will PS3's GPU be more modern than PS2's GS for its time?

[j^aws]

I do not think PlayStation 3 will fully go the way the first patent showed: namely a wholly CELL based system with a CELL based Visualizer.

Are we only just coming to this realisation?

I'm sure you're aware of this but just to clarify, for the GPU to be Cell based, it just needs to run 'software Cells' and doesn't have to be the Visualizer from the patents or configured like a S|APU from the CELL CPU. They only need a consistent ISA to run 'software Cells'!

Just ask DeanoC about ISA's, G5's and Xenon CPU!

 

[j^aws]

hmm?

Could someone translate the text in that diagram please. Thanks in advance!

Happy New Year!

 

[Panajev2001a]

I do not think PlayStation 3 will fully go the way the first patent showed: namely a wholly CELL based system with a CELL based Visualizer.

Are we only just coming to this realisation?

I'm sure you're aware of this but just to clarify, for the GPU to be Cell based, it just needs to run 'software Cells' and doesn't have to be the Visualizer from the patents or configured like a S|APU from the CELL CPU. They only need a consistent ISA to run 'software Cells'!

I give up .




The NV5X architecture was not developed with CELL in mind or using the Apulet ISA/runnign the same Apulets and required several years of work and hundreds of engineers to be completed: wouldn't you think that nVIDIA would need more than 50 engineers for a bit more time to convert all their Shading cores to support the execution of Apluets/Software cells ?

You will see more and more collaboration between CPU and GPU in the PC when the I/O bus that connects them gets faster and faster (PCI-Express is already pushing towards it and PCI-Express 2.0 goes one step forward in the same direction delivering even more bandwidth).

This is done without having the PC GPU adopt the x86 ISA.

The point of having the GPU be CELL based was the point behidn the Visualizer: using the strengths of distributed processing well built in the CELL architecture to their full extent.

Eventually the CELL philosophy really leads to a BIG CELL IC which works on general purpose processing, rendering, A.I., Physics, etc... attached to a co-processor which only provides less and less hardwired functions (in the future you will want programmable AA, programmable Texture Filtering, etc...).

 

[Dave Baumann]

I'm sure you're aware of this but just to clarify, for the GPU to be Cell based, it just needs to run 'software Cells' and doesn't have to be the Visualizer from the patents or configured like a S|APU from the CELL CPU. They only need a consistent ISA to run 'software Cells'!

I’m fairly sure you know that’s not what Pana was talking about.

I do not think PlayStation 3 will fully go the way the first patent showed: namely a wholly CELL based system with a CELL based Visualizer.

Are we only just coming to this realisation?

That got me laughing. Some, who will remain nameless, were giddy as schoolboys in hopes of a revolutionary, CELL-based rasterizer. I don't think they wanted to give that up.

Name me, I would have loved to see a REYES based set-up with a pure CELL based chipset .

Would have it been cool ? Yes.

Is it the best thing for the PlayStation 3 taking timing, cost, etc... into consideration ? Probably not, hence the switch to the Toshiba solution and then to the nVIDIA solution.

I wanted to reply to put my perspective down on where I've been coming from over the past year or so of discussions but I also wanted a wider discussion so I've opened a new thread here.

 

[passerby]

Could someone translate the text in that diagram please. Thanks in advance!

The original article. http://pc.watch.impress.co.jp/docs/2004/1228/kaigai146.htm
Worth taking a look since it has a link to one Intel slide that sounds and looks very, very familiar.

The Intel presentation says how they are going 'many-core' looking at numbers up to 10-100. Should really click on that slide for the rest of the so-familiar 'key words'.

Rest of the article just briefly summarizes problems, solutions of going 'many-core' and TDP(Thermal Design Power). The slide version linked to is of course, a 'Cell example'. Note that Goto is just making 'educated speculations'.

I have trouble understanding what he is really trying to say there, since he refers to 'performance' and 'performance/TDP' seperately. I assume he means the ratio. In which case...


- single core: high performance, p/tdp low
- multicore, 'sophisticated cores' - cores more complex, high performance, p/tdp not too high.
- multicore, 'simpler cores' - high p/tdp, scalable, performance/efficiency suffers.
- 'customized' CPU core - many specialised multimedia CPU cores linked to a general purpose CPU. very high p/tdp if used properly.

I still think the Intel slide is more fun...

 

[V3]

It is not going over budget, as you yourself said they already left wiggle room in the budget.

