Return to rasterizer for PS4?

[tenken]

Hi, watching the cell road map I was wondering if it will make sense for sony to return to a rasterizer for PS4. According to IBM they will have a 2 ppe 4 spe 45nm cell ready by 2010 at 1 tflop at current clock speeds and more if they raise it to 4ghz or 5ghz. If sony uses this there could be a lot of redundency and unnessecery expenses by putting in a full blown future gpgpu.

If articles are correct cell is supposed to be excellant at ray-tracing, vertex shading and pixel shading but sucks at texturing. Nnidia has already ported over cg for cell. So say for a future game work load there could be the 2 ppes (hopefully beefier with ooe) and 6 spes working on your ai, physics, sound etc code while you could divide the remaining 26 cores for graphics like a unified shader, say 8 for pixel, 8 for vertex and the remaining 10 for ray-tracing or whatever setup.

Then there could be a separate chip similar to the xenos daughter die but for rasterizing with programmable tmus (for texturing) rops and any other gpu logic needed with a nice blok of edram. If the get down to 32nm I suppose they could have it packaged with cell like the EE + GS. It should be cost efficient and if sony is learning from nintendo with the WII they should be wary of costs and look to launch something no more than $349 at profit and focus on the gameplay aspect.

I will expect microsoft to continue with smp maybe 6-8 cores with the latest gpgpu. A nice battle ground for the cell vs gpu debate. So feaseable or wrong all over.

 

[nonamer]

Very bad idea. Even if the power of a Cell2 matches that of a GPU, it's still not going very efficient at doing pixel shading in particular and other shader related work. I would still go with a full GPU.

 

[Laa-Yosh]

It's not just pixel shading (which it actually shouldn't be bad at) but texture sampling, filtering, caching; triangle setup, alpha blending, lots of interpolations and such. In short, all the fixed function computing power of a GPU would have to be run in software and it'd be a waste to use the CPU for this.

 

[Shark Sandwich]

It seems likely that Sony considered such a setup with the PS3. All the early hype about Cell focused on how Cell chips were supposed to work together. For a while, up until PS3 was unveiled, Sony denied that their dealings with Nvidia had anything to do with Playstation 3. Then, PS3 turned out to have only one Cell and an Nvidia GPU.

My guess, and I have heard this from other forum members here and elsewhere, is that a multi-Cell no-GPU setup was one possible candidate for PS3. They likely decided against such a setup for performance reasons. You can't beat a dedicated GPU for graphics.

Sure, in 5 years from now, IBM will have a third-generation Cell chip that probably runs at 6+ GHz. Don't forget, though, that GPUs will advance just as much, if not more. GPUs seem to advance even faster than CPUs. In 5 years, using a Cell in place of a GPU will probably make even less sense than it does now.

 

[inefficient]

Sure, in 5 years from now, IBM will have a third-generation Cell chip that probably runs at 6+ GHz. Don't forget, though, that GPUs will advance just as much, if not more. GPUs seem to advance even faster than CPUs. In 5 years, using a Cell in place of a GPU will probably make even less sense than it does now.

The thing is, GPUs are becoming more and more CPU like. And if we get a Cell with 32+ SPUs, that CPU is going to show analogs to a GPU technology in several ways.

Their paths are beginning to appear to converge somewhat. GPU's certainly will not be able to maintain the year over year performance trend they have maintaining forever. It's just a matter or when.

So when you say "Don't forget, though, that GPUs will advance just as much". You can't really take that fact for granted. The performance discrepancy on embarrassingly parallel problems between a GPU and a CPU will get smaller and smaller as CPUs get more cores and GPU execution units become more general purpose.

The gap "probably won't" be anywhere near closing in just 5 years. But predicting the future is a very tricky thing.

 

[Titanio]

For a while, up until PS3 was unveiled, Sony denied that their dealings with Nvidia had anything to do with Playstation 3.

Not that it's relevant, but the day their collaboration with nVidia was announced, they confirmed it was for the GPU in the next PlayStation.

 

[Dunchan25]

The thing is, GPUs are becoming more and more CPU like. And if we get a Cell with 32+ SPUs, that CPU is going to show analogs to a GPU technology in several ways.

Their paths are beginning to appear to converge somewhat. GPU's certainly will not be able to maintain the year over year performance trend they have maintaining forever. It's just a matter or when.

So when you say "Don't forget, though, that GPUs will advance just as much". You can't really take that fact for granted. The performance discrepancy on embarrassingly parallel problems between a GPU and a CPU will get smaller and smaller as CPUs get more cores and GPU execution units become more general purpose.

The gap "probably won't" be anywhere near closing in just 5 years. But predicting the future is a very tricky thing.

