Predict: The Next Generation Console Tech

Status
Not open for further replies.
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.



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! :D

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


XDR is has a programmable width from 32 to 2 bits,
 
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.

I'm also unaware of any CPU's other than the cell that have an XDR interface, and Cell doesn't have an integrated GPU inside, and also very unlikely to be used in anything except maybe the PS4.
 
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?



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?

No I'm talking XDR, both have clock speeds up to 7.2GHz and voltages around 1.8V, both have 2Gb chips on their roadmaps though GDDR5 now has one ready for production. XDR though is running at 3.2GHz in millions of machines for 5 yrs now where GDDR5 still has nothing in production running over 1.5GHz. Basically, GDDR5 is now catching up to XDR yrs after production.

Relative prices are from updated BOM costs of the PS3 and may not reflect actual price, but then one shouldn't equate GDDR5 prices in relation to GDDR3 prices either. As I said in another post, it has yet to be seen if GDDR5 will get the same kind of subsidies that GDDR3 got.
 
I don't think it would be that difficult for Nvidia to attach an XDR memory controller to a GPU and adapt it to make it work with Rambus instead of JEDEC standard. It would probably cost more as it will be a custom solution, but not that much more.

I think the PS4 will be more like this:

Enhanced Cell with more PPEs all with at least two threads. SPEs with a slightly larger instruction set and maybe even out-of-order execution (though I'm not certain it will be worth the extra silicon), but sticking 8 for a small die size. Combined with an Nvidia GPU, probably a Kepler derivative with similar compute power to a GTX480. Shared L3 cache, big, but not enough to blow the budget.

4GB XDR2 on a 128bit interface which should be good for over 100GB/s bandwidth which is plenty for 1080p120.

Basically Sony will be looking to build on what turned out to be a relatively successful design, I mean if PS3 had launched with an G80 class GPU it would be far and above the best console in terms of power so just moving away from G70 will be a big boost for Sony. Unified memory will be another big boost. The PS3 has a lot of architecture deficiencies and if Sony can work methodically to remove bottlenecks they won't need to beef it up too much to get 1080p120. They could even look at low end raytracing for their internal studios if they play it right. The silicon budget doesn't even have to be that big if they make the right architectural improvements.
 
I'm also unaware of any CPU's other than the cell that have an XDR interface, and Cell doesn't have an integrated GPU inside, and also very unlikely to be used in anything except maybe the PS4.

The CELL chip in PS4 probably will though and I don't see why Microsoft wouldn't be a target partner for XDR2, its a perfect fit for any next generation console.
 
I don't think it would be that difficult for Nvidia to attach an XDR memory controller to a GPU and adapt it to make it work with Rambus instead of JEDEC standard. It would probably cost more as it will be a custom solution, but not that much more.

I think the PS4 will be more like this:

Enhanced Cell with more PPEs all with at least two threads. SPEs with a slightly larger instruction set and maybe even out-of-order execution (though I'm not certain it will be worth the extra silicon), but sticking 8 for a small die size. Combined with an Nvidia GPU, probably a Kepler derivative with similar compute power to a GTX480. Shared L3 cache, big, but not enough to blow the budget.

4GB XDR2 on a 128bit interface which should be good for over 100GB/s bandwidth which is plenty for 1080p120.

I

Basically Sony will be looking to build on what turned out to be a relatively successful design, I mean if PS3 had launched with an G80 class GPU it would be far and above the best console in terms of power so just moving away from G70 will be a big boost for Sony. Unified memory will be another big boost. The PS3 has a lot of architecture deficiencies and if Sony can work methodically to remove bottlenecks they won't need to beef it up too much to get 1080p120. They could even look at low end raytracing for their internal studios if they play it right. The silicon budget doesn't even have to be that big if they make the right architectural improvements.

I don't think there'll be a need to target 1080p/120hz. HDMI1.4a's specifications for 3D max out at 720p/120hz (or variations thereof like a double wide 720p image) and I've seen no indications that there's any plans for the introduction of a new standard or that there's any desire or demand for one anyway. 1080p/60hz, 720p/120hz and 720px2/60hz are the max of what this box will have to target and I fully expect something like 1280x1080 w/ improved MLAA horizontally scaled to 1080p with a full 1080p 2D HUD (á la GT5) to be a more common target.

XDR2 @ 128 bit should actually be able to give you >200GB/s bandwidth.

My design ideas are quite similar to yours, though I'd rather have a larger number of simpler SPEs than a small number of complex SPEs.
 
I don't think there'll be a need to target 1080p/120hz. HDMI1.4a's specifications for 3D max out at 720p/120hz (or variations thereof like a double wide 720p image) and I've seen no indications that there's any plans for the introduction of a new standard or that there's any desire or demand for one anyway. 1080p/60hz, 720p/120hz and 720px2/60hz are the max of what this box will have to target and I fully expect something like 1280x1080 w/ improved MLAA horizontally scaled to 1080p with a full 1080p 2D HUD (á la GT5) to be a more common target.

