IBM power7 in PS4 ?

[green.pixel]

while PC's with mightier GPUs have far less flexibility, which devs don't care about as they don't want that much flexibility at the moment!

Far less flexibility, what? I think you meant far more . I don't see how being OpenCL/CUDA/DX Compute-ready with hundreds of unified stream processors and soon with CPP and coherent memory arch. is less flexibilty than what devs have on consoles.

And don't you think that devs would want that kind of flexibility on consoles? They don't care at the moment solely because their development is consoles-centric and on consoles they simply don't have powerful PC tech, they would care if they did have; also, they need their games to sell well, and unfortunately for all of us, that's not possible atm on the PC for a variety of reasons, some reasonable and some somewhat obscure. As Timothy Farrar said on his twitter feed "PC devs+players miss point and blow perf+bandwidth on resolution using nearly same engine as console". All those TFLOPS and transistors are IMO, almost being useless with that kind of business model.

In an ideal world, we would have just one "big" game system with the most powerful hardware available, being refreshed at a 2-3 years rate as technology advances and faster and more flexible tech arrives, ALL the devs in the world would develop and had much less hassle with just one platform, they would have been given the bestest tools possible to exploit all that silicon underneath. There's "just" couple of problems. First, no matter what company, all they want is profit for themselves, not the "universal benefit" and "good" for an end user. Second, Earth's economy can't work like that and would collapse in a second.

 

[corduroygt]

I'd say an 8-12 core PPC based chip, where each core can run SPU code as fast as an SPU, meaning they have 256kb of LS, which could also be used as registers (that's a lot of registers LOL), and a unified shared cache of a few megs, whatever fits the silicon budget. Also each core would have PPE instructions, branch prediction, etc. This might indeed be IBM's replacement for cell, or it may not even be possible or efficient to achieve SPU performance once you add the extra logic to it. At least this way, developers could just compile it and even if they don't use the exotic SPU architecture, they will have more than 1 core to run their programs.

 

[archangelmorph]

That's what I meant, when I said dev need relearn ( not that dev forgot ) how to put those SPUs to use with new code. I mean you can give those SPUs any tasks, but considering the cost of larger number of SPUs is transistors from GPU, can you give those SPUs jobs that'll give better visual improvement compare to just using better GPUs.

I don't see why you think CPU transistor budget will affect GPU transistor budget in such a significant way.. It's not as simple as saying "lets put in less SPUs so we have more silicon for the GPU"..
Also there's alot you can do with an array uber-fast general purpose stream processors (including weird and wonderful systems and algorithms we haven't even thought of yet) to make both your game look better and also create a richer world "simulation" which I think will have a much greater impact on the quality of games going into the next gen..

Can't the 8 SPUs in the current form of Cells be enough to handle those jobs instead of going, say to 32 SPUs Cell ?

We can always find more work for the hardware to do.. The problem is getting the hardware to do the work faster and smarter. So as far as the question of whether 8 SPUs is enough? no chance.. What about 32?.. 64..? 512..?
As far as there's money to burn there's always work that can be done to improve the quality of a product regardless of how powerful (or not) the hardware maybe..

With better expected GPU do you have jobs for those 32 SPUs that gamers can recognize and enjoy ? Do you see any clear benefits that you can do with 32 SPUs Cell that no GPUs in the world can't do for games ?

That's not the question you should be asking.. The question you should be asking is are there going to be new solutions to problems that will exist in the future (that will provided added value to the quality of the product, whether through the visuals or the complexity of the world simulation..) that you can handle efficiently on the CPU without having to push all your data across the bus to the GPU..? The answer more than not depends entirely on the rest of the hardware configuration as a whole & not just in whether you have a "stronger" CPU or GPU..

Or are you still wait and see like how dev approached PS3. The current Cell was like a solution looking for problems, instead of a solution to problems, which GPUs are.

