NVIDIA Sony Graphics Interview

[Sonic]

Why would Nvidia integrate CELL into the GPU at the very last minute? Think about it in terms of understanding the concept of CELL. CELL is an architecture designed to be scalable, powerful, but also to be a CPU. I'm sure you could customize CELL to the point where it resembles nothing like the CELL CPU going into PS3.

Basically, Sony doesn't need to have CELL be the GPU because the CPU will be CELL. Nvidia and Sony both understand that the GPU will be awesome on its own and it would be foolish to complicate if further for it to be CELL based. There will need to be customizations made to the chip for it to communicate and interface with the CELL CPU in PS3, but that doesn't mean it will be comprised of CELL.

IBM, Sony, and Toshiba all designed CELL for their own purposes. The adaptability of CELL has yet to be seen and we will see in the future if it can fill as many roles as some think it can.

 

[Megadrive1988]

I tend to agree with Sonic on that one. that's the most likely approach.

Panajev, are you in agreement or not? forgive me, I know you've stated your thoughts MANY times, but I tend to forget since you explain so thouroughly the architectures

 

[Trawler]

One other limiting factor for modificatons to the next gen nv core is timeframe.

I suspect Sony would like this chip delivered ASAP - perhaps for a 1st Q launch in Japan next year.

What type of timeframes are we talking about for a next gen GPU to go from tapeout to production silicon? Up to 6 months (if you include revisions)?

So what does that give nv/Sony? Perhaps a little over 6 months for modifications? Hardly long enough to make any substantial changes. Maybe embedded RAM and a few other tweaks?

Or am I way off base with my timeframes?

 

[Mythos]

And we’ve been working with them to produce a customized version that is customized specifically to connect that to the cell processor, so that they could work together.

I think we are looking at the same speech with customization aspect of the NV part with Cell but saying different things. The Cell is a stream processor and the GPU is a form of that. So we are looking at some similar aspects of each Chip part. Can we say that it will be cell like well.. yeah. Can we say that it will be a customized solution of the PC part GPU well..yeah. Each aspect of the Chips (Cell/GPU) have to meet with each other to effectively work effectively in the total system.

....so in terms of not just the 3D graphics but also in terms of the media processing, in terms of integration of different media types, and in terms of the role it going to play especially relative to what they see happening with the cell processor…

So it’s a pretty ambitious thing that we are doing with 3D graphics technology. And we are thrilled to be cooperating on the graphics side.

We will be rolling a new graphics solution out as a GPU in addition to what we are doing with Sony.

This last quote sounds like 1. There is deep intergration involve on both the Celluar aspect and the GPU apect of the PS3 system. 2. There is an exchange and introduction of new technology from both the Cell side and the GPU side going into each others next generation products.

 

[ninelven]

erased

 

[YeuEmMaiMai]

Edited by moderator:


Flaming and trolling will not be tolerated on this board. This is your first warning. If you do so again you will be banned.

 

[SedentaryJourney]

Something I've wondered about the GPU in the PS3 is if CELL is handling all the vertex shading, does that make all the vertex shaders on the GPU redundant or even disposable? And wouldn't that equal more headroom for pixel shading?

 

[V3]

Basically, Sony doesn't need to have CELL be the GPU because the CPU will be CELL. Nvidia and Sony both understand that the GPU will be awesome on its own and it would be foolish to complicate if further for it to be CELL based.

Even from the patent those pixel engine aren't cell, they were just attach to the PU like all the APU, to form something they called visualiser. This visualiser unit is the lowest client side spec, which could be the lowest spec for Cell device.

There will need to be customizations made to the chip for it to communicate and interface with the CELL CPU in PS3, but that doesn't mean it will be comprised of CELL.

The CPU is still the Power Processor, not the entire Cell, That's why those SPUs were called APUs, because they are attach to the Power Processor, in similar way this NV GPU will attach to the Power Processor in term of their cellular architecture.

Regardless of the final chips configuration, the NV GPU is still treated simlarly as the other attach device that can be further attached to Cell, to form what ever device intended for the job.

 

[Sonic]

V3, I'm sure you know what I meant. I don't want to get in the specifics of CPU architecture or architectures in general. Basically I was saying that Sony doesn't need CELL to be the GPU when they have a perfecttly fine one from Nvidia that is probably up to the task in a much better way.

CELL is still the heart of PS3, regardless of which portion will be the CPU. That's just like us saying that Emotion Engine is the CPU in PS2 but really only the MIPS portion is.

