Predict: The Next Generation Console Tech

[eastmen]

I'd hardly call the 400SP DX11 GPU in a Llano a "sub par graphics portion", its certainly a huge step up from both RSX and Xenos. Even with Ontario, a chip targetted at netbooks the GPU is a monster relatively speaking, low end discrete performance in a netbook is insane. Lllano will be limited by bandwidth in the real world of course, but in a console that's less of an issue because you can use faster RAM like GDDR5 and XDR2 and don't have to target particularly high rendering resoultions.

In both Ontario and Llano, roughly 2/3 or more of the die is being dedicated to the GPU portion, and in both cases the GPU portion is going to decimate everything else in their target market.

I'm not so sure how much of a step up it is http://www.anandtech.com/show/3933/amds-zacate-apu-performance-update

I wouldn't be surprised if a faster clocked waternoose/xenos wouldn't keep up or out perform these chips .


What would be the point of a new console with graphics similar to todays consoles ? Its not like with the wii that we got motion control . The 360/ps3 are getting motion control .

I would much rather have something that surpasses todays consoles by a huge margin. In all seriousness how much would a 12 core bobcat + 5770 actually cost in a next gen console. How much would the sucessor to the 5770 cost .

I think a single chip console from sony or ms would be the death of consoles esp if all 3 go with barely better than ps3/360 consoles. limiting us to low end performance of these fusion and sandy bridges would just let the pc become the dominate platform again. You can't stop progress and the high end of pcs will be way advance of whats capable on next gen consoles this way where as new consoles are normaly on par with the newest pc stuff of the time.

 

[Squilliam]

The original xbox 360 pushed much more wattage at the wall

I just don't see the point of intergrated video on a next gen console. It wouldn't make sense for anything but a wii 2 type set up. You can make something much better with a two chip approach.

The original Xbox 360 was designed for performance at all costs. Now how much power a console uses in standby or use does actually matter. I don't believe the market will be kind to a console which is considered grossly wasteful of power, especially the European market.

The more you raise your power the more you increase your costs. More power requires a more expensive, more complicated PCB, more elaborate cooling, possibly more warranty costs due to heat and on top of that you have to pay more for the CPU / GPU components. Theres simply no need to push the envelope, most people would be quite happy with a nice improvement in visuals in 3D and more in 2D.

The consoles only really got sane in terms of cost once they both had their first major revisions. The next generation will probably follow a similar power budget and form factor to the current consoles as they exist now. The power wars are dead, theres no going back to the idea of a razor/blade model uber expensive machine in order to make it back with accessories and games.

The above makes a lot of sense.
Just about all of the better reasoning in this thread tries to optimize the the cost/performance/power draw triad. But that really requires that you set up some (case by case) limits before you can start to get to specifics. TIme of availability, and thus process technology needs to be set as well.

We can also use cues from tidbits we have learnt here. For instance it is possibly cheaper to go with an on die pool of XRAM and cheaper DDR3 on a 128 bit bus than it is to get lower volume, higher power GDDR5 chips. One of the biggest costs in terms of power is moving data, and moving a lot of data off chip is the most expensive.

We also know that Microsoft leans on the GPU for Kinect, so if the same applies for the next generation then a low latency connection to the GPU would be ideal (thanks Sebbi).
So a chip with good data locality ought to out perform a chip which doesn't have as good data locality simply because it'll be more efficient.

A console could launch on 28nm with questionable (30-50%) yields if a shrink to 22nm is on the cards in the following year. I believe that for instance if someone was using Global Foundries, like Microsoft for instance they could afford to lose money on the first year with a large CGPU with on die cache with questionable yields if they can fix that issue within 12 months.

If 1920/1080 or 1280/720 in 3D is all they need to target, it shouldn't be too hard to ensure that a console would have enough die space, beit 16 or 24MB.

 

[eastmen]

The original Xbox 360 was designed for performance at all costs. Now how much power a console uses in standby or use does actually matter. I don't believe the market will be kind to a console which is considered grossly wasteful of power, especially the European market.

