Practicality of a CPU/GPU hybrid in consoles

[Fox5]

Assuming MS and SONY continue along the same path that they're going, their consoles will still consist of a fairly large but separate CPU and GPU. I think for backwards compatibility sake they'll stay with the same architectural design and just increase the number of PPEs, SPEs, shader units, etc.

Nintendo may take the single chip CPU+GPU model and stay within a certain die size to keep costs down since they likely will not be aiming for the most powerful console. Also Nintendo has already said they would like to continue to keep the size of the console small so smaller chips would help in that regard.

I think there's a high possibility Microsoft (And possibly Nintendo) have required that all code by heavily abstracted from the hardware, so they're free to make whatever hardware changes they want later. I wouldn't be surprised to see the next Xbox emulate 360 games, and with far better success than the 360 emulated xbox.

The wii will likely be reduced to a single chip system (for its major chips) by the end of its life if it continues its current rate of success. The n64 was made into a system on a chip (iQue) within 6 years of its launch I believe, and the Wii is based on Gamecube hardware and already looks reasonable to be made into a SOC with current tech.

 

[Capeta]

Yeah I could see the Wii chipset being a SiP or SoC when they move to 65nm or smaller. Even at 90nm you could already see that the CPU would almost fit onto the same package as Hollywood and the eDRAM. In fact I think that's one of the main reasons why the Wii chipset was designed the way it was.

 

[Techno+]

hybrid CPU-GPU on die combinations are the way to go. At 2012 we will be at 22 nm, (intel said 22 nm 2011, provided the rest are one year late in manufacturing process designs, most companies will be at 22nm), i.e we will have 16 core CPUs. I can imagine 4-6 normal CPU cores and the rest being used as GPU cores ( or ppu+gpu whatever is best). Just give me one good reason why we will need separate CPUs and GPUs for future consoles. The GPU will end up like the FPU, integrated into the CPU.

 

[Capeta]

Intel also said the P4 would reach 4GHz.

 

[TheChefO]

hybrid CPU-GPU on die combinations are the way to go. At 2012 we will be at 22 nm, (intel said 22 nm 2011, provided the rest are one year late in manufacturing process designs, most companies will be at 22nm), i.e we will have 16 core CPUs. I can imagine 4-6 normal CPU cores and the rest being used as GPU cores ( or ppu+gpu whatever is best). Just give me one good reason why we will need separate CPUs and GPUs for future consoles. The GPU will end up like the FPU, integrated into the CPU.

I'd put the likelihood of next gen console launching with 22nm tech in 2012 about equal to ps3's chances of repeating a 70million+ lead on their closest competitor this gen.

not - gonna - happen

I think we will be lucky to get 32nm in 2011 for next gen consoles.

2007 - 65nm
2009 - 45nm
2011 - 32nm*

*assuming everything goes smoothly. I have my doubts as these shrinks are approaching the physical limits of affordable material to reliably switch electrical current without bleeding into neighboring traces/transistors.

We'll see.

 

[nonamer]

hybrid CPU-GPU on die combinations are the way to go. At 2012 we will be at 22 nm, (intel said 22 nm 2011, provided the rest are one year late in manufacturing process designs, most companies will be at 22nm), i.e we will have 16 core CPUs. I can imagine 4-6 normal CPU cores and the rest being used as GPU cores ( or ppu+gpu whatever is best). Just give me one good reason why we will need separate CPUs and GPUs for future consoles. The GPU will end up like the FPU, integrated into the CPU.

22nm is the crossing over point for existing DUV (deep ultraviolet) to either EUV (extreme ultraviolet) or nanoimprint lithographic techniques, barring another near miraculous extension of DUV lithography. Seeing how EUV is problematic as hell right now and nanoimprinting lithography is extremely immature, 22nm could very well be the process node where Moore's Law dies. Last minute salvation could still happen though, and may even be expected, but don't expect people to push towards the potentially very nastly and hellishly expensive 22nm node by 2012 for something as low end as consoles.