 

[V3]

V3, I'm sure you know what I meant. I don't want to get in the specifics of CPU architecture or architectures in general. Basically I was saying that Sony doesn't need CELL to be the GPU when they have a perfecttly fine one from Nvidia that is probably up to the task in a much better way.

The graphic tasks will be divided between SPU and GPU. I'll assume developer will choose, which tasks goes where.

What I am addressing is some people here still insisting that NV GPU will be cell base, I am just saying that it needn't be that way, because the GPU will be attached to cell, regardless of the final chip configuration. So I was affirming your point, with my reason.

The final config for PS3 is still up in the air, but its lowest config from the info gather thus far is this, cell + NV GPU. But they can go further if budget stretch, upto config like Broadband Engine + Cells + NV GPUs which probably the upper limit of PS3 specs.

 

[rabidrabbit]

If the "Cell" CPU is 1 Teraflops or even 256 GFlops, and if the GPU is just a next gen nVidia PC GPU modified for PS3, would the GPU be a bottleneck in PS3?
Would the 1 Teraflops go to waste?
I mean if you take a PC with a high-end nVidia GPU, and a mid-end Intel/AMD CPU, there's usually not that much to be gained if you replace the CPU with a top speed one. Sure ther will be improvements, but at some point the GPU will become the limit.
Will a next gen nVidia GPU be that powerful that if you buy one next year, it'll still be the fastest component in your PC when PC processors reach 256 Gflops -> ?

If the "Cell" is such a beast as hyped, wouldn't the next gen nVidia GPU have to also be quite a speed demon to match the processing power of "Cell". Surely it would have to be a much bigger leap than what we've seen for a long time in nVidia's and ATi's roadmaps.

But as PC's of near future (year or two ahead) do not seem to match that hyped 256Gflops - 1 Tflops of "Cell", and if the next gen nVidia GPU is primarily designed for these PC's, with targeted speeds that would compliment said PC tech, wouldn't that mean this same next gen nVidia GPU tech would be too slow for the hyped "Cell" speeds?
If the nVidia next gen GPU is just slightly modified for PS3 as suggested here, could it become the bottleneck in PS3?

Or could they just make the next gen nVidia tech scale well, maybe there will several of thjose processors inside the PS3.

Someone said the "Cell" would do the vertex stuff, while GPU is more just a "pixel engine". Then the 1 Teraflop "Cell" would have more to do in graphics.

 

[GwymWeepa]

If the "Cell" CPU is 1 Teraflops or even 256 GFlops, and if the GPU is just a next gen nVidia PC GPU modified for PS3, would the GPU be a bottleneck in PS3?
Would the 1 Teraflops go to waste?
I mean if you take a PC with a high-end nVidia GPU, and a mid-end Intel/AMD CPU, there's usually not that much to be gained if you replace the CPU with a top speed one. Sure ther will be improvements, but at some point the GPU will become the limit.
Will a next gen nVidia GPU be that powerful that if you buy one next year, it'll still be the fastest component in your PC when PC processors reach 256 Gflops -> ?

If the "Cell" is such a beast as hyped, wouldn't the next gen nVidia GPU have to also be quite a speed demon to match the processing power of "Cell". Surely it would have to be a much bigger leap than what we've seen for a long time in nVidia's and ATi's roadmaps.

But as PC's of near future (year or two ahead) do not seem to match that hyped 256Gflops - 1 Tflops of "Cell", and if the next gen nVidia GPU is primarily designed for these PC's, with targeted speeds that would compliment said PC tech, wouldn't that mean this same next gen nVidia GPU tech would be too slow for the hyped "Cell" speeds?
If the nVidia next gen GPU is just slightly modified for PS3 as suggested here, could it become the bottleneck in PS3?

Or could they just make the next gen nVidia tech scale well, maybe there will several of thjose processors inside the PS3.

Someone said the "Cell" would do the vertex stuff, while GPU is more just a "pixel engine". Then the 1 Teraflop "Cell" would have more to do in graphics.

If Cell does come close to estimates, then yeah the GPU could become a bottleneck (this coming from a layman), perhaps Sony chose Nvidia partly due to its SLI technology...or maybe Nvidia will tweak the GPU to only have pixel shaders while Cell does all the vertex shading and perhaps all the lighting.

 

[Alejux]

I don't thin the GPU will become a bottleneck. There is A LOT of processing to be done in a game. Some of which have almost no direct relation with graphics. Even if we use a full 1 teraflop/s of processing power, we wouldn't come close to what we really need for perfect physics simulations and a decent artificial intelligence.