The more you raise your power the more you increase your costs. More power requires a more expensive, more complicated PCB, more elaborate cooling, possibly more warranty costs due to heat and on top of that you have to pay more for the CPU / GPU components. Theres simply no need to push the envelope, most people would be quite happy with a nice improvement in visuals in 3D and more in 2D.

The consoles only really got sane in terms of cost once they both had their first major revisions. The next generation will probably follow a similar power budget and form factor to the current consoles as they exist now. The power wars are dead, theres no going back to the idea of a razor/blade model uber expensive machine in order to make it back with accessories and games.

The ps3 launched at even higher power requirements. Not only that but many tasks like 1080p video playback should require alot less resources next gen than this gen so while doing other tasks power usage can be greatly scaled back. I personaly don't mind a 200w consolse and i doubt many people who buy at launch will care either. Esp if they launch at 28nm and 22nm is a year or so off. They wil lget power savings from that shrink.

We can also use cues from tidbits we have learnt here. For instance it is possibly cheaper to go with an on die pool of XRAM and cheaper DDR3 on a 128 bit bus than it is to get lower volume, higher power GDDR5 chips. One of the biggest costs in terms of power is moving data, and moving a lot of data off chip is the most expensive.

We also know that Microsoft leans on the GPU for Kinect, so if the same applies for the next generation then a low latency connection to the GPU would be ideal (thanks Sebbi).
So a chip with good data locality ought to out perform a chip which doesn't have as good data locality simply because it'll be more efficient.

A console could launch on 28nm with questionable (30-50%) yields if a shrink to 22nm is on the cards in the following year. I believe that for instance if someone was using Global Foundries, like Microsoft for instance they could afford to lose money on the first year with a large CGPU with on die cache with questionable yields if they can fix that issue within 12 months.

If 1920/1080 or 1280/720 in 3D is all they need to target, it shouldn't be too hard to ensure that a console would have enough die space, beit 16 or 24MB.

So does anyone have benchmarks of a high performance CGPU ? Does anyone know how any of the CGPU's scale once you go past 1 or 2 GPU cores ?

I don't see a CGPU system being capable of giving users a graphical reason to upgrade to new machines

 

[brain_stew]

I'm not so sure how much of a step up it is http://www.anandtech.com/show/3933/amds-zacate-apu-performance-update

.

Those are benchmarks for AMD's netbook chip, Ontario/Zacate, the GPU component in Llano (which is what I was comapring to Xenos/RSX) has 5x the SPs of that part. Just take a look at 5670 reviews (since its the same graphics core) if you want to see what that part can do when not bottlenecked by memory bandwidth, it blows Xenos away in multiplatform games and that's with the added PC overhead.

Sure, AMD's netbook APU doesn't outperform Xenos + Xenon yet but come on, the fact that its even up for debate is impressive enough, this is a 9w/18w part we're talking about and the 360 still sucks 100w+. Previous netbook designs struggle to run Quake 3 at native resolution.

Too many people seem to be under the impression just because Nintendo succeeded building a low power cheep console console with no loss t launch that everyone's going to suddenly go that way. Nintendo has been doing that for years with varying levels of success and it hasn't changed Microsoft and Sony's strategy. .

No one has suggested they use the Wii strategy of simply shrinking and overclocking your previous generation hardware though, far from it. What has been suggested is something more akin to the design strategy behind the GCN. That's a console that was small, cheap, low power and yet it was still very competitive in the graphics department because Nintendo focused on increasing efficiency and removing bottlenecks not using the latest cutting edge silicon with no concern for power consumption, heat output or size. Using that sort of design philosophy doesn't preclude fantastic graphics and performance, but it does ensure that they don't come at the expense of everything else (cost, power consumption, heat, size, reliability etc.)

There's a balance to be had between the Wii's approach and the PS3 approach.

Also, Nintendo ahven't been utilising a strategy like the one we saw with the Wii "for years." The Wii was designed completely differently to all their previous consoles and with the 3DS we can see that they've reverted to their old design approach.

 

[brain_stew]

So does anyone have benchmarks of a high performance CGPU ? Does anyone know how any of the CGPU's scale once you go past 1 or 2 GPU cores ?

