Predict: The Next Generation Console Tech

Status
Not open for further replies.
the power 7 use eDram has an L3 cache (same function as the L3 on phenom, i7 and others), with the trade-off that it's slower but bigger and very dense.
that article is pointless though. They copy-pasted Power7's specs and pretend it's the PS4 version
http://arstechnica.com/hardware/new...er7-twice-the-muscle-half-the-transistors.ars

The author is not afraid of ridicule. besides, it's an enterprise CPU. socket 1366 would already be too expensive for a console. That CPU though is higher end than Xeon and Opteron and competes with Nehalem EX, Itanium and Sparc.
Its 32MB L3 is meant for big server loads and the 100GB/s are achieved through two four-channel memory controllers.

a CPU the size of a GPU, with a 512bit bus to memory, optimized for running high end databases with high throughut, in a console? I doubt that :p

A very custom version, why not. but it would look more like a Xenon with a new architecture, not a copy-paste. What next? an Ultrasparc T3, or a mainframe?
 
I wouldn't rule out a console CPU core based on Power 7. The cores themselves are 28mm^2 (256KB level 2 cache included) in a 45 nm process. Ditch some of the more exotic features (like BCD floating point) and make a wider vector unit. That would probably be about 12-15 mm^2 per core in 22nm. Stick 8 or 12 of these on a chip (32/48 threads) with a fair amount of cache, and you would have a pretty capable main processor.

Cheers
 
I wouldn't rule out a console CPU core based on Power 7. The cores themselves are 28mm^2 (256KB level 2 cache included) in a 45 nm process.
For comparison an SPE is less than 6.5 mm^2 in a 45 nm process.

So in the same space you could have either one Power7 core (unmodded) with 256 kB cache or four SPEs with a total of 1 MB LS.

Which is the best thing to have on a console if the die space is fixed?
 
For comparison an SPE is less than 6.5 mm^2 in a 45 nm process.

So in the same space you could have either one Power7 core (unmodded) with 256 kB cache or four SPEs with a total of 1 MB LS.

Which is the best thing to have on a console if the die space is fixed?

And consoles do not need the DP. What is the estimated figure of flops for Power7? Wikipedia seems to imply that it's 258.6 GFlops for the 8 core version (I suppose that's DP?). It's not much more than Cell BE if single.
 
What if Microsoft makes a new change in the architecture of the CPU now that AMD is licensing fabless designs? The situation is different than the one in 2003 for the PowerPC processors and now that IBM has closed the Cell development team perhaps it has affected to Microsoft next console current design.

What do you think about a 10-Core Bobcat running at 3.2Ghz with SSE5?
 
What if Microsoft makes a new change in the architecture of the CPU now that AMD is licensing fabless designs? The situation is different than the one in 2003 for the PowerPC processors and now that IBM has closed the Cell development team perhaps it has affected to Microsoft next console current design.

What do you think about a 10-Core Bobcat running at 3.2Ghz with SSE5?
The only thing that I can think of that would keep them from doing it is BC but I was under the impression that emulating a CPU is a somewhat easy thing to do in comparison to a GPU. What would they do about VMX128 though?
 
Or is this all by the by, and it will all end up in the cloud?

http://www.mcvuk.com/news/36665/Square-Enix-Consoles-set-for-extinction

Are we going back to the dumb terminal?

The big 3 might feel the same way but something tells me that they would actually work on both hardware and the server side stuff at the same time. Maybe the world will be ready for this in 2012 or maybe it won't. I expect the next gen consoles to try and blend both. It sounds too radical of a shift right now but 2 years from now who knows.
 
I think next gen we'll see the occasional developer dabble into off-site processing (cloud), but I don't see entire games being processed off-site being the norm for a looooong time if ever.
 
Again, I guess, it depends on the application. Branch Heavy code would run better on the Power7 and vice versa (considering, they'd upgrade the SPEs)

Well, the application is games in 95% of the cases on consoles and the other applications are usually not that taxing on the CPU anyways.

