The Playstation 3 Before RSX

[Color me Dan]

I remember the hype leading up to the Playstation 3 launch, and I was gobbling up news and rumours left and right. But this was well over a decade ago now. I hardly remember what I ate last night. But I do remember a rumour that Sony was initially vying to get the Cell processor, or a version of it, to act as the graphics processor.

So does anyone happen to know more about what the rumoured pre-RSX based graphics solution was supposed to be? Or when it was dropped? And when was nVidia approached for a solution?

The nVidia RSX wasn’t too shabby, but I remember getting the impression that it was sort of a late adjustment. Though that impression might be because the vaunted Cell processor made a semi-custom 7800GTX seem humdrum by comparison. Especially when the Xbox 360 ATi Xenos GPU featured unified shaders, and that interesting eDRAM solution, which was definitely more “next-gen”.

 

[Sega_Model_4]

They would only use Cell or two of them. Who made Sony change its mind was ICE team.

“For a while, [PS3 had] no GPU, it was going to run everything with SPUs. The ICE team proved to Japan that it was just impossible. It would be ridiculous. Performance-wise, it would be a disaster. That’s why they finally added the GPU, closer to the end.”

“I think that the hardware guys focused too much on getting the [Sony’s proprietary processor] CELL working that the GPU project [for PlayStation 3] ultimately fell behind.”

https://archive.ph/uocPt#selection-1311.23-1311.196

 

[Color me Dan]

I've got to say that the last quote has me really curious. I wonder what that GPU project looked like. The SPU's had features that were useful for graphics processing, like the vector processing capabilities for one thing. But you would have needed more than that. Just having an extra Cell processor on GPU duty feels to obviously unfit for purpose. Sony must have been working on a more tailored version. Do you happen to have more sources?

 

[Sega_Model_4]

I've got to say that the last quote has me really curious. I wonder what that GPU project looked like. The SPU's had features that were useful for graphics processing, like the vertex processing capabilities for one thing. But you would have needed more than that. Just having an extra Cell processor on GPU duty feels to obviously unfit for purpose. Sony must have been working on a more tailored version. Do you happen to have more sources?

I believe it would be similar to what the PS2 does, with Cell doing vertex transformations and another Cell doing rasterisation. There are not much sources out there, besides these, probably because it was something changed during development (and the cause of the console delay to 2006).

Take a look at Coppeti's material, he writes about the architecture of various consoles.

PlayStation 3 Architecture | A Practical Analysis

An in-depth analysis that explains how this console works internally

www.copetti.org

 

[Color me Dan]

I believe it would be similar to what the PS2 does, with Cell doing vertex transformations and another Cell doing rasterisation. There are not much sources out there, besides these, probably because it was something changed during development (and the cause of the console delay to 2006).

Take a look at Coppeti's material, he writes about the architecture of various consoles.

PlayStation 3 Architecture | A Practical Analysis

An in-depth analysis that explains how this console works internally

www.copetti.org

That was a really enjoyable read. Thank you for that. I'm gonna have to read about the other consoles as well, that author has a knack for clear and concise, yet engaging, technical writing!

I suppose we might never really know what the initial plan was unless Sony, or a Sony insider, opens up on the topic. And I equally suspect nVidia will remain mum about any insights they might have gleaned as they got attached. But boy would I love to know how the Cell + Cell solution would have performed, even theoretically.

On the other hand it seems a bit of a shame that the Cell GPU project seemed to have remained until the last moment. There could have been some potential for Cell to offload RSX, and for RSX to reach for the XDR memory more regularly, with some more tweaking if they'd been given time.

 

[Sega_Model_4]

That was a really enjoyable read. Thank you for that. I'm gonna have to read about the other consoles as well, that author has a knack for clear and concise, yet engaging, technical writing!

I suppose we might never really know what the initial plan was unless Sony, or a Sony insider, opens up on the topic. And I equally suspect nVidia will remain mum about any insights they might have gleaned as they got attached. But boy would I love to know how the Cell + Cell solution would have performed, even theoretically.

On the other hand it seems a bit of a shame that the Cell GPU project seemed to have remained until the last moment. There could have been some potential for Cell to offload RSX, and for RSX to reach for the XDR memory more regularly, with some more tweaking if they'd been given time.

Apparently this ended up being used later. Developers such Naughty Dog used part of the SPUs to assist RSX.