No one will have that data until one is employed in a closed box. You simply have no way of exploiting that tight integration on the PC. These chips are also limited to slow DDR3 RAM in the PC space for the forseeable future, whereas you can use them with super quick XDR2 (or similar) in a closed box and quickly get rid of the bandwidth bottleneck (which is the number one thing holding back their performance at present).

I think the fact that both Tim Sweeney and Andrew Richards strongly argued for the move to integrated CGPUs despite having vastly different views on the future of graphics is quite telling. If both of their wildly different approaches would benefit immensely from the elimination of the CPU-GPU latency then I feel that covers most usage cases. They even suggested that if a two chips solution was necessary to meet performance goals then dual CGPUs was desirable rather than discrete CPU and GPU components and it makes sense. If your CPU and GPU are sharing a lot of workloads (and whether you feel a new CELL will be taking over more graphics tasks or a GPU will be taking over more CPU tasks, its inevitable) then your first priority has to be to make sure that communication between them is as painless and latency free as possible.

 

[Xenus]

No one has suggested they use the Wii strategy of simply shrinking and overclocking your previous generation hardware though, far from it. What has been suggested is something more akin to the design strategy behind the GCN. That's a console that was small, cheap, low power and yet it was still very competitive in the graphics department because Nintendo focused on increasing efficiency and removing bottlenecks not using the latest cutting edge silicon with no concern for power consumption, heat output or size. Using that sort of design philosophy doesn't preclude fantastic graphics and performance, but it does ensure that they don't come at the expense of everything else (cost, power consumption, heat, size, reliability etc.)

There's a balance to be had between the Wii's approach and the PS3 approach.

Also, Nintendo haven't been utilizing a strategy like the one we saw with the Wii "for years." The Wii was designed completely differently to all their previous consoles and with the 3DS we can see that they've reverted to their old design approach.

You totally missed the point. I wasn't suggesting that they would use the Wii strategy. I was suggesting something to the Gamecube which has been Nintendo's design strategy for years. However the fact still stands that Sony and Microsoft have never gone that strategy. The fact remains just cause Nintendo won this round doesn't mean they will jump to copy them in every aspect. Obviously there were flaws in the PS3 and Xbox360 strategy and there is a power envelope that is physically possible to coll in a box of set size with a given budget for cooling. However I see Sony and Microsoft pushing close to that and their console isn't going to be profitable on launch like the Gamecube.

 

[Urian]

About Microsoft next generation I believe that they will choose a 32nm PowerPC A2 derivate with 8 cores runningt at 3.2Ghz. The PowerPC A2 design with 16 cores is 428mm^2 at 45nm and with a density of 1.43 Bilion transistors, reduce the numbers to the half with 8 cores and then made the CPU using a 32nm process.

 

[Squilliam]

The ps3 launched at even higher power requirements. Not only that but many tasks like 1080p video playback should require alot less resources next gen than this gen so while doing other tasks power usage can be greatly scaled back. I personaly don't mind a 200w consolse and i doubt many people who buy at launch will care either. Esp if they launch at 28nm and 22nm is a year or so off. They wil lget power savings from that shrink.

Those 200W consoles also lost > $200 at launch in the case of the PS3 and at least >$100 for the Xbox 360 with further issues from reliability. Now if we're going to be seeing a higher BOM from new interfaces etc then the one thing which they can sacrafice to get the system into a palatable $299 or $399 price point has to be performance, especially when not all of the market is keen for having the outright fastest console, proof positive being the PS2 and Wii. I don't mind a 200W console but I like many other people aren't really willing to pay for it, especially as we probably aren't going to be seeing a repeat of the massive losses at the start of this generation.

So does anyone have benchmarks of a high performance CGPU ? Does anyone know how any of the CGPU's scale once you go past 1 or 2 GPU cores ?

I don't see a CGPU system being capable of giving users a graphical reason to upgrade to new machines

You know... the highest performance CGPU is currently the Xbox 360. Heaps of people are upgrading to it. We have already got good precedent. In the PC space we don't have benchmarks yet because its new, however look at the anandtech Sandybridge sneak preview, it performs admirably and performs similarly to a low end discrete GPU like the 5450.