IMO, the only bottleneck in the PS3, will be the lack of memory.

 

[jvd]

I don't thin the GPU will become a bottleneck. There is A LOT of processing to be done in a game. Some of which have almost no direct relation with graphics. Even if we use a full 1 teraflop/s of processing power, we wouldn't come close to what we really need for perfect physics simulations and a decent artificial intelligence.

IMO, the only bottleneck in the PS3, will be the lack of memory.

well a gpu can become a bottleneck.


Think about rendering doom 3 on a 5ghz p4 and a geforce 3 .... which do you think is the bottle neck in that system ?


So even if the cell chip is 1tflop the gpu may not be able to keep up with it


Though i expect them to get paper spec of 256gflop and a actual sustained in game rate of around 60-70gflops (which would still be great performance

 

[London Geezer]

Memory will always be the main bottleneck in consoles, but it's easy to create examples when either the CPU is the bottleneck, or the GPU, or something else.

 

[DemoCoder]

I don't buy the notion that a GPU needs to be "clean sheet" for a console anyway. In the past, this was the case because they could take advantage of the fact that they only needed to support 480p out and design the GPU around that framebuffer size (e.g. edram)

There are two major differences between consoles and PC game systems: display device and input device.

However, next-gen consoles should support 720p, 1080i, and 1080p which puts them in the same realm as PC graphics. Thus, the use cases between the architectures are similar in terms of display.

Internally, the GPU's don't need to differ signifcantly from PC designs, except in how they interface with the CPU and memory/IO buses, and maybe some timing tweaks. As far as pipeline configuration goes, I don't see any huge gains that a "clean" chip could have vs a "tweaked" GPU.

Ergo, I see no big advantage in a hypothetical Xenon GPU "designed from the beginning for console" (which itself is somewhat bogus, since I would bet, it's a PC design part refactored and borrowed over 12+months anyway). For example, if you were going with eDRAM, on the PC just dictate that you support 1920x1080 or 1920x1200 max, and there's your design point.

A hypothetical NV5x, with memory and northbridge logic deleted, and replaced with something CELL/PS3 specific isn't going to face any insurmountable performance issues.

I'm of the opinion that two much inter-chip communication is bad for performance anyway. I like asynchronous stream designs.

 

[London Geezer]

I'm of the opinion that two much inter-chip communication is bad for performance anyway. I like asynchronous stream designs.

Funny you say that, i was thinking about this the other day, and i didn't know whether there are any advantages in asynchronous systems, like GC, as opposed to synchronmous ones like PS2. I thought synchronous designs were always preferred, but it seems there are some advantages to asynchronous ones.

Do u have time to elaborate?

 

[MfA]

Well true asynchronous isnt an option, since it doesnt break down gracefully ... if you can stop pushing information in once the FIFO is full then it is semi-synchronous, even if it is very large and goes through memory.

 

[n00body]

too narrow a viewpoint

The people stressing over the GPU being a bottleneck are making a fuss over an unlikely problem. The way I see it, they're spending too much time thinking in traditional computer terms and not in terms of CELL's peculiar new design philosophy (especially with all the examples of GFX cards running in PCs). They don't consider that CELL will have its own capabilities and quirks that could be vastly different from what we have become used to.

With that in mind, my theory on how the GPU will integrate is that it will interface directly with CELL and be treated as though it were just another PE (albeit a highly specialized one). That way, the GPU would be directly controlled by CELL and would reap the benefits of CELLS high speed bus. It would also allow CELL's other PEs to work with the GPU if it became necessary.

Any thoughts on this would be most appreciated.

 

[j^aws]

Re: too narrow a viewpoint

The people stressing over the GPU being a bottleneck are making a fuss over an unlikely problem. The way I see it, they're spending too much time thinking in traditional computer terms and not in terms of CELL's peculiar new design philosophy (especially with all the examples of GFX cards running in PCs). They don't consider that CELL will have its own capabilities and quirks that could be vastly different from what we have become used to.

With that in mind, my theory on how the GPU will integrate is that it will interface directly with CELL and be treated as though it were just another PE (albeit a highly specialized one). That way, the GPU would be directly controlled by CELL and would reap the benefits of CELLS high speed bus. It would also allow CELL's other PEs to work with the GPU if it became necessary.

Any thoughts on this would be most appreciated.

What your proposing is basically a CELL based GPU. You might want to search the forums as this has been discussed many times here...but here's a recent thread to get you started,

http://www.beyond3d.com/forum/viewtopic.php?t=19299