 

[tmp]

(complete dual thread instead of just dual issue? I never understood what was going on there)

Freescale promises 760 MIPS for dual issue PPC core at 400 MHz. This is somewhat optimistic, of course. Dual threading, instead of dual issue, would give 2 x 400 MIPS. Not much better efficiency, and you'd have to separate your workload into two threads.



I'm predicting increased integration, although more with eDRAM than CPU/GPU because of two words. Thermal Envelope. Very fast buses are hard to design, hard to route, tend to require more expensive PCBs and can use significant amounts of power. Which means larger slabs of aluminum/copper and noisier fans. And devices that have tendency fail spectacularly, when placed into places where airflow is obstructed.

Floating body effect based DRAMs seem promising, since tradionally to use eDRAM, you needed to choose eDRAM process which was more expensive and had poorer logic performance than bulk CMOS. ZRAM type memory requires high performance SOI, which is not much of a sacrifice since performance difference seems to be about on par with cost difference. And nice, carefully designed eDRAM allows ripping some logic out of caches, memory controllers and latency hiding. Saving both the die area and power that was used for these functions.

So CELL(local storage) type solution with advanced eDRAM(for GPU also) combined with reasonable sane(Wii) thermal envelope seems to be the way to go.



It seems to me that in this generation MS solved the cooling problem by being the first to market(no competition, so it didn't matter all that much. Respin NOW.), Sony by making an impressive and expensive piece of consumer electronics(I dig Bluray, btw, because it spins slower than DVD), and Nintendo by not being all that powerful.

 

[nonamer]

I'd put the likelihood of next gen console launching with 22nm tech in 2012 about equal to ps3's chances of repeating a 70million+ lead on their closest competitor this gen.

not - gonna - happen

I think we will be lucky to get 32nm in 2011 for next gen consoles.

2007 - 65nm
2009 - 45nm
2011 - 32nm*

*assuming everything goes smoothly. I have my doubts as these shrinks are approaching the physical limits of affordable material to reliably switch electrical current without bleeding into neighboring traces/transistors.

We'll see.

While I can't comment on your belief of the PS3's success, your claims on the 32nm node is wrong, as of a few days ago. High-k dielectrics, as described here, will make 32nm an easily achievable reality, given you have to balls or money to make a chip by the 2011-2012 timeframe. 22nm has other problems and I too doubt we'll see any consoles with that at launch.

 

[TheChefO]

While I can't comment on your belief of the PS3's success, your claims on the 32nm node is wrong, as of a few days ago. High-k dielectrics, as described here, will make 32nm an easily achievable reality, given you have to balls or money to make a chip by the 2011-2012 timeframe. 22nm has other problems and I too doubt we'll see any consoles with that at launch.

I appreciate the insight but how does that make my statement wrong?

I stated it would be difficult (I have my doubts, lucky) to get 32nm by 2011 for consoles - not that it would be impossible.

I'm sure intel will have fabs pumping 32nm by this time frame but they are always ahead of Chartered, TSMC, IBM, AMD, etc on successful process shrinks.

 

[nonamer]

I appreciate the insight but how does that make my statement wrong?

I stated it would be difficult (I have my doubts, lucky) to get 32nm by 2011 for consoles - not that it would be impossible.

I'm sure intel will have fabs pumping 32nm by this time frame but they are always ahead of Chartered, TSMC, IBM, AMD, etc on successful process shrinks.

The problem has been solved now, or at least by the end of this year. It's hard to believe that by 2011-2012, it still won't be solved for the rest of the industry.

 

[Acert93]

2007 - 65nm
2009 - 45nm
2011 - 32nm

It totally depends what partners MS and Sony have.

Intel had production 65nm in 2005, with 65nm production eclipsing 90nm in Q3 2006. And right around the corner at the end of 2007 Intel will have 45nm production, with high K dielectics, on the market. Intel has also been bullish on 32nm in 2009 (the timeframe Intel has for Keifer) as well as the possibility of "3D" (tri-gate) transistors. The roadmaps indicate 22nm in 2011 from Intel. AMD had 65nm at the end of 2006 and is is talking about production 45nm in the middle of 2008 and they are working with IBM for 32nm and 22nm technologies through 2011. TSMC does lag, but will be turning out 65nm in the first half of this year for GPUs; ditto Charter on 65nm in the first half of 2007.