No one will have that data until one is employed in a closed box. You simply have no way of exploiting that tight integration on the PC. These chips are also limited to slow DDR3 RAM in the PC space for the forseeable future, whereas you can use them with super quick XDR2 (or similar) in a closed box and quickly get rid of the bandwidth bottleneck (which is the number one thing holding back their performance at present).

Are we likely to see XDR2 in a console? Will it measure up from a cost perspective against royalty free, bulk fabbed memory technology such as GDDR5 and DDR3?

 

[Entropy]

About Microsoft next generation I believe that they will choose a 32nm PowerPC A2 derivate with 8 cores runningt at 3.2Ghz. The PowerPC A2 design with 16 cores is 428mm^2 at 45nm and with a density of 1.43 Bilion transistors, reduce the numbers to the half with 8 cores and then made the CPU using a 32nm process.

The PowerPC A2 is a heavily multicore/multithread network/server processor with relatively low performance and power draw per thread. Its qualities and intended use lie quite far from gaming consoles.

 

[Shifty Geezer]

Are we likely to see XDR2 in a console? Will it measure up from a cost perspective against royalty free, bulk fabbed memory technology such as GDDR5 and DDR3?

If you're targeting low cost, no more than $50 on RAM at launch or whatever, no. If you're targeting high performance, 256 GB/s, the savings in chips, mobo and cooling for a given very high bandwidth suggests 'yes' to me. If a system design decides to go with one large pool of very fast unified RAM, 2+ GBs of 200+ GB/s, XDR2 seems the logical choice. If a system decides to go with a low-cost design, split pools of slower RAM or eDRAM or negative high bandwidth requirements seems necessary.

 

[upnorthsox]

Are we likely to see XDR2 in a console? Will it measure up from a cost perspective against royalty free, bulk fabbed memory technology such as GDDR5 and DDR3?

XDR2, probably not. but then there's still XDR which is about the same price as GDDR3 but equal to GDDR5 in performance. If it weren't for the current litigation I would not have been surprised to see Nividia use it for Fermi and beyond. DDR3 won't be in a console, I'm pretty darn sure of that. And before we get into a pricefest over GDDR5, let's see if the Tiawanese/S Korean governments are willing to bankroll another round of mem chip dumping like they did for GDDR3/DDR2/3. So far they haven't.

 

[upnorthsox]

If you're targeting low cost, no more than $50 on RAM at launch or whatever, no. If you're targeting high performance, 256 GB/s, the savings in chips, mobo and cooling for a given very high bandwidth suggests 'yes' to me. If a system design decides to go with one large pool of very fast unified RAM, 2+ GBs of 200+ GB/s, XDR2 seems the logical choice. If a system decides to go with a low-cost design, split pools of slower RAM or eDRAM or negative high bandwidth requirements seems necessary.

This should go without saying but I do see why you have to say it. To underscore the point, GDDR3 was still new when MS spec'd it for the 360 and Sony was the first major buyer of XDR for the PS3.

 

[Teasy]

You totally missed the point. I wasn't suggesting that they would use the Wii strategy. I was suggesting something to the Gamecube which has been Nintendo's design strategy for years. However the fact still stands that Sony and Microsoft have never gone that strategy. The fact remains just cause Nintendo won this round doesn't mean they will jump to copy them in every aspect. Obviously there were flaws in the PS3 and Xbox360 strategy and there is a power envelope that is physically possible to coll in a box of set size with a given budget for cooling. However I see Sony and Microsoft pushing close to that and their console isn't going to be profitable on launch like the Gamecube.

AFAIR GameCube wasn't profitable at launch, the hardware was sold at a small loss initially.

 

[brain_stew]

Are we likely to see XDR2 in a console? Will it measure up from a cost perspective against royalty free, bulk fabbed memory technology such as GDDR5 and DDR3?

While the per chip cost may be higher, you have to look at the savings it can allow elsewhere.