If their solution can fit in that wiggle room, they would go with their own solution. But it didn't fit. That's why they go with NV, a cheaper solution that fit their budget.

If you think that nVIDIA is in to provide only Pixel Engines/ROPs then you are mistaken.

Look at NV40, there is a lot more stuff that makes a GPU than just Pixel Engines/ROPs.

Pixel Engines, is whatever it needs to shade pixels, so if it needs other stuff other than ROPs it will be there.

That is a fancy way of looking at things: it is not that nVIDIA's solution is superior, it is the most affordable.

Not exactly, superior and inferior is just speculation, but because Sony went with NV, NV solution is definitely meet their budget.

Plan A might have not passed budgeting (internal... Visualizer), but I am not aware that plan B (which jad been the official plan for quite a while) went overbudget as the contractors for plan B seemed to be a bit surprised at the coup that nVIDIA managed.

I think nVIDIA solution might have out-performed and out-featured plan B, but Toshiba had the lower cost (being Sony/SCE's partner, having already done work with XDR, Redwood and CELL they could afford to push the envelope without needing too much R&D money [less money than what Sony/SCE deemed as maximum R&D budget for the GPU]).

Being partner, doesn't always mean lower cost. NV undercut them. Like I said before, there are no motives to waste money on something better, performance or features wise. If they've contracted Toshiba to make something at their specs, it'll be good enough.

I also think that nVIDIA might have convinced Sony/SCE because of their change on the royaltes issues: it might have been what Sony/SCE needed (plus a nice big argument by nVIDIA about where the future of GPUs was headed and how the GPU they selected so far was nowhere near, according to nVIDIA of course, what Sony/SCE needed for PlayStation 3) to finally change their minds.

nVIDIA did not even seem a likely winner of the contract in the two weeks before the announcement was made, not at all.

That kind of argument won't win over Sony Pana. Like I said NV undercut Toshiba, to provide cheaper solutions, and won the contract. It might be better too, but that's speculation.

 

[j^aws]

I do not think PlayStation 3 will fully go the way the first patent showed: namely a wholly CELL based system with a CELL based Visualizer.

Are we only just coming to this realisation?

I'm sure you're aware of this but just to clarify, for the GPU to be Cell based, it just needs to run 'software Cells' and doesn't have to be the Visualizer from the patents or configured like a S|APU from the CELL CPU. They only need a consistent ISA to run 'software Cells'!

I give up .

Just trying to make my point not to discount a CELL based NV GPU under that definition too quickly.

...
The NV5X architecture was not developed with CELL in mind or using the Apulet ISA/runnign the same Apulets and required several years of work and hundreds of engineers to be completed: wouldn't you think that nVIDIA would need more than 50 engineers for a bit more time to convert all their Shading cores to support the execution of Apluets/Software cells ?
...

Sure...but to put things into perspective, NV were using HUNDREDS of engineers to design the core GPU architecture and ALONGSIDE them, over the past 2 YEARS, have used 50 NV engineers to customize the GPU for the PS3 and ALONGSIDE them, an undisclosed number of SONY engineers.

This is done without having the PC GPU adopt the x86 ISA.

Sure..but the x86 wasn't designed with the same CELL vision in mind.

Any project and the PS3 is no different has a specification, a delivery time and allocated budget/resources. Of course, these will all be subject to change througout the course of the project. And all projects have several contingency plans to meet the spec and delivery time's at critical MILESTONES. But just because one of the resources is now OFFICIALLY NV, doesn't mean any of the project specs and delivery timescales have been altered. Re-allocation of resources are always being juggled in projects to MEET the spec and delivery date. Which noone here know without being NDA'd.

Anyway, did you get a chance to read my link on the TAOS operating system?

http://www.beyond3d.com/forum/viewtopic.php?p=438551#438551

They get around the uniform ISA problem in a heterogeneous multi-processor environment by compiling to a 'virtual' processor with no overheads in translation due to a very efficient nano-kernel running on each processor. I know the CELL press releases mentions that the CELL processors can run multiple operating systems. If this means multiple nano-kernels or equivalent on each core, then they could be borrowing many ideas from TAOS for CELL.

 

[j^aws]

Could someone translate the text in that diagram please. Thanks in advance!

The original article. http://pc.watch.impress.co.jp/docs/2004/1228/kaigai146.htm
Worth taking a look since it has a link to one Intel slide that sounds and looks very, very familiar.