I'd agree with you if it wasn't for the fact that I don't .. There were and still are plenty of problems the Cell was designed to handle efficiently and its an interesting point to note that these problems generally make up the wider load of games-related processing.. I don't see how a chip that is considerably well optimized for things like vertex processing, skinning, culling, collision, physics, encoding/decoding audio/video streams etc.. is not a solution to a set of problems that have existed in games from the beginning.. Also it's worth noting that Cell is much more future-proof as far as being much more adaptable to new types of work loads than having to somehow wangle all your data into a shader-friendly format & fit it into your GPU registers..

Overall though my point remains that a new Cell wouldn't require any new "learning" of how to apply processes to work with the hardware as most of the practical/theory in this area has been done this gen. It'll only be a case of scaling existing tasks to take advantage of more raw power (i.e. spreading your physics, skinning & vert processing across more cores for example..) & making use of the hardware to implement new forms of processes that can further enrich the quality of your product (in the ways I mentioned earlier)..

 

[Shifty Geezer]

IWe can always find more work for the hardware to do.. The problem is getting the hardware to do the work faster and smarter. So as far as the question of whether 8 SPUs is enough? no chance.. What about 32?.. 64..? 512..?
As far as there's money to burn there's always work that can be done to improve the quality of a product regardless of how powerful (or not) the hardware maybe..

Herein lies the rub, and from that context V3's point makes a lot more sense. Given the finite budgets of next-gen, will the cost of getting 8 SPU's fully loaded and a monster GPU mashing pixels be that limit or will we have cash to spare that, if we had more processing cycles to play with, we could have produced a more impressive engine? PS2 had more processing power than most devs cared to really tap, which goes to show that just because it's there, doesn't mean it'll get used.

That said, I supposed ideally with enough processing power development becomes much cheaper. You can use higher-level tools and code, and turn offline production tasks to realtime tasks. eg. If we had enough system performance, we wouldn't need the expense of precomputing lightmaps as we could render GI in realtime!

Given the way games are being made at the moment though, I expect if the GPU was very fast and capable, anything beyond 8 SPU's worth of CPU processing will go unused. Sharing that processing over 16 cores may make things easier for devs though who won't have to optimise so strongly.

 

[liolio]

Sony were better off putting better GPU in PS3 while going for smaller CPU. PS3 had some large silicon real estate ~ 500 mm^2. Even if the trade off between CPU and GPU are quite small this gen, next gen, 32 nm Cell will be around or less than 70 mm^2. if Sony stuck with Cell, they can put 400 - 500 mm^2 into GPUs real-estate.

Highly unlikely due to thermal dissipation /power consumption issues and yields such a chip would have.

 

[Dregun]

I think when comparing the 360 vs the PS3 we can see how much impact a flexible and powerful CPU can have on graphics. The Cell thrives on assisting the GPU, it always will; they could throw in an uber fast GPU and keep the same Cell and it would still be able to provide assistance for the GPU.

Think of it this way, if the PS4 came out with just a more powerful version of the Cell could we expect an increase in graphics? If the PS4 just came out with a more powerful GPU could we expect an increase in graphics? While this is a very general comment I think it shows that because of the Cell and its ability to help the GPU no matter what they beef up we will see significant gains out of it. While the processor in the 360 is a powerful one it cannot do as much in assisting the GPU because of its design, therefore a Xbox4 with just a stronger CPU would do very little in increasing the graphic fidelity unless they went with a different CPU architecture.

PC = Console does not work! Trying to compare what equipment PC users buy or how PC devs develop games is like comparing people who buy airplanes and those who work on assembly lines for automobiles. In the PC world the CPU does very little to assist in graphics so of course a faster/stronger GPU would be where the majority of gamers would spend their budget. If however the PC had a very programmable and flexible CPU that assisted in the rendering of graphics I think we would see a very different buying habit from the PC gamer market.

In fact it wasn't until recently that the gains from having a stronger CPU in a PC started to have less and less impact on the overall performance of gaming. PC = Console is only good for comparing the performance of individual aspects of technology; not the systems as a whole.

I would like to see a Cell that is a minimum of 2x more powerful then it is now and a nice beefy GPU. On paper who would have thought RSX could output games such as Killzone/Uncharted 2; on paper it couldn't but with the assistance of Cell it did.