We know we'll have 4Gb chips ready for launch, but there's no roadmap for 4Gb GDDR5 chips yet (they should be there, but there's not a guarantee) so there's the potential you could use half as many chips.

It'll allow you to use a single cheap 128 bit memory bus while delivering >200GB/s of bandwidth. If you want that sort of bandwidth out of GDDR5 you're going to need a single 256 bit link or two separate 128 bit links and both of them solutions will increase board complexity and the number of RAM chips required. So if your targets are >200GB/s of bandwidth then I believe yes, it will indeed be the cheaper solution, especially over time as the lower board complexity will pay dividends later on when trying to target smaller form factors.

If 1 or 2 of the next generation consoles go with XDR2/XDR3 (and honestly, both Microsoft and Sony really should, its the best option out there) then the chips will be produced in mass quantities by default. They'll still be pricey at launch but shouldn't be prohibitively expensive as the years roll by.

AFAIR GameCube wasn't profitable at launch, the hardware was sold at a small loss initially.

Only something in the region of $5/$10 though and that was with a super low $200 launch price and only for a very short time. It only took them a couple of years to get a profitable $99 machine out of that design. That's the right sort of compromise Sony and Microsft need to be making. No masses losses on launch and a machine that is able to target price points like $200/$100 a few years later without taking a hit.

However, it does show that this fallacy that Nintendo has only ever produced consoles with >100% and without competitive graphics solutions to be complete crap. The Cube's best titles blew away anything produced on the PS2 and were surprisingly competitive with the best Xbox offerings, and that's despite the fact it was much smaller and cheaper to produce than both those competing solutions.

 

[Squilliam]

If you're targeting low cost, no more than $50 on RAM at launch or whatever, no. If you're targeting high performance, 256 GB/s, the savings in chips, mobo and cooling for a given very high bandwidth suggests 'yes' to me. If a system design decides to go with one large pool of very fast unified RAM, 2+ GBs of 200+ GB/s, XDR2 seems the logical choice. If a system decides to go with a low-cost design, split pools of slower RAM or eDRAM or negative high bandwidth requirements seems necessary.

Given the overall costs of everything, and the need to turn a profit eventually, wouldn't the focus instead be on improving the main die(s)? What I mean is that it seems to be simpler and cheaper to make the CPU, GPU, CGPU smaller given the fact that they could probably count on at least a couple of die shrinks during the lifetime of the system than it is to make a wide memory bus cheaper or a large number of ram chips cheaper. So wouldn't the likely design constraints for the next generation consoles involve narrower buses, perhaps even 64 bits as an extreme, few RAM chips and relatively abundant on die cache for both the CPU and GPU?

XDR2, probably not. but then there's still XDR which is about the same price as GDDR3 but equal to GDDR5 in performance. If it weren't for the current litigation I would not have been surprised to see Nividia use it for Fermi and beyond. DDR3 won't be in a console, I'm pretty darn sure of that. And before we get into a pricefest over GDDR5, let's see if the Tiawanese/S Korean governments are willing to bankroll another round of mem chip dumping like they did for GDDR3/DDR2/3. So far they haven't.

If it wasn't for that litigation, would XDR be that cheap? Perhaps the reason for the price is that companies are unwilling to use that technology.

Anyway it doesn't make sense that XDR would have the performance of GDDR5 given a similar size memory bus yields similar performance with GDDR3. Are you talking about XDR2? Also where are you getting the relative prices?

 

[Squilliam]

While the per chip cost may be higher, you have to look at the savings it can allow elsewhere.

We know we'll have 4Gb chips ready for launch, but there's no roadmap for 4Gb GDDR5 chips yet (they should be there, but there's not a guarantee) so there's the potential you could use half as many chips.

It'll allow you to use a single cheap 128 bit memory bus while delivering >200GB/s of bandwidth. If you want that sort of bandwidth out of GDDR5 you're going to need a single 256 bit link or two separate 128 bit links and both of them solutions will increase board complexity and the number of RAM chips required. So if your targets are >200GB/s of bandwidth then I believe yes, it will indeed be the cheaper solution, especially over time as the lower board complexity will pay dividends later on when trying to target smaller form factors.

