Predict: The Next Generation Console Tech

[Inuhanyou]

Is there any performance?

HarDHarHar

 

[MarkoIt]

Someone on neogaf is claiming that the "final" flop number is 1.3 TF (10 CUs@1Ghz).
Have you guys considered that using SRAM is reasonable, since AMD has never used eDRAM technology? Anyway, Aegis said something also about Durango not being an APU+Discrete. What would be the point of having 2 die-size, if both are really tiny (50mm^2 + 150mm^2)?

So, we may have Durango: A fully HSA APU with the following specs.

8 Jaguar-cores @1.6 Ghz
16-32 mb of shared cache.
10CUs@1 Ghz,
256bit bus
8 GB of DDR3/4
totale die-size = 250 mm^2, TDP 125-150W.

I want to give a chance to DDR4.

 

[bkilian]

Neat, this seems to link back to that leaked vision document a few months back:
http://www.engadget.com/2013/01/09/microsoft-and-samsung-demos-xbox-display-fills-room-with-images/

I'd love something like this in my TV room.

 

[XpiderMX]

Someone on neogaf is claiming that the "final" flop number is 1.3 TF (10 CUs@1Ghz).
Have you guys considered that using SRAM is reasonable, since AMD has never used eDRAM technology? Anyway, Aegis said something also about Durango not being an APU+Discrete. What would be the point of having 2 die-size, if both are really tiny (50mm^2 + 150mm^2)?

So, we may have Durango: A fully HSA APU with the following specs.

8 Jaguar-cores @1.6 Ghz
16-32 mb of shared cache.
10CUs@1 Ghz,
256bit bus
8 GB of DDR3/4
totale die-size = 250 mm^2, TDP 125-150W.

I want to give a chance to DDR4.

Aegis said no way 1+tflop, 1.3tflops look like 1+tflop.

 

[MarkoIt]

Okay, a reliable rumor is that Durango GPU has at least 3 sections of custom external hardware to help in gaming tasks. One said to be inside the GPU, and that "you can guess some of them easily" or something like that. Well, I cant, lol.

Fascinating! Special sauce seems to be very real.

Edit: One of the "sections of external hardware" could be the ESRAM I suppose.

Edit2: Could another be a tesselation add on or something?

I think it can access with high bandwidth (~ 1 TB/s) a common cache with the CPUs, allowing for fast data transfer between the 2. A beefed up tessellation unit, and the third I would go for ROPs and texture improvements.

 

[Helmore]

Someone on neogaf is claiming that the "final" flop number is 1.3 TF (10 CUs@1Ghz).
Have you guys considered that using SRAM is reasonable, since AMD has never used eDRAM technology? Anyway, Aegis said something also about Durango not being an APU+Discrete. What would be the point of having 2 die-size, if both are really tiny (50mm^2 + 150mm^2)?

So, we may have Durango: A fully HSA APU with the following specs.

8 Jaguar-cores @1.6 Ghz
16-32 mb of shared cache.
10CUs@1 Ghz,
256bit bus
8 GB of DDR3/4
totale die-size = 250 mm^2, TDP 125-150W.

I want to give a chance to DDR4.

I think that a chip like that would have a TDP lower than that. Another thing, what did you have in mind for that shared cache? As L3 cache and then those 8 Jaguar cores still had 4 MB L2 cache? 32 MB of cache would be around 90 to 120 mm² or something, right? At least if you use SRAM on 28 nm, it would be a lot smaller if it was eDRAM, but I'm not sure if GlobalFoundries or TSMC has eDRAM on 28 nm.

Are we even sure where Durango will be fabbed?

 

[MarkoIt]

I think that a chip like that would have a TDP lower than that. Another thing, what did you have in mind for that shared cache? As L3 cache and then those 8 Jaguar cores still had 4 MB L2 cache? 32 MB of cache would be around 90 to 120 mm² or something, right? At least if you use SRAM on 28 nm, it would be a lot smaller if it was eDRAM, but I'm not sure if GlobalFoundries or TSMC has eDRAM on 28 nm.

Yes, like an L3 cache. Some pages before it was said that SRAM is 6 times less dense, so your numbers may be right. This is why I think 16 MB is more likely. Altought, I'm not sure how much can help with reducing bandwidth usage.

 

[pjbliverpool]

2. A beefed up tessellation unit, and the third I would go for ROPs and texture improvements.

If AMD had these improvements available now (fundamental improvements to basic elements of the GPU) why would they be holding them back from their premium desktops GPU's?