 

[Carl B]

In response to the above:

http://forum.beyond3d.com/showpost.php?p=1374707&postcount=3208

 

[V3]

I'm not so sure about that. PS3 seems to follow some ideals of PS2, with a very powerful, flexible CPU performing some fo the graphics tasks. the graphics problems exist and need processing to solve them, and Sony chose to deal with lots of that processing on the CPU rather than on the GPU. Taken as a whole system, the programmability of PS3's graphics is unequaled this gen. The problem is it needs developers to make use of it. In contrast, XB360's system hit a good flexibility with far less hassle, while PC's with mightier GPUs have far less flexibility, which devs don't care about as they don't want that much flexibility at the moment! They just want to render pretty graphics, and using the conventional methods makes that possible and easy on their hardware of choice.

PS3 isn't a bad design per se; just, relative to the approach of other systems, it's not a good fit for the current market.

Remember, they disable one of the SPU in Cell {what a waste of real estate that was). Also Sony were able to reserve one (and a half ?) SPU. So even Sony thought Cell was too much for PS3. It's not too suprising since RSX has vertex shaders. If this was PS2 like design Cell would have handle the vertex job and help with pixel job.

Anyway programmability is all well and good, but it was too costly for PS3. And all those programmability that they put into PS3, it still ended breaking the Playstation family compatability, that they work so hard and waste a lot of money to maintain in the few early PS3 designs.

 

[jonabbey]

Remember, they disable one of the SPU in Cell {what a waste of real estate that was). Also Sony were able to reserve one (and a half ?) SPU. So even Sony thought Cell was too much for PS3. It's not too suprising since RSX has vertex shaders. If this was PS2 like design Cell would have handle the vertex job and help with pixel job.

Raising the yield of Cell by allowing them not to toss out CPUs with a bad SPU on board is not a waste at all. GPU manufacturers all do the same thing.

 

[V3]

I don't see why you think CPU transistor budget will affect GPU transistor budget in such a significant way.. It's not as simple as saying "lets put in less SPUs so we have more silicon for the GPU"..

Next gen, transistors budget will be far more significant than current gen. This gen we have around 500m budget, next gen it'll be somewhere inbetween 4-8 b transistors. Any inefficiency or flaw in the design will make the system handicap be more obvious.

Also there's alot you can do with an array uber-fast general purpose stream processors (including weird and wonderful systems and algorithms we haven't even thought of yet) to make both your game look better and also create a richer world "simulation" which I think will have a much greater impact on the quality of games going into the next gen..

Here is the problem, it is good to design a system to handle some new algorithms that may arise in the future, but not at the expense of what algorithms that are already available. Beside research into GPGPU is going ahead so you better be sure that there are something out there that will make putting 32 SPUs be worth it, if you're to go ahead with it.

We can always find more work for the hardware to do.. The problem is getting the hardware to do the work faster and smarter. So as far as the question of whether 8 SPUs is enough? no chance.. What about 32?.. 64..? 512..?
As far as there's money to burn there's always work that can be done to improve the quality of a product regardless of how powerful (or not) the hardware maybe..

If you're going to view this from unlimited resources sort of design you're not going to see my argument. You can always give Cell more jobs to do, but are they useful jobs ? Jobs that GPU can't do ? Because Cell is more expensive compare to GPU in term of transistors budget.

That's not the question you should be asking.. The question you should be asking is are there going to be new solutions to problems that will exist in the future (that will provided added value to the quality of the product, whether through the visuals or the complexity of the world simulation..) that you can handle efficiently on the CPU without having to push all your data across the bus to the GPU..? The answer more than not depends entirely on the rest of the hardware configuration as a whole & not just in whether you have a "stronger" CPU or GPU..

I think this is the wrong way to go about it. This is the recipe to end up with overengineered product.

I would make sure the system runs the solutions that are available and run it upto standard. As for new solutions that may rise in the future, if it worked on the configuration than use it, if not it will have to wait for another gen of consoles.