At worse I would expect 32nm in 2011. True, as you get down to 16nm and beyond it looks like things will become much, much trickier. But with 450mm wafers coming down the pipe (for Intel) and new tricks and techniques (like tri-gate transistors, ZRAM, etc) and the full bore ahead push for multi-core architectures there is a lot of potentual in the next 5 years.

Edit: See nonamer brought up the same point I did With Intel with production 32nm in 2009, the real point should be: If others cannot get 32nm by 2011 do you really expect MS to partner with someone else, especially with the potential power reductions Intel has planned between now and then? It makes sense that MS and others target 32nm or smaller, as 45nm is only 2 steps below what they have now. With certain areas seeing increased diminishing returns aiming for a process node where you can alleviate this some would be a better goal than making less of a jump traditionally aimed for, only to exacerbate the situation even more. Better to wait on the technology that can achieve your vision than to just churn out a new product that offers little compelling reason to jump on. And this is one of the problems with suggestions for the 2009 and 2010 timeframe, in that it isn't only chips that have to reach this point, but with the shift to digital distribution there needs to be a storage medium that is silicon to fit into the typical console model (or at least hopes of one within the first 2 or so years so they can begin cutting prices). I saw that SanDisk was offering a 32GB Solid State HDD... for $600. With digital distribution being an option (probably not only) means to distribute content on the next consoles I am guess you will want to aim for 500GB for HD oriented material. Of course Sega fans like their console launches fast and furious.

 

[TheChefO]

It totally depends what partners MS and Sony have.

Intel had production 65nm in 2005, with 65nm production eclipsing 90nm in Q3 2006. And right around the corner at the end of 2007 Intel will have 45nm production, with high K dielectics, on the market. Intel has also been bullish on 32nm in 2009 (the timeframe Intel has for Keifer) as well as the possibility of "3D" (tri-gate) transistors. The roadmaps indicate 22nm in 2011 from Intel. AMD had 65nm at the end of 2006 and is is talking about production 45nm in the middle of 2008 and they are working with IBM for 32nm and 22nm technologies through 2011. TSMC does lag, but will be turning out 65nm in the first half of this year for GPUs; ditto Charter on 65nm in the first half of 2007.

At worse I would expect 32nm in 2011. True, as you get down to 16nm and beyond it looks like things will become much, much trickier. But with 450mm wafers coming down the pipe (for Intel) and new tricks and techniques (like tri-gate transistors, ZRAM, etc) and the full bore ahead push for multi-core architectures there is a lot of potentual in the next 5 years.

Edit: See nonamer brought up the same point I did With Intel with production 32nm in 2009, the real point should be: If others cannot get 32nm by 2011 do you really expect MS to partner with someone else, especially with the potential power reductions Intel has planned between now and then? It makes sense that MS and others target 32nm or smaller, as 45nm is only 2 steps below what they have now. With certain areas seeing increased diminishing returns aiming for a process node where you can alleviate this some would be a better goal than making less of a jump traditionally aimed for, only to exacerbate the situation even more. Better to wait on the technology that can achieve your vision than to just churn out a new product that offers little compelling reason to jump on. And this is one of the problems with suggestions for the 2009 and 2010 timeframe, in that it isn't only chips that have to reach this point, but with the shift to digital distribution there needs to be a storage medium that is silicon to fit into the typical console model (or at least hopes of one within the first 2 or so years so they can begin cutting prices). I saw that SanDisk was offering a 32GB Solid State HDD... for $600. With digital distribution being an option (probably not only) means to distribute content on the next consoles I am guess you will want to aim for 500GB for HD oriented material. Of course Sega fans like their console launches fast and furious.

Interesting - So if Intel is able to mass produce 32nm by 2009, then 2010 should allow for TSMC, AMD, Chartered, etc to be on the same process. Just in time.