The Intel presentation says how they are going 'many-core' looking at numbers up to 10-100. Should really click on that slide for the rest of the so-familiar 'key words'.

Rest of the article just briefly summarizes problems, solutions of going 'many-core' and TDP(Thermal Design Power). The slide version linked to is of course, a 'Cell example'. Note that Goto is just making 'educated speculations'.

I have trouble understanding what he is really trying to say there, since he refers to 'performance' and 'performance/TDP' seperately. I assume he means the ratio. In which case...


- single core: high performance, p/tdp low
- multicore, 'sophisticated cores' - cores more complex, high performance, p/tdp not too high.
- multicore, 'simpler cores' - high p/tdp, scalable, performance/efficiency suffers.
- 'customized' CPU core - many specialised multimedia CPU cores linked to a general purpose CPU. very high p/tdp if used properly.

Thanks for translation!

I still think the Intel slide is more fun...

You mean this,

CELL inside, INTEL outside (TM)?

 

[Gubbi]

CELL inside, INTEL outside (TM)?

Intel won't make the mistake of putting a 1960s programming model on their system though (wrt. the memory model).

Cheers
Gubbi

 

[Panajev2001a]

If CELL APU's were considered by nVIDIA to be much better than their Vertex and Pixel Shader ALUs to spend a lot of effort and money replacing NV5X's ALUs with modified/enhanced APUs for a GPU contract they were not even sure of winning, it makes you think why NV5X is not CELL based .

Sure...but to put things into perspective, NV were using HUNDREDS of engineers to design the core GPU architecture and ALONGSIDE them, over the past 2 YEARS, have used 50 NV engineers to customize the GPU for the PS3 and ALONGSIDE them, an undisclosed number of SONY engineers. Wink

I do not think this is how it happened.

The GPU was not prepared in parallel with close collaboration of nVIDIA's engineers and SCE's engineers for that long (nVIDIA did follow closely PlayStation 3's development and started plannign for features and things that SCE might like in a console environment, but that is part of the "trying to win the contract" thing): the solution that was supposed to GO in PlayStation 3 officially was not nVIDIA's solution.

The custom GPU nVIDIA is providing to SCE started development when NV5X development was finishing down and nVIDIA was trying to win PlayStation 3's contract away from Toshiba.

I think that nVIDIA put their engineers to work on NV5X and if others were free during the ehavvy phase of R&D they were pout to work on NV4X revisions and early R&D for NV6X.

However what you talk about is not a simple customization of their architecture, something that would have required more engineers on their part.

There would the a good case for a lawsuit on the part of Toshiba if it was true that Sony was funding with engineers and resources (including CELL IPs) a customized NV5X while they had already told Toshiba that their GPU was going to be THE GPU used by PlayStation 3.

 

[j^aws]

If CELL APU's were considered by nVIDIA to be much better than their Vertex and Pixel Shader ALUs to spend a lot of effort and money replacing NV5X's ALUs with modified/enhanced APUs for a GPU contract they were not even sure of winning, it makes you think why NV5X is not CELL based .

I thought you gave up! I've already stated, I'm not bothered either way but I think I've made my point not to discount a CELL based GPU too quickly! Especially when we've come full circle with the above argument!

For PC, the NV5x is NOT going to be CELL based. For PS3, it make's sense for it to be CELL based as you've already agreed to.

So we're now debating the CELL ISA and how that could be developed for a GPU for PS3 to the required spec and with the available budget/resources/engineers?

I've already given examples of ways for this ISA to work. Whether the contract remained in-house, went to Toshiba, ATI, or who ever...

1. In your above example, how do you know it won't be easier to modify NV ALUs than it would be to modify S|APUs for the CELL ISA? Would modified S|APUs be better at PS work than modified NV ALUs? They've chosen NV so...

2. I've already mentioned they could have a CELL template like the Toshiba MeP template here...

Toshiba has made a flexible MeP architecture to have modified ISA's. Why wouldn't CELL have something like this considering it's use in a similar market? Why wouldn't NV's next gen NV5x have a similar type of template to modify ISA's considering they're going into the same market as mentioned by Huang? Why are other companies looking at making customized ISA's, e.g. Tensilica's Extensa?

3. What about compiling to a 'virtual' processor CELL ISA as mentioned earlier like TAOS? In the press releases, STI have already stated that CELL is an 'OPEN' platform? How about making that CELL ISA 'OPEN' and introduce CELL to a wider market? How about encouraging HW/SW innovations for the CELL ISA to make inroads into the x86 ISA?

