Predict: The Next Generation Console Tech

[Heinrich4]

If MS do decide to go with eDRAM again.. Is it possible to see a 256MB implementation? IIRC, the DICE presentation at GDC 2011 had a 1080p g+z buffer 4xMSAA at 158MB. I believe 1080p is what people really want to see next gen, but I remember reading a post (I think it was Al's) saying that devs should just sort of stick to 720p instead of making a huge memory sacrifice for 1080p and then sacrificing even more memory that could have been used to make full resolution transparency's.

So I was wondering if 256MB is an option.... Or is it a bit to far-fetched?

It would be great have a console with 256MB of eDRAM, but unfortunately it is virtually impossible in cost and die size:

Link talk about SRAM
http://ww.realworldtech.com/beta/forums/index.cfm?action=detail&id=119583&threadid=119523&roomid=2

2.9Billion transistors for 364MB at 22nm showed by intell recently:
http://www.smallformfactors.com/articles/id/?4331

eDRAM(1/4 SRAM area) vs SRAM:
http://www2.renesas.com/process/en/edramadvantages.html

 

[AlNom]

They are but they consume a hell lot, no?

nVidia parts consume a lot because of the scalar setup. There's a reason why the raw throughput of AMD is so much higher. It's just that in practise, the scalar setup ends up being much more efficient for PC - devs aren't necessarily targeting the VLIW4/5 setup so it sort of balances out in multi-platform. AMD's architecture is much denser.

It would still require plenty of bandwidth to make a proper use of all this compute power, no (honest question)?

ALU ops don't use much bandwidth (miniscule). Texture ops do. Anyways, the L1/L2 caches are significantly better and more plentiful than before, so bandwidth requirements (to main memory) have been alleviated to a degree.

 

[Shifty Geezer]

Sorry, but where is this coming from?

Eurogamer.

 

[AlNom]

Microsoft have apparently got another revision in store for us with 3D support. Is there any chance they'll copy the NES 6 internals and give us an intermediate upgrade?

Eurogamer.

This new change is set to occur through a firmware update. While it is unclear exactly how Microsoft will enable full 3D, there is more than enough bandwidth to sustain it, even on HDMI 1.1, let alone the 1.2 the Xbox 360 uses.

mmm....

I suppose it just makes life easier for multiplatform. I'm not sure people will notice the difference in practise aside from the games that render both eye-views at full res i.e. 1280x720 packed into 640x720, but then again there are games that do half-res anyway for true parallax. For 2D+depth cases (typically full res per eye), having 1280x1470 would mean skipping the downscale (packing)-> upscale.

 

[liolio]

nVidia parts consume a lot because of the scalar setup. There's a reason why the raw throughput of AMD is so much higher. It's just that in practise, the scalar setup ends up being much more efficient for PC - devs aren't necessarily targeting the VLIW4/5 setup so it sort of balances out in multi-platform. AMD's architecture is much denser.

Indeed I remember that Nvidia use more logic for a given SIMD width but achieve better hardware utilization. I also remember that Nvidia ALUs run at twice the frequency than the rest of the GPC (and there is a third frequency domain for ROPs and L2 right?).

ALU ops don't use much bandwidth (miniscule). Texture ops do. Anyways, the L1/L2 caches are significantly better and more plentiful than before, so bandwidth requirements (to main memory) have been alleviated to a degree.

Honest answer
I assumed that that kind of power would not be only use by pixel/vertex/geometry/etc shaders, but also for compute and post processing matters. All this would be done as texture access and I assumed that you may want to avoid to be texture fetch limited as it happens on the 360 (on the good side as Sebbbi or Joker454 you can consider it as free ALUS cycles). But I guess it depends on the kind (and thus as you exdplained the amount) of RAM manufacturers plan to use, DDR3 or GDDR5, I guess it could be an issue with DDR3 especially if EDram is not involved.


Sweet dreaming a bit, I hope PC sales volumes will go down a bit more and that Intel would be willing to sell one of its chip for cheap for the sake of keeping their fabs busy :=)
I want 1 (or 2) Haswell core + 16 larrabee V3 cores + some L3, one single chip, a culv part @22nm running @ 2.2 or 2.3 GHz all this backed by 4GB of RAM and 16GB of on board flash memory.
EDIT
For the sake of dreaming properly Intel can fit 16 element on a ring bus, so Let move to 2 Haswell cores 14 "larrabee next" cores.

 

[Butta]

This post is likely better placed here than in the general thread...

--------------------------------------------------------------------------------

Just wondering if it would be possible for Nintendo to include 1GB of GDDR3 for the CPU and 1GB of GDDR5 for the GPU. This would resolve any issues with CPU latency attributed to to GDDR5 and low bandwidth for GPU attributed to GDDR3.

I hadn't seen it suggested most likely because of the cost split pools. But it seems that getting a GPU with 1GB of GDDR5 would be far more affordable than 2GB. Also, the once cheaper/faster memory is available, I imagine that they could combine both into a single pool?

 

[AlNom]

DDR3 would be a better fit for the CPU (or graphics if it comes to that). What you're also suggesting does imply more complexity with a second memory controller - this would make it highly unlikely to merge the two pools later down the road as there are compatibility concerns. It's probably safe to say that devs prefer just the single memory pool.

 

[liolio]

This post is likely better placed here than in the general thread...

--------------------------------------------------------------------------------

Just wondering if it would be possible for Nintendo to include 1GB of GDDR3 for the CPU and 1GB of GDDR5 for the GPU. This would resolve any issues with CPU latency attributed to to GDDR5 and low bandwidth for GPU attributed to GDDR3.

