processing, its good for vertex processing but thats about it

The one in PS3 can, in theory, do a bit over 200. Gflop ratings are almost useless though.

Each SPE at 3.2ghz is around 25gflops, PPE is supposedly ~38 (IIRC that number isn't correct in reality -- it's like 25 or so as well). Someone with a better memory can probably tell you the exact numbers.

[Razor1]

opps sorry should have been a 2 typo

[Techno+]

so is there a way cells gfx processing capabilities, and is there a way to utilize all that vertex power?

[Techno+]

is it good for texture mapping?

[Razor1]

It can be used to some degree, simliar to what CPU's are doing now in PC's, but the cell is more proficient at vertex calculations then x86 cpus.

[Arun]

Originally Posted by Techno+

so is there a way cells gfx processing capabilities, and is there a way to utilize all that vertex power?

Generally, and for the vast majority of cases (99.9%+), no, you'll be triangle setup limited way before that.

is it good for texture mapping?

Nope, no dedicated bilinear interpolation hardware and insufficient latency hiding, among other things. Sure, it's better than a Pentium, but it's not on the same order of magnitude as a dedicated GPU, in terms of speed. So, no.


Uttar

[Arwin]

Originally Posted by Techno+

hello guys

I have two questions about the cell processor, which i hope you would answer.

1) How many GFLOPS is CEll capable of?

As mentioned, a little over 200 Gflops, with each of the 7 SPEs rated at about 26.

2)How is gfx processing dine on Cell?

That really depends on the game developer. In theory, a developer can create a game using only the RSX for graphics, but conversely a developer can also use the Cell exclusively.

The real power is that the load can be distributed over both systems, each using the other system to their own strengths. I think probably one of the most common setups will be this:

the Cell do a lot of physics and animation calculations, as well as maybe some calculations on what does and does not need to be drawn. Here and there, it may also pre-process textures, or even generate them.

The resulting vertex and texture data will be read from XDR memory by the RSX chip. The RSX will then render the scene, with maybe some textures and other data also cached in GDDR3 memory, and output the rendered scene to XDR memory, or first to GDDR3 and then copied to XDR, depending on where the bandwidth is needed most. Then finally the Cell can do some post-processing, which is then read again by the RSX and displayed.

But like I said, some of these steps are going to be optional for some games, and implemented differently in others. Just giving you an idea of how they might work together.

Note that I've just based this on what I learnt on the internet, and I'm not a 3D expert.

[Arwin]

One last thing ... Purely looking at Gflops ratings of the GPU, SPE and PPE doesn't quite tell the whole story either. The SPEs are less efficient for certain routines and more efficient for other routines than the GPU, the GPU is less efficient for certain routines and more efficient for other routines than the SPE, the PPE is less efficient for certain routines and more efficient for other routines than the SPE, and the PPE is less efficient for certain routines and more efficient for other routines than the GPU.

And in some of these cases, the difference in efficiency can be really, really big too. To give an oft-quoted example, the SPE is extremely efficient at FFTs and is optimised for data streaming, is helped a lot by its local store for certain type of operations, and while not as good at branch prediction as the PPE, it still has more branch logic than the GPU ALUs (hopefully I got my acrynoms right).

And of course the overall bandwidth design through the flexio, individual memory management units on the different parts and the Rambus setup all contribute a lot to the overall performance of the system.

But anyway to summarise, the SPEs are capable and flexible streaming (co-)processors that can help out in various math and/or streaming heavy tasks, which include certain types of graphics (post-)processing, vertex data manipulation and animation, physics calculation, and so on. It can take many bits of hard work out of the hands of the GPU (and PPE) so that the RSX can focus more on the core rendering, and they seem to be setup to work together quite efficiently.

[Techno+]

so that's way the gfx on PS3 look so good, since it is actually a collaboraton between Cell and RSX, so Cell does help in generating vertices and textures and also helps in post-processing.

Thanx a lot

[nelg]

Originally Posted by Arwin

That really depends on the game developer. In theory, a developer can create a game using only the RSX for graphics, but conversely a developer can also use the Cell exclusively.

That is true for any CPU, not just Cell. I have a game in beta testing that uses any CPU that can run MS's ref-rast. We started testing a year ago and testers are reporting that they have recently got thorough the opening credits.

[Guden Oden]

Originally Posted by Razor1

the cell processor isn't that good for graphics processing, its good for vertex processing but thats about it

That depends on how you define "graphics processing", really.

Someone did a rather fast raytracer (ie: multiple times faster than on a regular CPU) that runs on a cell blade server a while back, there was a discussion thread in one of the console-related forums about it I believe. Also, the game warhawk generates volumetrically rendered cloud textures on the fly on one cell SPE, a task that you pretty much can't do on a GPU today, or at least not very efficiently (or someone would surely have done it already).

So it really depends. For certain specialized types of graphics processing it excels. For general polygon 3D rendering, it's fairly "meh" I think we can safely proclaim, but then again that's not a problem because it's not there to sit and draw polys anyway.

[[maven]]

Originally Posted by Guden Oden

Also, the game warhawk generates volumetrically rendered cloud textures on the fly on one cell SPE, a task that you pretty much can't do on a GPU today, or at least not very efficiently (or someone would surely have done it already).

Well, it has already been done, but it makes for much better Pro-Cell PR to ignore or forget that, see for example here.

[zed]

the problem with that is its a demo (which wouldnt run on my system )
ie the cpu is devoted to doing just that,
now in a typical game u will also want the cpu working on AI/sound processing etc
but if the cpu is busy doing the clouds then it mightnt have enuf cycles left over to perform all these other things

with cell u have the 6 cores u might as well make use of them

i believe a nice use of cell would be like the ultimate vertex processor like a more advanced version of the geometry shader in d3d10
or as a shading calculator, millions rays/sec

[Razor1]

Originally Posted by Guden Oden

That depends on how you define "graphics processing", really.

Someone did a rather fast raytracer (ie: multiple times faster than on a regular CPU) that runs on a cell blade server a while back, there was a discussion thread in one of the console-related forums about it I believe. Also, the game warhawk generates volumetrically rendered cloud textures on the fly on one cell SPE, a task that you pretty much can't do on a GPU today, or at least not very efficiently (or someone would surely have done it already).

So it really depends. For certain specialized types of graphics processing it excels. For general polygon 3D rendering, it's fairly "meh" I think we can safely proclaim, but then again that's not a problem because it's not there to sit and draw polys anyway.

Ah yes I agree it would be good at raytracing, was assuming more texture based approachs aren't that great on Cell

[Guden Oden]

Originally Posted by [maven]

Well, it has already been done, but it makes for much better Pro-Cell PR to ignore or forget that, see for example here.

Ah, yes. The great cell-uberness conspiracy! Yes, we cell fans always ignore whatever obscure graphics demo that may exist on some almost totally unknown website out there somewhere in order to preserve our views of cell's uniqueness... Sheesh! Yes, I admit I visit that site every day and just plain lied when I said it's never been done.

I thought it fully obvious that my comment referred to GAMES, but apparantly that was not the case.

Anyway, I can't check how accurately this thing may mimic warhawk's cloud rendering (which I've only seen in blurry streaming videos anyway) since the program crashes immediately when run. Yay, score one for the gpgpu guys vs cell, I guess. Or not.