These are just my ramblings but you should get my gist!

...
There would the a good case for a lawsuit on the part of Toshiba if it was true that Sony was funding with engineers and resources (including CELL IPs) a customized NV5X while they had already told Toshiba that their GPU was going to be THE GPU used by PlayStation 3.

This is what lawyers are there for when signing contracts and sub-contracts! And remember PS3 project != CELL project.

 

[Panajev2001a]

For PC, the NV5x is NOT going to be CELL based. For PS3, it make's sense for it to be CELL based as you've already agreed to.

I did not agree to that .

 

[Panajev2001a]

1. In your above example, how do you know it won't be easier to modify NV ALUs than it would be to modify S|APUs for the CELL ISA? Would modified S|APUs be better at PS work than modified NV ALUs? They've chosen NV so..

You knwo what is easier ?

Take the ALUs and the control logic and all the interfaces betwen Shading ALUs, TMUs, etc... that nVIDIA has developed and customize the architecture around those items .

There is more stuff to be done on the GPU than to modify it so that it becomes a CELL processor and share Apulets with the CELL based GPU.

Adding Shading ALUs and/or TMUs, adding e-DRAM, etc...

 

[j^aws]

For PC, the NV5x is NOT going to be CELL based. For PS3, it make's sense for it to be CELL based as you've already agreed to.

I did not agree to that .

Well thanks for confirming this to me...now is the perfect opportunity to say we shall 'agree to disagree (TM)'!

Until now, I've read your post as , yes it makes sense to have a CELL based GPU, yes that would be cool...etc, etc, etc..and where we are disagreeing is that delivery time, resources/budget to meet that spec wouldn't be possible for PS3 but will be for PS4/ CELL 2.0? Hence my above post for various resolutions on the CELL ISA for PS3...

And no, I'm not getting into a game of semantics here...

1. In your above example, how do you know it won't be easier to modify NV ALUs than it would be to modify S|APUs for the CELL ISA? Would modified S|APUs be better at PS work than modified NV ALUs? They've chosen NV so..

You knwo what is easier ?

Take the ALUs and the control logic and all the interfaces betwen Shading ALUs, TMUs, etc... that nVIDIA has developed and customize the architecture around those items .

There is more stuff to be done on the GPU than to modify it so that it becomes a CELL processor and share Apulets with the CELL based GPU.

Adding Shading ALUs and/or TMUs, adding e-DRAM, etc...

My preference for overall flexibility, simplicity and wider market penetraion is to make a CELL ISA on a 'virtual processor', a CELL VM, an Open standard. Like the Java VM but working closer to the metal for performance.

Then have various 'stream' programming languages, Cg, Brook etc. to compile to this CELL ISA, taking advantage of BOTH CPU+GPU being 'stream processors' to get a distributed 'software renderring' analogy. Making it open would let different stream languages develop in this CELL ISA environment Kinda like how C/C++ flourished under GNU but have the equivalent 'Stream-C/C++' flourish!

 

[Panajev2001a]

For PC, the NV5x is NOT going to be CELL based. For PS3, it make's sense for it to be CELL based as you've already agreed to.

I did not agree to that .

Well thanks for confirming this to me...now is the perfect opportunity to say we shall 'agree to disagree (TM)'!

Until now, I've read your post as , yes it makes sense to have a CELL based GPU, yes that would be cool...etc, etc, etc..and where we are disagreeing is that delivery time, resources/budget to meet that spec wouldn't be possible for PS3 but will be for PS4/ CELL 2.0? Hence my above post for various resolutions on the CELL ISA for PS3...

I want to be clear.

I said that I would have liked a CELL based GPU... eventually I'd like to make of somethign like the Visualizer the only IC in the chipset if it were possible, fusing graphics and general purpose processing... no need of a Broadband Engine and a Visualizer... just a sea of shading ALUs and PUs with a pool of fixed function units as co-processor basically.

This is not what we can get IMHO in PlayStation 3: it is something we can look forward for the more distant future.

NV5X is not CELL based and at this point it would not make sense (financially and engineering wise) for the PlayStation 3 GPU to be CELL based as it is based on the NV5X architecture.

 

[MrSingh]

ohhh.. the RS would have sucked so badly compared to the NV offering.

