RSX still in development, no silicon produced yet.

[mckmas8808]

Dr. Evil he said both about the explosion. He said in the beginning of the demo it was just the CELL, then at the end at night time he said CELL and RSX.
________
Hot Box Vaporizer

 

[London Geezer]

Dr. Evil he said both about the explosion. He said in the beginning of the demo it was just the CELL, then at the end at night time he said CELL and RSX.

Why would the daylight scene be Cell only and the night scene be Cell+RSX?

 

[eDoshin]

Dr. Evil he said both about the explosion. He said in the beginning of the demo it was just the CELL, then at the end at night time he said CELL and RSX.

Why would the daylight scene be Cell only and the night scene be Cell+RSX?

Well Duh!!! You need RSX to render all the black spaces.

 

[Laa-Yosh]

"The key to this demo was that the imagery was ray-traced in realtime solely via a _ pair of Cell processors_.

Correct me if I'm wrong but wasn't that demo using raycasting? Which means that one ray is cast for each pixel, but no bounces or anything else is involved. I'd like to remind you that Wolfenstein 3D was also using realtime raycasting, although Carmack has only cast one ray for each column and that's why there was only a single planar floor and a fixed height ceiling.
Nevertheless, it means that while Wolfenstein calculated a maximum of 320 rays per frame, and run really well on an average 486, the Cell processors would probably need at least a million rays per frame... which is still quite impressive. But if I recall it correctly, then websites describing it as raytracing are only building up source material for more flamewars...

Actual raytracing for effects like ambient occlusion, global illumination and subsurface scattering requires many casted rays for each pixel (or block of pixels when using undersampling), and even with 50-100 rays per sample, you'll probably end up with noise. SSS might also require bounces, which increase the number of rays even more.

Which brings up the Molina head demo. Can anyone confirm if the SSS solution was calculated in real time? I'd rather think that it was precalculated for the static mesh, especially because it was so amazingly good looking. Then they've just read the cached values for each vertex from a texture or per-vertex data...

 

[Titanio]

Which brings up the Molina head demo. Can anyone confirm if the SSS solution was calculated in real time? I'd rather think that it was precalculated for the static mesh, especially because it was so amazingly good looking. Then they've just read the cached values for each vertex from a texture or per-vertex data...

It's a good question if Cell calculated it once, or was updating it in "realtime" (or every few frames). I doubt there'd be much point in highlighting that cell was calculating that if it wasn't doing the latter - a precomputed static set of data could be generated on any computer really, given enough time, and just loaded and fed into the gpu when needed.

Phil Harrison's comment on it suggests they were doing it in realtime:

Eurogamer: Was most of what we saw really just showing off the graphics capabilities - stretching the RSX graphics part rather than the Cell chip? The assumption is that Cell is there for complex physics and AI...

Phil Harrison: You're right; obviously Cell allows you to do complex collisions, physics, dynamics, simulations, all of those things. Though, the Getaway demo was a good example of how you can have a living city brought to life as a result. Although it was pretty graphics, most of that power was actually Cell-based.

The Doc Ock head - the Alfred Molina head - is actually more of a Cell demo than it is a graphics demo, because we're calculating hugely complicated light sources in real-time on the Cell, even to the point where we calculate the angle at which light enters the skin, the way that the light is then coloured by your blood, and the way that it is then reflected back out. It's something called transmission. Skin is hugely complicated - if I put my finger over a light, for example, you can see that the light is coming through my skin. We were simulating that - emulating, simulating, kind of a fine line - we were simulating that on the Doc Ock head demo.

If it was just a static data set, there'd be no point in cell doing any calculations - just do them offline and store it for use when needs be. What do you think?

 

[mckmas8808]

Why would the daylight scene be Cell only and the night scene be Cell+RSX?

No I was trying to give you the time frame that he said it in. Better yet here is the video at 16 seconds he explains what the CELL is exactly doing. Then in the end he says something about the RSX and CELL. Someone more expirenced can break it down I just know what I heard.

http://media.ps3.ign.com/media/748/...14/614835/sonycon_demos_gasstation_wmvlow.wmv
________
DRUG TESTING

 

[Shifty Geezer]

He talks about Cell rendering volumetrically. I'm pretty sure what he's getting at is Cell mathematically modelling an explosion in terms of air temperature, density, and so forth. The heat haze and light are produced as a result of the explosion, and this information is rendered through RSX.