XDR2 @ 128 bit should actually be able to give you >200GB/s bandwidth.

My design ideas are quite similar to yours, though I'd rather have a larger number of simpler SPEs than a small number of complex SPEs.

1080p120 will be the 'new' tech that Sony will introduce with PS4. They will call it something like headache free/ultimate 3D and release a bunch of 1080p 600Hz TVs to go alongside it. I think the standard for internal developers will be 1080p60 with 3D and third parties will target 1080p30 without 3D. 1080p is much more important than 60fps as long as 30fps is v-synced.

The other thing with making improvements to the current architecture than whole sale changes is that it will make for an easier upgrade path for developers and they will get up to speed much quicker than they did with PS3 or PS2. Moving to ARM or x86 will make developers learn a whole new bag or tricks (well less for x86) and institute a whole bunch of costs in training and retraining that they don't need right now. It could push a third parties to go for a wait and see approach which would spell doom for PS4. I think had third parties known what was in store for PS3 in terms of complexity and market failures then they would have taken a similar approach and we could ave seen the end of Sony console development. Luckily for Sony they have remedied the situation, but their redemption won't be complete until they release a decent power, low cost box like PS2. They need the "Two hundred and ninety-nine dollars" moment again to retake their position as a market leader.
 
If XDR and XDR2 is so fantastic, why haven't either AMD or Nvidia implemented an architecture which takes advantage? If they can get a 128bit card to perform just as good if not better than a 256bit card, why wouldn't they take advantage? What is the marketing not telling us?
 
If XDR and XDR2 is so fantastic, why haven't either AMD or Nvidia implemented an architecture which takes advantage? If they can get a 128bit card to perform just as good if not better than a 256bit card, why wouldn't they take advantage? What is the marketing not telling us?

The amount of litigation that Rambus have brought against Nvidia and their board partners probably has more than a little something to do with that. It'd probably piss off their board partners as well since there'd be fewer suppliers to choose from and less supply generally.

Sony on the other hand have been one of Rambus' key partners for several years now (both the PS2 and PS3 used RAM supplied by Rambus) and they've already created an XDR interface in partnership with IBM, so they're much more cosy partners. They can build on the work they've already done. I don't really see what Sony has to gain by moving to a completely new memory standard?
 
Indeed Sony likes Rambus. Flex-io was also their doing. However Rambus has sued just about every manufacturer of RAM and Nvidia among others. Basically they vigorously defend their patents cause they make most of their money from Royalties and haven't made too many friends in the business by doing so.
 
If XDR and XDR2 is so fantastic, why haven't either AMD or Nvidia implemented an architecture which takes advantage? If they can get a 128bit card to perform just as good if not better than a 256bit card, why wouldn't they take advantage? What is the marketing not telling us?

There is no love lost between Rambus and many other companies including Nvidia, AMD, Intel and generally any JEDEC partner.

Sadly, while Rambus have got great tech, their legal department all need to be sacked. They create trouble within the industry protecting patents that they own through questionable means. The worst part though is that Rambus are one of the few DRAM companies that are still innovating rather than waiting around for JEDEC to get their arse in gear. Stuff like the terrascale initiative wouldn't happen any where else. GDDR5 was spec'd to compete with XDR, but missed and there is nothing planned to compete with XDR2, while Rambus are already working on a spec for XDR3 for 1TB/s over 128bits.

I fully expect Sony to use XDR2 in PS4 and MS to look at it for Xbox 3. To not look at XDR2 would be negligent as it offers basically everything a games console needs in terms of memory bandwidth and ability to reduce costs (small bus width). Having a 256bit bus would hamper any attempts by MS to reduce their GPU costs.
 
If XDR and XDR2 is so fantastic, why haven't either AMD or Nvidia implemented an architecture which takes advantage? If they can get a 128bit card to perform just as good if not better than a 256bit card, why wouldn't they take advantage? What is the marketing not telling us?

Well GDDR is AMD's baby, using something else would be more damning than if a TT champion got beat using it on TV. Nividia on the otherhand has other issues, they're paying royalties either to AMD, their competitor, or Rambus, who they are currently in litigation with. In the end I think they decided to gamble on the Tiawanese bankrolling another round mem chip dumping.

XDR2 simply didn't/hasn't get the blessing of the Chinese and without that it's not going to get manufactured unless you bring in your own large buyer like Sony.
 
Indeed Sony likes Rambus. Flex-io was also their doing. However Rambus has sued just about every manufacturer of RAM and Nvidia among others. Basically they vigorously defend their patents cause they make most of their money from Royalties and haven't made too many friends in the business by doing so.

It is also true that many DRAM makers were found guilty of price fixing and were using their cartel to pressure RAMBUS out of the market (in the Intel-RAMBUS days)... so they have been emplying anti-competitive practices against RAMBUS.
 
