Masayuki Chatani (SCEI CTO) interview

[Shifty Geezer]

Is there no such algorithm as splitting scenes in parts to fit in 512MB of RAM in PS3 for good parallelism?

I use a raytracer called RealSoft. It worked on a PIII 800 with 512 Mbs RAM. This particular renderer doesn't tessellate NURBS so gets perfect results while also being memory efficient for models. Dividing a frame up into smal sections (tile rendering ) is possible across a network.

Not an ideal solution for a professional graphics outfit, but PS3 could be a cheap and capable way for entry-level CG creation if the software was there.

 

[Laa-Yosh]

Is there no such algorithm as splitting scenes in parts to fit in 512MB of RAM in PS3 for good parallelism?

Rendering times would start to increase. If you don't have the whole 2 gigs, you'll have to swap out to a harddrive. It KILLS speed.
Besides, I thought PS3 has 256MB of RAM, the rest is VRAM...

Is it more expensive than 4 Opterons + 2GB RAM machine for a typical render farm?

Quad processor machines aren't good, because you start to loose system bandwith speed and such. 2CPUs, 2 gigs per CPU for highend work. 1 GB for everything else.


Guys, PS3's and most likely Cell's memory system was designed for processing streamed data. Most commercional renderers available on the market are not working this way, they need memory to load the scene file and textures and other data. PRMan can be set up so that it can work with less memory, but it's always a tradeoff for speed.

We'll see how a Cell-based server/workstation works out in practice, but I have my doubts ever since I've learned the 90% performance drop for DP. It really isn't worth pushing, by the time they can introduce it on the market, regular PCs will be a better buy.

 

[Shifty Geezer]

Besides, I thought PS3 has 256MB of RAM, the rest is VRAM...

256 MB XDR, 256 MB DDR, all accessible by Cell directly (and RSX too). The XDR is lower latency to keep feed the streaming SPUs, but they can still work on DDR such as outputting data there. It's really just a division between 256 MBs fast RAM and 256 MBs faster RAM

 

[Wicked_Vengence]

Two things while reading through...

First, isn't Procedural Synthesis patented by MS, so what does it matter if the Cell would be better?

Second, rendering. You do realize that current Cell technology limits it to 256MB of XDR memory don't you? That makes it pretty useless to replace typical rendering setups...

 

[Tacitblue]

I think the dual Cell server IBM had at E3 had 512MB per Cell.

 

[Titanio]

First, isn't Procedural Synthesis patented by MS

No. Procedural Synthesis is an old and wide field. They've patented specific techniques relating to their CPU to aid procedural synthesis, a completely different point.

Second, rendering. You do realize that current Cell technology limits it to 256MB of XDR memory don't you? That makes it pretty useless to replace typical rendering setups...

Who's talking about replacing..? And Cell can access the full 512MB of memory. It has faster access to 256MB of that.

 

[BOOMEXPLODE]

DeanoC, or anyone who wants to guess: what about having Cell do T&L or just completely taking over for the vertex shaders? I think RSX has some pretty powerful vertex shaders, but how does their performance compare to VS running on Cell? NV40 had 6 vertex shaders, I think G70 has 8? Or possibly 12, it's not official ATM.

 

[nAo]

Why would you 'waste' CELL cycles to do some vertex shading when you have a GPU designed for it?
IMHO it's better to use CELL for other, even complementary, tasks.

 

[Titanio]

Why would you 'waste' CELL cycles to do some vertex shading when you have a GPU designed for it?
IMHO it's better to use CELL for other, even complementary, tasks.

Indeed, certainly moving ALL VS to Cell makes little sense since you'd just be idling perfectly good VS logic on the GPU. VS on Cell in tandem with the GPU, however, could make a lot of sense and let you do some cool stuff.

 

[version]

Why would you 'waste' CELL cycles to do some vertex shading when you have a GPU designed for it?
IMHO it's better to use CELL for other, even complementary, tasks.

1 SPE equal 8 vertexshaders speed, 32 pixelshader better design

 

[nAo]

VS on Cell in tandem with the GPU, however, could make a lot of sense and let you do some cool stuff.

Agreed.

1 SPE equal 8 vertexshaders speed, 32 pixelshader better design

What?