Maybe we'll get a console with cell chips of specific tasks (cell for lights, cell for sound, cell for physics, cell general purpose) all working with a modern GPU. By that time everyone with understand how to get the most out cell so these task specfic cells could be ideal

 

[Laa-Yosh]

The thing is, GPUs are becoming more and more CPU like. And if we get a Cell with 32+ SPUs, that CPU is going to show analogs to a GPU technology in several ways.

There's a lot more to GPU's than programmable parts and those are quite hard to efficiently emulate on a generic processor. You'll need the actual rasterizing and other units I've mentioned in actual hardware to keep the speed of today's solutions.

 

[Titanio]

I wonder if it would be worth Sony asking someone like nVidia to integrate the necessary silicon for GPU-level texturing/interpolation/filtering/rasterisation etc. right onto a Cell chip. And then sticking two of them into a machine. So the SPUs would do all the computational stuff shaders do, and have the back-up of the more fixed-function stuff GPUs benefit from.

In the same way shaders have become unified, processing generally could become unified under such a architecture. We're already seeing it happen in an early, less sophisticated sense with people offloading stuff to the GPU (like physics)..having just one homogenous chunk of processing power to load balance as the developer sees fit seems like a logical, potential, end-game. I don't know if the time would be right with PS4 or not, though..it kind of depends how quickly things evolve from here.

 

[PiNkY]

I suppose if it would have been feasible this gen, sony would have done it. A 2 Cell configuration would probably have been noticably cheaper.

 

[Iron Tiger]

I suppose if it would have been feasible this gen, sony would have done it. A 2 Cell configuration would probably have been noticably cheaper.

The most credible rumor I read was that Sony was planning a multi-CELL configuration with something like a "GS2", but a polling of developers showed a desire to move towards a more simplified architecture with programmable shaders and such.

 

[taucias]

The most credible rumor I read was that Sony was planning a multi-CELL configuration with something like a "GS2", but a polling of developers showed a desire to move towards a more simplified architecture with programmable shaders and such.

It's not a rumour, it's a fact. The original patent was for a Visualizer using a custom CELL style chip:
http://www.aceshardware.com/read.jsp?id=60000291

That link explains more.

 

[emacs]

Hi, watching the cell road map I was wondering if it will make sense for sony to return to a rasterizer for PS4.

dedicated hardware for graphics does have its benefits. i am, however, keenly interested to see what future iterations of the Cell can do.

 

[inefficient]

It's not a rumour, it's a fact. The original patent was for a Visualizer using a custom CELL style chip:
http://www.aceshardware.com/read.jsp?id=60000291

That link explains more.

That patent does not resemble Cell at all as we know it today.

It only seemed relevant back in 2003 back when we had almost no information.

Whatever that design is for, it clearly has nothing to do with the Cell BE enabled PS3.

 

[Crossbar]

I wonder if it would be worth Sony asking someone like nVidia to integrate the necessary silicon for GPU-level texturing/interpolation/filtering/rasterisation etc. right onto a Cell chip. And then sticking two of them into a machine. So the SPUs would do all the computational stuff shaders do, and have the back-up of the more fixed-function stuff GPUs benefit from.

Sounds appealing!!!!

Two 32 SPE-Cells, where the SPE instruction set has been expanded with some graphics specific instructions and some ROP-units, etc have been added.
By having just one type of processing unit you can reduce the development cost for the (PS4?) and you get higher production volumes for that component faster.

64 SPEs in total should be enough to do some decent ray-tracing.

 

[Capeta]

That patent does not resemble Cell at all as we know it today.

It only seemed relevant back in 2003 back when we had almost no information.

Whatever that design is for, it clearly has nothing to do with the Cell BE enabled PS3.

Actually it does, sounds like revisionist history to me. I don't think the patent mentioned the Broadband Engine, Visualizer and CELL for nothing.

 

[Bigus Dickus]

It might be possible to add 'fixed-function' GPU components to the CELL architecture to make it more suited for raster rendering, but I'm not really sure what the purpose would be. That's is equivalent to telling nVidia or AMD/ATi to start with the CELL architecture and design a GPU around it. Is that really the best approach to designing something optimized for a task? I would think the result would be a decent GPU with some die space occupied by non-graphics related logic, that isn't utilized much, while sacrificing some GPU specific logic that could have been included otherwise.

Or looking at it the other way, if it was used for general purpose "CPU" code, the resulting chip might be a decent CPU with some die space occupied by graphics related logic, that isn't utilized much, while sacrificing some general purpose logic that could have been included otherwise (beefier BPU, deeper cache, whatever).