I hadn't seen it suggested most likely because of the cost split pools. But it seems that getting a GPU with 1GB of GDDR5 would be far more affordable than 2GB. Also, the once cheaper/faster memory is available, I imagine that they could combine both into a single pool?

Well if I were to go with two seperate pool of RAM I would favor 512MB for the VRAM and 2GB of DDR3 ram as the main RAM and it should cost almost the same.

 

[Heinrich4]

Sweet dreaming a bit, I hope PC sales volumes will go down a bit more and that Intel would be willing to sell one of its chip for cheap for the sake of keeping their fabs busy :=)
I want 1 (or 2) Haswell core + 16 larrabee V3 cores + some L3, one single chip, a culv part @22nm running @ 2.2 or 2.3 GHz all this backed by 4GB of RAM and 16GB of on board flash memory.
EDIT
For the sake of dreaming properly Intel can fit 16 element on a ring bus, so Let move to 2 Haswell cores 14 "larrabee next" cores.

In your (our..if you put 8GB RAM) dream could be this cpgpu(for 2014,22nm/14nm customized for consoles etc)?

http://www.thegurureview.net/tag/opencl
http://www.techdelver.com/latest-an...ure-computing-intel-32-core-processorcpu.php/

Im still thinking about early posts....counting evolution software too...I dont know,but kepler*(4 to 5Tflops) is coming this year (Maxwell 15 Tflops in 2013!) and i think for next gen consoles at 1Tflops for 5/6 year cycle could be very disapointing for gamers.

* http://www.neogaf.com/forum/showpost.php?p=27728980&postcount=45

 

[AlNom]

Intel is going to be a step or two ahead of everyone, so I wouldn't count on those process nodes by 2014 unless the consoles were using their h/w.

 

[3dcgi]

Is this even possible?

Not yet.

 

[Heinrich4]

Not yet.

And when this is possible? Under 22 nm process? In 2012/2013?

 

[3dcgi]

And when this is possible? Under 22 nm process? In 2012/2013?

I don't know when it will be possible.

 

[sebbbi]

nVidia parts consume a lot because of the scalar setup. There's a reason why the raw throughput of AMD is so much higher. It's just that in practise, the scalar setup ends up being much more efficient for PC - devs aren't necessarily targeting the VLIW4/5 setup so it sort of balances out in multi-platform. AMD's architecture is much denser.

PC shader code tends to be pretty inefficient, since you can't target just one graphics chip. You have to make something that doesn't misbehave on any chip. Basically you do not know exactly how many registers you want to use, because you do not know how that affects the thread counts (different architectures have different space for registers). And you do not know the exact ALU/TEX ratio you are targeting. Vectorizing the code can be slightly slower in some cases for NVIDIA architectures (if you need extra instructions to get things properly packed in a vector for example), but for ATI architectures you get huge gains.

We console developers are used to spending even a whole day in getting rid of the last one/two bottleneck ALU microcode instructions in a performance critical shader. VLIW seems be much better choice for future consoles, as developers will be pushing the utilization near the max in just a few years.

 

[RudeCurve]

I could see the next Xbox going with more eDRAM put on a separate die initally again...

 

[Dr. Nick]

I could see the next Xbox going with more eDRAM put on a separate die initally again...

I would have thought the use of deferred rendering in all of the high end engines (BF3, Crysis, Killzone 3 etc, etc) and the development of MLAA/DLAA and other post process AA would serve as an argument against using eDram again. Anyone want to give a detailed explanation on why Microsoft should use eDram again next time.

 

[RudeCurve]

I would have thought the use of deferred rendering in all of the high end engines (BF3, Crysis, Killzone 3 etc, etc) and the development of MLAA/DLAA and other post process AA would serve as an argument against using eDram again. Anyone want to give a detailed explanation on why Microsoft should use eDram again next time.

http://forum.beyond3d.com/showthread.php?t=60118

http://forum.beyond3d.com/showpost.php?p=1550862&postcount=5799

64MB of eDRAM should be sufficient for 1080p with 4X FSAA...also bakcwards compatibility would be 100% perfect.

 

[hoho]

64MB of eDRAM should be sufficient for 1080p with 4X FSAA...also bakcwards compatibility would be 100% perfect.

That depends how many and how "fat" buffers you want to use. 1080p with 4x MSAA would give you 8 bytes per sample with 64MB eDRAM.

Like you can read from those threads eDRAM isn't bad on it's own, just the implementation in XB wasn't exactly great. Had they enabled more flexible use for it it would have been significantly better.

 

[AlNom]

Always in cycles. http://forum.beyond3d.com/showthread.php?t=48155

Also rewinding 3 years ago...

I was actually just talking about GPU+EDRAM. IMO the only reason it was separate on 360 (and thus had limitations compared to PS2) was that there was no fab able to combine them cost effectively during the design stage.

 

[Dr. Nick]

What I said before falls more inline with my belief that Microsoft and Sony won't be willing to lose as much money as they did this generation on hardware. My question wasn't about that fact that you can do deferred rendering on the 360 and MLAA on the 360 it's more about why should they include eDram when you don't need it to provide AA and when deferred rendering seems to be the way forward in game development removing some of the advantages of eDram. Is being able to do a great deal of alpha blending enough of a reason to have it?

I know that the reasons to include eDram was used in part because of forward rendering. When the biggest names in multiplatform graphical achievement are using deferred rendering it seems less necessary to include the eDram. So back to my question of Why should Microsoft include eDram next gen? Sebbi gave a good answer on the memory foot print thing but I would just point to Sony's system and the use of deferred rendering there and things seem to be fine without eDram. I'm not trying to cause a console argument I'm just being a penny pincher but I'm sure Microsoft and Sony have penny pinchers of their own.

On a side note I just wanted add that I think Sony might even have AMD and PowerVR in the running for their next console.