Beside those new solutions tend to be a way to do jobs more efficiently on the current hardware configuration, so what ever your current hardware configuration is, there will be new solutions anyway. But it is poor hardware design to make dev think of new solution so your product can catch up to competitor that had released their hardware a year earlier.

I'd agree with you if it wasn't for the fact that I don't .. There were and still are plenty of problems the Cell was designed to handle efficiently and its an interesting point to note that these problems generally make up the wider load of games-related processing.. I don't see how a chip that is considerably well optimized for things like vertex processing, skinning, culling, collision, physics, encoding/decoding audio/video streams etc.. is not a solution to a set of problems that have existed in games from the beginning.. Also it's worth noting that Cell is much more future-proof as far as being much more adaptable to new types of work loads than having to somehow wangle all your data into a shader-friendly format & fit it into your GPU registers..

There is optimized then there is overkill. GPU or cheap dedicated hardware already cover some jobs you mentioned. For the rest Xbox 360 managed to do them on smaller CPU. Games like GTA4 where there are large scale world simulation going on everytime, just failed to perform any better on PS3. Because at the end of the day, Playstation runs games, not physics. Even if Cell could run physics simulation faster than its rival, it is overkill when it still runs the game as choppy as its rival. Future proofing is nice, but future proofing is the same as looking for problems. Which I said what Cell is doing, so in a way you seems to agree with me at the end

Overall though my point remains that a new Cell wouldn't require any new "learning" of how to apply processes to work with the hardware as most of the practical/theory in this area has been done this gen. It'll only be a case of scaling existing tasks to take advantage of more raw power (i.e. spreading your physics, skinning & vert processing across more cores for example..) & making use of the hardware to implement new forms of processes that can further enrich the quality of your product (in the ways I mentioned earlier)..

I still don't see any jobs that future GPU can't handle to a satisfactory level and you keep saying new forms of processes which imply "learning" new algorithms, finding out what works and what doesn't. Given how good the GPU performance to cost ratio compare to other solutions that pave the way into GPGPU research, wouldn't better GPU enriches games more compare to 32 SPUs Cell ?

 

[V3]

Highly unlikely due to thermal dissipation /power consumption issues and yields such a chip would have.

I wasn't talking about a single 400 mm^2 GPU, For PS4, if it is like PS2 and PS3, it should have two 200 mm^2 chips, I am saying this should be GPUs. 32 nm Cell should be smaller than EE+GS chip that was in the early PS3 iterations. They can integrate it with the rest of the southbridge. Cell should have enough FLEX I/O lane for it. Power consumption will increase either way.

It should be easier to combine the two same GPUs into a single chips compare to combining 32 SPU Cells and a GPU into a single chip as process improves.

 

[V3]

Raising the yield of Cell by allowing them not to toss out CPUs with a bad SPU on board is not a waste at all. GPU manufacturers all do the same thing.

They should have make 4 or 6 SPUs Cell and yield will be good all around. They could make up some of the performance difference with clock speed instead. If you need to disable a single core, not just bit of parts of the chip, eg cache, etc, the design is just too big for the application. Sony paid for this if you see their bottom line. So with new management I am not too sure they're eager to do it all over again.

 

[turtle]

They should have make 4 or 6 SPUs Cell and yield will be good all around. They could make up some of the performance difference with clock speed instead. If you need to disable a single core, not just bit of parts of the chip, eg cache, etc, the design is just too big for the application. Sony paid for this if you see their bottom line. So with new management I am not too sure they're eager to do it all over again.

So in that vein, is Xenos a failure in your eyes? It has (had?) 64 shader pipelines, with 16 disabled for yield. It was 186mm2. Cell was something like 220mm2.

Or, to put it another way, if they made these chips, what would they do with the chips that only had 3 or 5 functioning SPEs? What about the ones that couldn't reach 4ghz clockspeed on all units? Remember Cell has limited uptake, and is mainly (not only, but mainly) used in the PS3.