Cell is not producing visuals, but the simulation needed to create geometry and visual effects to be rendered on RSX (or rather, whatever GPU they are using )

 

[mckmas8808]

I understand now. Hey Shifty is there any games that allows GPUs and CPUs to work together like this in the PC gaming market now?
________
BUY GRINDERS

 

[Laa-Yosh]

a precomputed static set of data could be generated on any computer really, given enough time, and just loaded and fed into the gpu when needed.

That probably wouldn't disturb a company that can live with a CG animation assumed to be realtime by the press and the fans.

Phil Harrison's comment on it suggests they were doing it in realtime:
The Doc Ock head - the Alfred Molina head - is actually more of a Cell demo than it is a graphics demo, because we're calculating hugely complicated light sources in real-time on the Cell, even to the point where we calculate the angle at which light enters the skin, the way that the light is then coloured by your blood, and the way that it is then reflected back out.
We were simulating that - emulating, simulating, kind of a fine line - we were simulating that on the Doc Ock head demo.

Notice the part in bold.
If you calculate a static SSS solution, it will hold information about how much indirect light a given point on the surface will get if it's surroundings are lit. This is only a part of a larger calculation, so he can say that the demonstration and the lighting is in real time.
However, this part is the single most expensive part of the calculation, taking several minutes for a 640*480 render on a PC-based workstation, and I doubt that Cell would be fast enough to calculate it in realtime.
It is too hard to see if the SSS updates properly with the deformations of the head during the lipsync animation in the demo. Nevertheless it's not deforming that much so they might as well get away with it.

Also note that the head is a very high quality model created by Sony Pictures Imageworks (a VFX studio) to be used as Molina's CG double in the Spiderman movie. Is it that much to assume that Imageworks provided not only the - 3D scanned - model and its textures, but also their - very very good - SSS solution?

Keep in mind that this does not invalidate neither the demo, nor the technique. It might as well get used in a game if it works well enough.

 

[jvd]

When you think of it, why wouldn't it be more than twice as fast, the bandwidth (35GB/sec) between Cell and RSX puts even PCI-Express x32 to shame. Same with X360's 22 GB/sec bandwidth between XCPU and Xenos. They're both faster than a single card implementation on a PC which is 16 GB/sec per card.......is that about right? SLI bandwidth is more than the dedicated bus on 360 but still fractionally slower than PS3's. And its clocked higher than G70.

the geforce 6800 ultra has 550mhz ddr ram with a 256bit bus giving it 35.2 gb/s bandwidth to its pool of 256 megs of ram .

The cell to rsx bus really means nothing , I would think even a standard pci bus to the cpu - gpu in a pc would have enough to carry the needed data and surely a agp interface would be mroe than enough . What really matters is buffer bandwidth and texture reads and writes bandwidth

 

[Berserk]

In the press conference Harrison clearly states that the Explosion demo was a demonstration of the "power of Cell and RSX", and mentions the RSX specifically in reference to the heat haze.

I don't really see where people are getting the idea that the Getaway demo was supposed to be only Cell either, there is nothing to indicate that, and before you pull out the interview where Harrison says something like, "That's mainly a Cell demo" he's talking in relation to the fact that the impressive thing about the demo isn't the graphics but the simulation of the city with the traffic and the pedestrians and how the Cell can handle all that "AI" kind of stuff, he never says the demo only uses Cell.

Hmm NOPE! He CLEARLY says this about the explosion demo:

"Now what this demonstration actually uses, is ONLY the cell processor. to render volumetrically all the performance required to simulate an explosion, thermal dynamics, heat, gas, smoke, fire."


The 1st Landscape demo was also done in realtime by Cell only. (second landscape was cell + RSX)


Well see for yourself. This is the complete Sony conference (1h30min)

http://www.ryoni.com/news/134/ARTICLE/1426/2005-05-17.html (streaming but very fast!)

 

[jvd]

"Now what this demonstration actually uses, is ONLY the cell processor. to render volumetrically all the performance required to simulate an explosion, thermal dynamics, heat, gas, smoke, fire."

well whats rendering the rest of the scene ? Acording to this quote the cell chip is only simulating the explosion , thermal dynamics , heat , gas , smoke and fire . So there is something rendering the rest of the scene

 

[marconelly!]