 

[swaaye]

If you search the forum you will find more talk about this. If I remember right there was some kind of Toshiba GPU before NVidia got involved.

For example,

Why did Sony use a G70 RSX instead of something better? *spawn

There was another source* that said Sony worked on another GPU in the style of the RS in the PS2 for PS3. The SPU's would do T&L and that would just rasterize/shade stuff. But it became to big so they had to do with nVidea instead. *that I can't find now....

forum.beyond3d.com

 

[Color me Dan]

If you search the forum you will find more talk about this. If I remember right there was some kind of Toshiba GPU before NVidia got involved.

I took your advice, and I've come across a few really interesting threads that contain some of the info I'm looking for. Granted, it's still mostly speculation and rumours. But I think the most interesting thing I found so far was a two part blog/article by the user Urian here on the forum (here is part I and part II through Archive.org). They seem to have compiled a development history of the RSX back in the day. And It's great reading backed up by some good articles and well reasoned guess-work.

Essentially it seems as though Kutaragi was aiming to use the PS2 EE and GS internals as a jumping off point for further development. Potentially leading to the PS3. Something Kutaragi layed out more or less verbatim (and Urian linked to) in an EE article. It's interesting reading in itself.

From that it seems as though a Toshiba derived GPU, with some possible family ties to the GS chip, was initially planned for in early stages of development. When and how Sony and Toshiba shifted gears towards a Cell is unclear. But as we all know Cell went from idea to reality, and the GPU work must have followed suit. It seems to be this patent that people refer to when they say they've heard of this Cell iteration. A.k.a. the "Cell Visualizer".

Now I'm a layman, so I can't make heads or tails of that patent and whether or not it would lead to a feasible and potentially competetive GPU. But my exceptionally vague understanding of it hints at some potentially interesting flexibility for developers. Even if the coding part might have been uhm... trickier still. A PS3 where two Cell dies, one with added hardware for rendering tasks, could share resources over the FlexI/O interconnect rather directly, and a shared pool of memory.

 

[Shifty Geezer]

I don't believe Cell existed as the evolved idea of PS2's paradigm. At the point in history when PS3 was being considered, GPGPU - non-graphics work on the GPU - was in its infancy, while data-centric workloads were clearly the future. There was obviously a need for processing that was more programmable than a GPU but faster at data crunching than a CPU, and a clean slate design that wasn't trying to take one existing paradigm and stretch it to another was worth a shot. The idea of Cell made sense and was seemingly a vision shared at IBM, although maybe they just went along with it for fun and monies.

Another key consideration for Cell was that it wasn't just a PPU and 8 SPUs, but it was a heterogenous multicore processor allowing for alternative cores to what we think of as SPUs - the Visualisers you talk of. Cores could be specialised and assembled on the Ring Bus for fast intercore communication.

Toshiba apparently had a GPU idea, but we know next to nothing about it other than it was dropped at the last minute and nVidia adopted in a clearly-less-than-ideal solution that added cost and complication.

However...here's the fun part. Rendering has moved on a lot from the techniques that existed at the end of PS2 and into PS3. GPUs are doing far more compute work and non-drawing work, and software-rendering provides more solutions than the fairly fixed hardware of the time. GPUs are bonkers wide now, but at the same time they can struggle to operate efficiently given reliance on processing things like pixel clusters. Cell potentially offers more programmability and granularity and faster core speeds. 65nm Cell was 120 mm² and apparently a SPE could run at 40° at 4 GHz.



Counting 6 complete node shrinks to 10 nm (I think by that point we're not seeing a full halving any more?), 120 mm² would have shrunk to 2mm². Given a GPU-like die area of 500 mm², that'd fit 250 Cells. At around 250 GF per Cell, that's 62 TF at 4 GHz.

Of course, such a rough number doesn't mean much. The Cell's would be evolved from that, more LS, etc. But on paper it's not impossible for Cell to have evolved into something meaningful.

The main issue was coding the thing, and competition with the conventional GPU market that was evolving faster and more in line with current thinking meant Cell was never going to succeed commercially. Regardless of its theoretical potential, it was a business dead-end. We can at this point only speculate. However, I don't think Kutaragi et al were entirely mental to consider it. Slightly, perhaps, but Kutaragi had faith
software engineers would do incredible things given the opportunity, and that would have been true if Cell hung around in a viable form for generations. I'd love to see what a Cell could do with the latest graphical thinking.