I think people are looking at this from the wrong vantage point. What CELL has demonstrated IMO is not that convergence is driving us towards unification, but rather that as our ability to stuff more transistors on a chip increases we will see more divergence and specialization. Obvious case in point: CELL uses an asymmetric design with one PPE and supporting SPE's, compared to the AMD/Intel/IBM(Xenos) approach of symmetric multi-core. I've postulated that we will see more of this in the future... after AMD and Intel initially stuff more cores on a chip they will reach a point where more 'general purpose' cores don't help much, and we will start seeing specialized cores. i.e., after the Intel Core Duo iterations reach 4 or 8 general purpose cores, I wouldn't be surprised to see SPU like "number crunching" support cores, optimized for streaming, FPU crunching, matrix manipulation, whatever. CELL is the first contemporary step in that direction.

So perhaps the question should not be whether CELL would make a good GPU, but rather is it a good starting point for a GPU architecture. I see no reason why it would be any better suited for that task than the massively paralleled designs AMD/ATi and nVidia have developed after decades of refinement.

 

[Shifty Geezer]

I don't think the worry about bits of chip sitting idle is necessarily too great. We already have that with VMX units and the like not being used 100% of the time. Similarly GPU's have moments where components are sitting idle. If combining the activities onto one chip increases overall utilisation, even though utilisation could never be 100%, it's a positive step. The idea of combined functions also has other benefits like scalability and a uniform programming model across different hardwares.

I think the real concern is whether adding key GPU functions to a SPE would provide performance comparable with the same amount of silicon real-estate as a discrete processor. The choice would be either one CPU @ 250mm^2 and one GPU @ 250 mm^2, or two CPGPUs at 500 mm^2. Or whatever sizes. Basically, given a cost per transistor, how much advantage does specialization bring you? And how much advantage does multi-function get you? Ultimately, load balancing could be much better on a combi-CPGPU. Just as unified shaders let a GPU switch between vertex and pixel workloads, graphically-enhanced SPEs could switch between vertex, pixel, and non-graphics workloads. Idling time would be minimal. You'd also have more targetable power where you want it. A graphically simple game would benefit from more available processing power, and a simple game could aply more SPEs to graphics work. That's something we're talking abut now with PS3, but taken to the next step. And the graphics enhancements may well find their use outside of graphics work too. Just because a texture-unit is called a texture-unit, doesn't mean it has to be accessing and interpolation textures!

 

[Bigus Dickus]

I don't think the worry about bits of chip sitting idle is necessarily too great. We already have that with VMX units and the like not being used 100% of the time. Similarly GPU's have moments where components are sitting idle. If combining the activities onto one chip increases overall utilisation, even though utilisation could never be 100%, it's a positive step. The idea of combined functions also has other benefits like scalability and a uniform programming model across different hardwares.

This sounds like an argument for using idle SPE's to assist in graphics work and idle GPU pipes/arrays/units to assist in e.g. physics calculations.

I think the real concern is whether adding key GPU functions to a SPE would provide performance comparable with the same amount of silicon real-estate as a discrete processor.

Bingo.

About the closest thing I could see would be something either half-way between dedicated CPU/GPU and symmetric cores for CPU/GPU or nowhere near either, depending on your perspective. And that would be a "general purpose" CPU core(s) like the PPE in CELL, plus a "general purpose" GPU core(s) containing dedicated texturing/ROP/etc. logic, plus an array of "SPE/vertex/pixel" cores that were available for both CPU and GPU tasks. Kind of an extension of the Xenon's unified pixel/vertex shader concept. Sounds a lot like "CELL + GS2 rasterizer" to me, and apparently that combination, at least for the forseeable future, isn't optimal.

It's an interesting idea though.

 

[Megadrive1988]

SCEI does not have to go back to just having a rasterizer in PS4. they can have a full blown GPU, but what they need is a highly custom, incredibly high-bandwidth (via embedded memory) graphics processor that is ahead of PC GPUs in performance as much as PS2's GS was ahead of PC GPU performance (though not image quality) in 2000.

the CELL-based Visualizer GPU shown in diagrams back in 2003 indicated a chip with 4 simpler PPEs, 16 APUs, 4 pixel engines (perhaps with many ROPs each),
image caches and maybe EDRAM.


Xenos daughter-die internal EDRAM bandwidth is 256 GB/sec so I'd expect both next-gen consoles to have several TB of EDRAM bandwidth and several hundred GB/sec of main system memory bandwidth.


as for the PS4 GPU itself, I don't know what type of rendering we can expect from it-- if it'll be a departure from what we've seen so far, or a massively more powerful implementation of current rendering.