 

[j^aws]

For PC, the NV5x is NOT going to be CELL based. For PS3, it make's sense for it to be CELL based as you've already agreed to.

I did not agree to that .

Well thanks for confirming this to me...now is the perfect opportunity to say we shall 'agree to disagree (TM)'!

Until now, I've read your post as , yes it makes sense to have a CELL based GPU, yes that would be cool...etc, etc, etc..and where we are disagreeing is that delivery time, resources/budget to meet that spec wouldn't be possible for PS3 but will be for PS4/ CELL 2.0? Hence my above post for various resolutions on the CELL ISA for PS3...

I want to be clear.

I said that I would have liked a CELL based GPU... eventually I'd like to make of somethign like the Visualizer the only IC in the chipset if it were possible, fusing graphics and general purpose processing... no need of a Broadband Engine and a Visualizer... just a sea of shading ALUs and PUs with a pool of fixed function units as co-processor basically.

This is not what we can get IMHO in PlayStation 3: it is something we can look forward for the more distant future.

NV5X is not CELL based and at this point it would not make sense (financially and engineering wise) for the PlayStation 3 GPU to be CELL based as it is based on the NV5X architecture.

Crytal clear...I'll just conclude with why, with all things said, the following still sounds contradictory for PS3, IMHO.

http://www.beyond3d.com/forum/viewtopic.php?p=437597#437597

 

[Panajev2001a]

Your argument relies on this, basically:

Premise 1: If Hofstee was talking about two-way comms between CPU<=>GPU,

Premise 2 (a conclusion based on the validity of premise 1): then these VS CELL threads should run on the GPU also, no?

---------

Conclusion: This should imply the GPU is CELL based.

This is what you are using to say my argument is contraddictory, isn't it ?

I call it a non sequitur on the passage from P1 to P2 and especially from P1+P2 to the Conclusion.

Your argument might be valid, but it does not seem sound to me as your premises do not appear more known than the conclusion you are trying to proove.

 

[j^aws]

Your argument relies on this, basically:

Premise 1: If Hofstee was talking about two-way comms between CPU<=>GPU,

Premise 2 (a conclusion based on the validity of premise 1): then these VS CELL threads should run on the GPU also, no?

---------

Conclusion: This should imply the GPU is CELL based.

This is what you are using to say my argument is contraddictory, isn't it ?

I call it a non sequitur on the passage from P1 to P2 and especially from P1+P2 to the Conclusion.

Your argument might be valid, but it does not seem sound to me as your premises do not appear more known than the conclusion you are trying to proove.

http://www.beyond3d.com/forum/viewtopic.php?t=19108

Hofstee above...

And conversely, from my point 3, if CPU=>GPU, would imply GPU is NOT CELL based...but hey, it's been an interesting discussion either way.

 

[Panajev2001a]

I do not see what Hofstee is saying is contraddicting what I am saying.

Whether CELL based or not, the GPU in PlayStation 3 will be programmable enough and with enough bandwidth between itself and the CPU to effectively act as a media co-processor: similarly to how it will work on Xbox 2.

The era of 1-way communication between CPUs and GPUs is ending IMHO.

Hofstee talked about things that were possible in the short term and then long term goals: he said that we would go away from textures (he wants an all procedural system [textures are still used in the most advanced off-line CG production today and the trend seems to continue]), do you believe that PlayStation 3 will go to a kind of rendering system that does not leverage textures (even if you limit this to saying "no bitmaps to fake detail" which would apply to normal mapping, detail texturing, displacement mapping , etc...) ?

You have a theory, you have certain premises: I ask you to expand on those premises .

3. However, if the VS is NOT done on the CELL CPU but on the GPU, then the entire GPU may OR may not be CELL based then, no?

I am not saying VS will not be done at all on the APUs, even if the GPU had VS units.

I am disputing your premise that says "if Vertex Shading is done on the CPU then the GPU uses CELL technology".

Your argument relies on this, basically:

Quote:
Premise 1: If Hofstee was talking about two-way comms between CPU<=>GPU,

Premise 2 (a conclusion based on the validity of premise 1): then these VS CELL threads should run on the GPU also, no?

---------

Conclusion: This should imply the GPU is CELL based.


This is what you are using to say my argument is contraddictory, isn't it ?

I call it a non sequitur on the passage from P1 to P2 and especially from P1+P2 to the Conclusion.

Your argument might be valid, but it does not seem sound to me as your premises do not appear more known than the conclusion you are trying to proove.