Games tend to be computation heavy so I think games would benefit more from 4 SPEs than 1 Power7 core. Another benefit of the SPEs are that they don´t depend on having an additional Level 3 cache to sustain performance, so that extra die space could be used for additional SPEs as well.
 
I know everybody is going to jump all over me for this, but it seems people like SPE's because they get used for graphics. Well, a much better solution would be to put those resources into the GPU in the first place. In other words if it's 4 more SPE's or 20% more on the GPU, the latter is far more effective at painting pretty pictures which face it, is what we actually care about. Because I dont see Cell doing any mondo physics games or something. But I do see it doing depth of field in Uncharted 2.
 
I know everybody is going to jump all over me for this, but it seems people like SPE's because they get used for graphics. Well, a much better solution would be to put those resources into the GPU in the first place. In other words if it's 4 more SPE's or 20% more on the GPU, the latter is far more effective at painting pretty pictures which face it, is what we actually care about. Because I dont see Cell doing any mondo physics games or something. But I do see it doing depth of field in Uncharted 2.

I don´t think people will jump on you, but you should take your reasoning to the general case because this is not unique to Cell.

Where is the best place to add additional performance? Maybe the benefits of scaling the CPU is hitting a wall, maybe it´s best to add those resources to a GP-GPU? Maybe the CPU will soon end up in one corner of the GPU die?
 
I know everybody is going to jump all over me for this, but it seems people like SPE's because they get used for graphics. Well, a much better solution would be to put those resources into the GPU in the first place. In other words if it's 4 more SPE's or 20% more on the GPU, the latter is far more effective at painting pretty pictures which face it, is what we actually care about. Because I dont see Cell doing any mondo physics games or something. But I do see it doing depth of field in Uncharted 2.

I think its a perfectly reasonable stance, Xenon is hardly going to win any awards for CPU performance but its held up just fine this generation, and its also much easier to "extract" that CPU power than it is with a Cell like design. If Sony had went with a simple Xenon like design and instead invested all that extra R & D and transistor budget on cramming in a consolised version of G80 or similar, Uncharted & Killzone wouldn't be the exception, they'd be the norm. Really, what has that investment in Cell bought them this generation outside of a couple of standout exclusives? Certainly not parity or superiority in multi platform titles, that's for sure.

Next generation you'll be able to offload stuff like physics to the GPU as well, further decreasing its importance even more. For me, a dual/quad core Sandy Bridge + the biggest most cutting edge GPU you can afford, would seem the best use of your budget from a purely theoretical standpoint. That way anyone can get great results, no matter how sloppy their code and if you want to push the system, investigate into new rendering algorithms and utilise heavily optimised and threaded code you can go ahead and do that, it'll just take part on the GPU rather than the CPU like it did with the PS2 and PS3. Its win, win in my book, the silicon is being put to use in all games right from the off, rather than only in a few select exclusive titles, but there's still plenty of room for pushing things with the increasingly general purpose nature of the GPU.

Put it this way, when building a PC gaming rig on a budget, where does the bulk of your money go if you want the best bang for your buck? Its of course the GPU every time, because modern games are heavily GPU bound and wth an increasingly diverse workload that can theoretically be shift to the GPU, I don't see that changing anytime soon, in fact I'd expect the exact opposite.
 
The only thing that I can think of that would keep them from doing it is BC but I was under the impression that emulating a CPU is a somewhat easy thing to do in comparison to a GPU. What would they do about VMX128 though?

and having that in mind we can safely assume that next xbox GPU would also use eDram. Perhaps enought to do "full" HD with 4xaa and at least 4 af without tiling.
 
How far can you really push a GPU when offloading more general purpose tasks from the CPU over to it?

Would a theoretical GP-GPU be quicker at GP game processing than a CPU? How would all that heavy GP stuff on a GPU affect the overall perfomance while you're trying to do graphics processing on it as well?

What are the limitations of current GP-GPU processing? Are there things that a CPU is just 100% better at than a GPU?

If a GPU is better at doing everything, why not spend your entire transistor budget on a whopping big GPGPU?
 
Status
Not open for further replies.
Back
Top