 

[version]

1 SPE equal 8 vertexshaders speed, 32 pixelshader better design

What?

why would use vertexshaders , when SPE is 8 times faster, or not ?
if gpu has only pixelshaders , it will be simple, and more pixelpower

 

[nAo]

why would use vertexshaders , when SPE is 8 times faster, or not ?

Vertex shader engines are custom hw designed to efficiently shade vertices, per any given transistor area VS will be more efficient than CELL on a similar fabbing process.

if gpu has only pixelshaders , it will be simple, and more pixelpower

But we know from Sony E3 conference this is not the case, RSX has vertex shaders.
The question is: since we have a GPU that supports vertex shading too why sould we left these vertex shaders unused?

 

[Wicked_Vengence]

Tacitblue: No to the 512MB per Cell. They way they are laid out, each cell can only address 256MB. The next step up (which IBM is talking about) is up to 2GB but hasn't been done yet. It could have had 512MB total...

 

[Tacitblue]

Here's an FYI for ya, It's a bit of an old topic but here's a thread and an article for you.

http://www.beyond3d.com/forum/viewtopic.php?t=23304

http://techon.nikkeibp.co.jp/article/NEWS/20050524/105016/

10x512mb chips per Cell, 2 Cells, 1 GB total.

 

[zidane1strife]

why would use vertexshaders , when SPE is 8 times faster, or not ?

Vertex shader engines are custom hw designed to efficiently shade vertices, per any given transistor area VS will be more efficient than CELL on a similar fabbing process.

if gpu has only pixelshaders , it will be simple, and more pixelpower

But we know from Sony E3 conference this is not the case, RSX has vertex shaders.
The question is: since we have a GPU that supports vertex shading too why sould we left these vertex shaders unused?

I'd hope it can at least aid in VSding, or that the RSX has substantial VS power. Right now many 360 games are missing just a bit more polys here and there to be truly uber, though there are some where it does suffice.

Even the luna demo would benefit from some polys here and there showing the need for more vertex power than is at present dedicated to such task. Maybe the faster clock allows it to genere sufficient geometry to allow for smooth objects, but if not, it's indeed a necessary thing.

 

[Shifty Geezer]

Seems to me Cell would ideally have been paired up with Xenos, to switch that whole GPU over to pixel shading and have Cell process the vertices. Unified shaders would offer more flexibility than fixed pipes.

 

[one]

Seems to me Cell would ideally have been paired up with Xenos, to switch that whole GPU over to pixel shading and have Cell process the vertices. Unified shaders would offer more flexibility than fixed pipes.

If unified shader is better, I guess Cell-based GPU might not be discarded in the first place... just my 2 cents.

 

[London Geezer]

Seems to me Cell would ideally have been paired up with Xenos, to switch that whole GPU over to pixel shading and have Cell process the vertices. Unified shaders would offer more flexibility than fixed pipes.

That's exactly the same thing i was thinking... And the same reason why i'm not too sure Sony did the right thing going for NVIDIA, whatever the outcome will be.
Ultimately PS3 games will look fantastic and everyone will be happy, but i can't help wondering the point of some of their decisions...

Or of course, as usual, i might just be missing something.

 

[Wicked_Vengence]

Tacitblue:
That was 512megabit, not 512megabyte. Translates to 64MB *2 for 128MB.

In the XDR memory system (which the cell uses for control), each channel can support a maximum of thirty-six devices connected to the same command and address bus (currently). The data bus of each device connects to the memory controller through a set of bi-directional point-to-point connections. In the XDR memory system, address and command are sent on the address and command bus at a rate of 800 Mbits per second (Mbps), and the point to point interface operates at a datarate of 3.2 Gbps. Using DRAM devices with 16 bit wide data busses, each channel of XDR memory can sustain a maximum bandwidth of 102.4 Gbps (2 x 16 x 3.2), or 12.6 GB/s. The CELL processor can thus achieve a maximum bandwidth of 25.2 GB/s with a 2 channel, 4 device configuration.

The obvious advantage of the XDR memory system is the bandwidth that it provides to the CELL processor. However, the maximum of 4 DRAM devices means that the CELL processor is limited to 256 MB of memory, given that the highest capacity XDR DRAM device is currently 512 Mbits (64MB).