The cell to rsx bus really means nothing , I would think even a standard pci bus to the cpu - gpu in a pc would have enough to carry the needed data and surely a agp interface would be mroe than enough.

So you're suggesting that the large bandwidth between Cell and RSX is a complete waste and is added only to increase the cost?
Somehow I doubt it.

 

[Tacitblue]

When you think of it, why wouldn't it be more than twice as fast, the bandwidth (35GB/sec) between Cell and RSX puts even PCI-Express x32 to shame. Same with X360's 22 GB/sec bandwidth between XCPU and Xenos. They're both faster than a single card implementation on a PC which is 16 GB/sec per card.......is that about right? SLI bandwidth is more than the dedicated bus on 360 but still fractionally slower than PS3's. And its clocked higher than G70.

the geforce 6800 ultra has 550mhz ddr ram with a 256bit bus giving it 35.2 gb/s bandwidth to its pool of 256 megs of ram .

The cell to rsx bus really means nothing , I would think even a standard pci bus to the cpu - gpu in a pc would have enough to carry the needed data and surely a agp interface would be mroe than enough . What really matters is buffer bandwidth and texture reads and writes bandwidth

Gotcha, however how does that explain the desire to have such a bandwidth increase between the CPU and GPU in this case? If there's no real benefit why use it? Throwing money at a level of integration between the 2 units which is uneccesary?

 

[jvd]

The cell to rsx bus really means nothing , I would think even a standard pci bus to the cpu - gpu in a pc would have enough to carry the needed data and surely a agp interface would be mroe than enough.

So you're suggesting that the large bandwidth between Cell and RSX is a complete waste and is added only to increase the cost?
Somehow I doubt it.

no but comparing it to the pc is a diffrent thing all together .

THe bandwidth is there so the rsx can acess the xdr ram. But the xdr ram is limited to another speed in comunications and has to travel through the cell chip and then to the rsx core . So comparing the numbes is meaning less esp when we don't know how much of the bandwidth to the ram the cell chip will require .

To many unknown factors to make a sugestion like that .

 

[Tacitblue]

AFAIK, the interface speed has as much to do with with using SPE's for post processing, dynamic blur for one I believe was mentioned. They said Cell can directly access objects on RSX for various functions. But yes I'd agree that part of the bandwidth is for accessing the XDR pool.

 

[jvd]

AFAIK, the interface speed has as much to do with with using SPE's for post processing, dynamic blur for one I believe was mentioned. They said Cell can directly access objects on RSX for various functions. But yes I'd agree that part of the bandwidth is for accessing the XDR pool.

right your going to need more bandwidth for all those things .


on a pc the cpu has acess to large pools of memory (normal gaming machine has 512 + of ram and now most are at 1gig ) and then the gpu has its own dedicated ram pool. The cpu also isn't doing post processing affects its mostly used to run the game code , the api and do geometery and what not . The needs for the two are highly diffrent .


Also don't forget that in june of this year the nv40 (6800ultra and the like ) will have been on store shelves for a year in large numbers .

So in another year when the ps3 comes out you will be comparing the ps3 to a 2 year old card . Yes it helps that its two 6800ultras in sli but that isn't very efficent and its not anywhere near close to a 2x jump in speed .

 

[Tacitblue]

As always though, a console irregardless of being frogleaped in performance by a PC, has a price/performance ratio that can't be ignored. Especially in my PC gaming experience where i've had to play the patch game ad nauseum. But i get where you're going 8)

 

[jvd]

As always though, a console irregardless of being frogleaped in performance by a PC, has a price/performance ratio that can't be ignored. Especially in my PC gaming experience where i've had to play the patch game ad nauseum. But i get where you're going 8)

That isn't true in all ways .


The gpu in the ps3 is most likely (By what they are saying ) lightly modified g70 which will be about 2x a 6800ultra and have new features over the 6800ultra like a 32 bit hdr . However the g70 will be out in a few months along with the r520 and by the time the ps3 comes out either refreshes of it will be hitting the shelves or the new gpus .

So yea u get alot of power for cheap but the real advantage is that from day one devs are coding for that hardware . When the ps3 comes out i will say devs will finaly start using the x800 and 6800 as the base for the games . That is where the real strength comes from .

 

[mckmas8808]

Do you guys think the way CELL is made that its going to take a little bit longer for PCs to past consoles than in the past?
________
Rudi Fischer