How wide is the address required for a 2/4Gb XDR2 chip? In the PS3 they have 32 bits per module, is 32 bit standard or could they use 16 bit as well? If there were 4Gb 16 bit address chips available, then surely 16Gb with 100GB/S transfer rate could be ample for a simple and cost effective design?

In any case with 200GB/S it seems there wouldn't be much need for on die cache. You could get away with a relatively puny on die cache structure which makes sense given that Sony already did that with the PS3 with the Cell processor and its paired XDR.

 

[brain_stew]

Given the overall costs of everything, and the need to turn a profit eventually, wouldn't the focus instead be on improving the main die(s)? What I mean is that it seems to be simpler and cheaper to make the CPU, GPU, CGPU smaller given the fact that they could probably count on at least a couple of die shrinks during the lifetime of the system than it is to make a wide memory bus cheaper or a large number of ram chips cheaper. So wouldn't the likely design constraints for the next generation consoles involve narrower buses, perhaps even 64 bits as an extreme, few RAM chips and relatively abundant on die cache for both the CPU and GPU?

Indeed they will, and that's precisely why going with XDR2 (or XDR3 if its available in time) makes so much sense. At the same bus width it delivers twice the bandwidth of GDDR5, so you can hit your bandwidth goals with a much narrower bus and fewer chips. GDDR5 with a 64 bit bus really isn't going to deliver enough bandwidth to ensure a significant leap over these current consoles (it'd only give you ~50GB/s, that's fine for the 720p Wii 2, not so much for PS4/Xbox3 imo) while targeting 2x the resolution. XDR2/XDR3 might be enough even on a 64 bit bus though since we're still talking in excess of 100GB/s of bandwidth which could be pretty effective in combination with a large shared L3 cahce utilising IBM's eDRAM tech.

http://www.rambus.com/us/technology/solutions/xdr2/xdr2_vs_gddr5.html

 

[brain_stew]

How wide is the address required for a 2/4Gb XDR2 chip? In the PS3 they have 32 bits per module, is 32 bit standard or could they use 16 bit as well? If there were 4Gb 16 bit address chips available, then surely 16Gb with 100GB/S transfer rate could be ample for a simple and cost effective design?

32 bit is indeed the standard, no idea if 16 bit chips are available but considering this stuff was basically made entirely to target next generation consoles, then if that's a demand, I wouldn't be surprised if it could be met.

In any case with 200GB/S it seems there wouldn't be much need for on die cache. You could get away with a relatively puny on die cache structure which makes sense given that Sony already did that with the PS3 with the Cell processor and its paired XDR.

The large on die cache is less about easing bandwidth concerns and more about having a zero latency interface between the integrated CPU and GPU. If you have a say 15MB-20MB of shared L3 cache (which doesn't take up all that much die space if you use IBM's tech) then both your CPU and GPU can work on your render targets (or physics simulations) in tandem without the developer ever having to worry about hiding latency, which will save major headaches (the number one headache if you believe Sweeney and Richards). The near limitless bandwidth is a nice bonus as well of course!

I'd ditch the external 128 bit link (for a 64 bit link) much sooner than I'd ditch the shared L3 cache.

 

[corduroygt]

The problem with XDR besides the possibly higher cost due to lack of popularity in the market is that there are no existing GPU designs that use it. On the PS3, RSX can only access the XDR by going through the Cell, which is not as fast as direct access of course. You'd have to design an XDR memory interface for the GPU, which might end up being costly. Also, you have to use 32-bit addressing if you want to be able to address up to 4GB of memory

 

[brain_stew]

The problem with XDR besides the possibly higher cost due to lack of popularity in the market is that there are no existing GPU designs that use it. On the PS3, RSX can only access the XDR by going through the Cell, which is not as fast as direct access of course. You'd have to design an XDR memory interface for the GPU, which might end up being costly. Also, you have to use 32-bit addressing if you want to be able to address up to 4GB of memory

Well these proposed consoles don't even have a discrete GPU. They'll be sharing the CPU's memory interface and L3 cache since they're on the same chip.