It is also true that many DRAM makers were found guilty of price fixing and were using their cartel to pressure RAMBUS out of the market (in the Intel-RAMBUS days)... so they have been emplying anti-competitive practices against RAMBUS.

Well Rambus did just come in and patent everything and then try and licence other company's tech back to them. It's pretty scandalous behaviour. Don't get me wrong, the other companies were just as bad fixing their prices to force Rambus out of the market.

My personal view is that Rambus still have some innovative designs, while the others sit around in committee after committee waiting for someone else to do the work. It's why all of the DDR specs are so middle of the road compared to what comes out of Rambus. I mean RDRAM was so far ahead of its time compared to SDRAM, it's just a shame it was so expensive (which we now know was down to price fixing rather than some inherent cheapness of SDRAM which we all thought lead to the situation).
 
Well Rambus did just come in and patent everything and then try and licence other company's tech back to them. It's pretty scandalous behaviour. Don't get me wrong, the other companies were just as bad fixing their prices to force Rambus out of the market.

My personal view is that Rambus still have some innovative designs, while the others sit around in committee after committee waiting for someone else to do the work. It's why all of the DDR specs are so middle of the road compared to what comes out of Rambus. I mean RDRAM was so far ahead of its time compared to SDRAM, it's just a shame it was so expensive (which we now know was down to price fixing rather than some inherent cheapness of SDRAM which we all thought lead to the situation).

To understand why the rest of the memory world is so dead set against working with Rambus, you have to know a at least a little about the patent system. For an idea to be patentable, it has to
A: be non-trivial. If the idea is at a level that someone "in the field" could reach by extension, or easy realization, it is not (should not be) patentable.
B: have "height". Again, the idea can't be to close to something existing or too trivial.

However, this means that what is patentable or not is up to the local patent office where you submit your application. There is no guarantee that the same idea will be treated the same even in patent offices in the same country, much less across the world. To be blunt, in the eyes of the rest of the world, the US standards when it comes to what level an idea has to have to be patentable is quite low. If something new comes up, you can quickly come up with a bunch of rather trivial ideas and effectively block the field. Furthermore, the US has rather odd litigation practises - if for instance Rambas wants to sue Toshiba, they can do it in some godforsaken hole with a drunkard judge, and that is where it will be handled. And Toshiba cannot say that they'd rather want the case to be handled by some kind of international patent court, such an entity doesn't exist. Anyone believe that for instance an Indian company, sued by an American company over a patent granted by an American patent office, in an American court has much of chance? It's a nightmare scenario for any non-US company.

The rest of the world is effectively forced to play by the rules set down by US patent standards and litigation practices, and we are not happy about it. At all.


I'm not sure someone born and bred in the US can understand the frustration of the rest of the (tech)world, and how abhorrent a company like Rambus is. Nobody has a problem with corporations developing new tech and patenting their good stuff for proprietary use or licensing, but Rambus is very artfully leveraging the (bizarre) US patent system and litigation legislation against the rest of the industry.
They are pariah, and ironically their methods is the one thing that has stopped a good idea that could have reached widespread use long ago. It's sad, really. Patent squatting killing something good. Then again, it's not the cure for cancer. I doubt that the lack of uptake has made any difference in the greater scheme of things. But damn, in this day and age, patents need to be handled at an international level.
 
Last edited by a moderator:
If you don't want the US legal system you don't have to do buisiness in the USA.

You make it sound as if ignoring the biggest computer electronics market in the world is somehow an option; its not. Any RAM spec. is dead in the water if it can't be used in the US, as is any console, any GPU or CPU platform, and you know it.
 
You make it sound as if ignoring the biggest computer electronics market in the world is somehow an option; its not. Any RAM spec. is dead in the water if it can't be used in the US, as is any console, any GPU or CPU platform, and you know it.
Can't you have a separate company in Taiwan just for producing the memory, and another one that acts just as a distributor to get around that?
 
Does anyone have (roughly) accurate figures on what the transistor count is per MB of IBM eDRAM tech? Power7 is a 567mm2 design at 45nm with 1.2bn transistors. Some die shots here:

http://aceshardware.freeforums.org/power-7-t891-15.html

Seems to be packed quite dense. Fitting ~20MB in a 2bn transistor 22nm console integrated CPU/GPU shouldn't be too difficult assuming it scales near linearly (I'm guessing it won't?). Looks like too big a design win to pass up, but I'd like some accurate figures to work with.

20MB would be enough to fit an entire 1080p framebuffer (assuming no msaa since MLAA and its derivatives will probably be much more popular next generation) and still have a decent amount of space left over to work as a generic CPU L3 cache as well. With MRTs becoming so popular it seems silly to go with much more than that as no amount of L3 cache is ever going to be big enough to fit all your render targets at once, so long as they can each easily be swapped in and out of cache then I think the solution is fine.
 
Last edited by a moderator:
Status
Not open for further replies.
Back
Top