 

[Acert93]

Interesting - So if Intel is able to mass produce 32nm by 2009, then 2010 should allow for TSMC, AMD, Chartered, etc to be on the same process. Just in time.

It doesn't work like that. And even if it did, if Intel hit 32nm in Q4 2009 who is to say the smaller ones would be ready in Q4 2010 (instead of Q1/2 2011). Not to mention some of the big changes on the horizon. Besides wiggle room (maybe everyone slides 12 months), there is always the possibility that GPUs will continue being "late commers" to new process nodes. When you don't control your own fabs you are very much at the mercy of others. One nice thing about hitting a new process node after it has proven stability is process maturity (better yields and performance) as well it means the next node could be on the horizon within 12-18 months, allowing for strategic price drops early in the lifecycle.

IMO, whoever MS partners with they won't only need to consider the above issues (along with architecture) but they also will need to be mindful of the companies roadmap beyond 32nm. Currently shopping the IP to TSMC, UMC, Charter, etc offers MS with some good options, but in 7 or 8 years how will these companies adapt to the new hurdles in process technologies? In this regards Sony is probably in a better position as they have partners in IBM, Toshiba, NEC, and AMD as well as have their own fabs to help dictate their own production schedules. Not necessarily cheaper, but having more control over your future can be vital.

 

[TheChefO]

It doesn't work like that. And even if it did, if Intel hit 32nm in Q4 2009 who is to say the smaller ones would be ready in Q4 2010 (instead of Q1/2 2011). Not to mention some of the big changes on the horizon. Besides wiggle room (maybe everyone slides 12 months), there is always the possibility that GPUs will continue being "late commers" to new process nodes. When you don't control your own fabs you are very much at the mercy of others. One nice thing about hitting a new process node after it has proven stability is process maturity (better yields and performance) as well it means the next node could be on the horizon within 12-18 months, allowing for strategic price drops early in the lifecycle.

IMO, whoever MS partners with they won't only need to consider the above issues (along with architecture) but they also will need to be mindful of the companies roadmap beyond 32nm. Currently shopping the IP to TSMC, UMC, Charter, etc offers MS with some good options, but in 7 or 8 years how will these companies adapt to the new hurdles in process technologies? In this regards Sony is probably in a better position as they have partners in IBM, Toshiba, NEC, and AMD as well as have their own fabs to help dictate their own production schedules. Not necessarily cheaper, but having more control over your future can be vital.

Good points. I can't see how it will make financial sense for both Intel(or AMD) and MS to collaborate on a next gen chip though.

 

[Capeta]

What about stacking a GPU on top of a CPU eg 3D die stacking?

http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2925

 

[3dilettante]

That would make for some nice real-time smoke effects just from plugging it in.

I don't think I've seen any serious proposals for stacking two highly active circuit regions on top of each other. Usually it's either memory or they stagger the chips so the hot spots get access to cooling.

 

[Capeta]

That would make for some nice real-time smoke effects just from plugging it in.

I don't think I've seen any serious proposals for stacking two highly active circuit regions on top of each other. Usually it's either memory or they stagger the chips so the hot spots get access to cooling.

Internally a LSI is just a bunch of layers of circuitry stacked on top of one another.

Think of stacked dies as MCM but vertically instead of horizontally. I don't see a problem with cooling if you know what you are doing.

 

[Gubbi]

Internally a LSI is just a bunch of layers of circuitry stacked on top of one another.

Think of stacked dies as MCM but vertically instead of horizontally. I don't see a problem with cooling if you know what you are doing.

So tell us, how are you going to sink 300W from 100 mm^2 ?

Cheers

 

[Capeta]

So tell us, how are you going to sink 300W from 100 mm^2 ?

Cheers

We've had cooling devices that could cool a 200W chip for years, another 100Ws isn't going to magically transform into some quantum affect wall, but to answer your question a nice heat spreader + water or liquid alloy cooling technology coupled to a nice radiator or large heatsink.

 

[Shifty Geezer]

Even if that's possible, why choose that solution over separate dies? I can only see it saving space, and not money or complexity which is the main concern in these consoles.