Expecting to use the whole of all chips on a wafer at perfect spec, especially a specialized chip and/or one over 200mm2, is fool-hearty (or you're nVIDIA). Chips towards the edge of a wafer will run at lower clock speeds, and defects will occur on any number of chips on a wafer of silicon. While the affected chips are hardly useless in most cases, what one does is either bin them as CPU/GPU manufacturers do to put them in specific products for chips that can meet a certain minimum spec, or in this case find the highest common denominator for the bulk of chips while still maintaining a certain performance standard. This is what they did with Cell. On it's earliest process and original die size, they found this (7 SPE, 3.2ghz) to be the best cost/performance yield per wafer compared to the perfect spec of 8 SPE at 4ghz in accordance to it's pipeline. Hardly an unreasonable compromise.

 

[3dcgi]

Xenos has 48 enabled shader pipelines and 3 redundant ones.

 

[sunscar]

Stray thought...

We could keep the current CELL with few changes, and allow it to play traffic cop for two GPUs - One, a DX11-level part, something cheap, maybe as powerful as a Radeon 5670, or a low-end or mobile FERMI, dedicated to specifically to physics processing (but could do more) and the other GPU could be something a little beefier dedicated specifically to graphics processing (but again could do more). It's feasible you could retain backwards compatibility this way if you view it as a necessity, but regardless that's quite a lot of power.

 

[V3]

So in that vein, is Xenos a failure in your eyes? It has (had?) 64 shader pipelines, with 16 disabled for yield. It was 186mm2. Cell was something like 220mm2.

I think that was the size of prototype Cell before some fixes, PS3 Cell was 235 mm2. Xenos failure in that sense pretty much speak for itself with how much it cost MS. Sony and MS chips were just too big for $300 consoles (Sony had to up theirs to $500 and people weren't too please).

Or, to put it another way, if they made these chips, what would they do with the chips that only had 3 or 5 functioning SPEs? What about the ones that couldn't reach 4ghz clockspeed on all units? Remember Cell has limited uptake, and is mainly (not only, but mainly) used in the PS3.

You throw them away, if you can't bin them for other uses. That's why they're so expensive.

Expecting to use the whole of all chips on a wafer at perfect spec, especially a specialized chip and/or one over 200mm2, is fool-hearty (or you're nVIDIA). Chips towards the edge of a wafer will run at lower clock speeds, and defects will occur on any number of chips on a wafer of silicon. While the affected chips are hardly useless in most cases, what one does is either bin them as CPU/GPU manufacturers do to put them in specific products for chips that can meet a certain minimum spec, or in this case find the highest common denominator for the bulk of chips while still maintaining a certain performance standard. This is what they did with Cell. On it's earliest process and original die size, they found this (7 SPE, 3.2ghz) to be the best cost/performance yield per wafer compared to the perfect spec of 8 SPE at 4ghz in accordance to it's pipeline. Hardly an unreasonable compromise.

Compromise ? Have you forgotten on how much money they lost per PS3 ? Sony shouldn't have to make cost/performance compromise after the design. They should have gone with 4 SPEs or whatever size chip that isn't going to bankrupt them from the start. They had the same problem with PS2 too, large chips. Though I don't remember them disabling any of the main part of the EE for yield, but there were shortage for PS2 too.

That's why for PS4, they should stick with the current Cell, higher clock with all SPEs intact.

 

[Carl B]

V3 you would be wrong to peg the $500 price tag on Cell alone though - there was a much more pertinent (and expensive) component included in the system to drive that price, lest we forget.

 

[ShadowRunner]

Compromise ? Have you forgotten on how much money they lost per PS3 ? Sony shouldn't have to make cost/performance compromise after the design. They should have gone with 4 SPEs or whatever size chip that isn't going to bankrupt them from the start. They had the same problem with PS2 too, large chips. Though I don't remember them disabling any of the main part of the EE for yield, but there were shortage for PS2 too.

That's why for PS4, they should stick with the current Cell, higher clock with all SPEs intact.

You are suggesting that the 3-4 additional spus in Cell are the scource of Sonys financial issues with regards to PS3?

There is a whole lot of lunacy going on around here at the minute....