 

[anexanhume]

I think that a chip like that would have a TDP lower than that. Another thing, what did you have in mind for that shared cache? As L3 cache and then those 8 Jaguar cores still had 4 MB L2 cache? 32 MB of cache would be around 90 to 120 mm² or something, right? At least if you use SRAM on 28 nm, it would be a lot smaller if it was eDRAM, but I'm not sure if GlobalFoundries or TSMC has eDRAM on 28 nm.

Are we even sure where Durango will be fabbed?

Agreed. I also assumed they would favor more CU's at lower frequency (800 MHz or so). I guess it gets back to what costs more, larger die that yields better or smaller die with poorer yields (now).

 

[ERP]

If AMD had these improvements available now (fundamental improvements to basic elements of the GPU) why would they be holding them back from their premium desktops GPU's?

Agreed anything new likely couldn't be exposed on a PC without significant performance penalty on existing titiles or requiring a new API to expose the functionality, or requiring developers to adopt the new feature.

Obvious changes would be things like ROPS being similar to the 360 part and be part of any "embedded" memory interface.
That could result in real "free" AA in the absence of tiling.

 

[nunomoreira]

If AMD had these improvements available now (fundamental improvements to basic elements of the GPU) why would they be holding them back from their premium desktops GPU's?

The future of games is very tessellation heavy and this is a console that needs to last quite a few years, so they need to think foward.

 

[MrFox]

If AMD had these improvements available now (fundamental improvements to basic elements of the GPU) why would they be holding them back from their premium desktops GPU's?

Which leads to my question: Why didn't AMD/Nvidia ever added an eDRAM pool to their GPU cards, is it limitation of what DirectX/OpenGL exposes which would makes it a bad idea for a PC GPU?

 

[nunomoreira]

Which leads to my question: Why didn't AMD/Nvidia ever added an eDRAM pool to their GPU cards, is it limitation of what DirectX/OpenGL exposes which would makes it a bad idea for a PC GPU?

if making drivers is hard right now imagine with edram, a bit to high a resolution and the games crashes, from what i heard the advantages are mostly saving money, a higher bandwidth is preferable.
they would probably be better of using the same die area to make another 128bit bus then edram, though is not that simple

 

[Proelite]

Someone on neogaf is claiming that the "final" flop number is 1.3 TF (10 CUs@1Ghz).
Have you guys considered that using SRAM is reasonable, since AMD has never used eDRAM technology? Anyway, Aegis said something also about Durango not being an APU+Discrete. What would be the point of having 2 die-size, if both are really tiny (50mm^2 + 150mm^2)?

So, we may have Durango: A fully HSA APU with the following specs.

8 Jaguar-cores @1.6 Ghz
16-32 mb of shared cache.
10CUs@1 Ghz,
256bit bus
8 GB of DDR3/4
totale die-size = 250 mm^2, TDP 125-150W.

I want to give a chance to DDR4.

1 ghz cus. That seems fishie.

 

[Ruskie]

It does sound alot like leaked roadmap, except for more RAM and even crappier GPU.

 

[Osamar]

The problem is that modern GPU's in the PC space at least are optimized to run last years games, and they all have similar memory configurations. It's still a bad way to compare even in that space IMO, but it's difficult to separate all of the factors.

I just think that when we're talking about consoles FLOPS (assuming they are in the same ballpark) is not going to be a good indicator of system performance, there will be a much bigger difference IMO in memory configuration, and how well utilized those ALU's will be.

So, in the end, the best is to have a balanced system without underutilized parts.
It is not easy to design a console with four of five years in advance.

 

[MarkoIt]

If AMD had these improvements available now (fundamental improvements to basic elements of the GPU) why would they be holding them back from their premium desktops GPU's?

We haven't see a new GPU from AMD since last January, so we don't really know what will be inside the next one. As someone pointed out, the improvements could lead to some performance penalty with current titles.

 

[Proelite]

It does sound alot like leaked roadmap, except for more RAM and even crappier GPU.

NVM.

 

[Proelite]

Someone on neogaf is claiming that the "final" flop number is 1.3 TF (10 CUs@1Ghz).
Have you guys considered that using SRAM is reasonable, since AMD has never used eDRAM technology? Anyway, Aegis said something also about Durango not being an APU+Discrete. What would be the point of having 2 die-size, if both are really tiny (50mm^2 + 150mm^2)?

So, we may have Durango: A fully HSA APU with the following specs.

8 Jaguar-cores @1.6 Ghz
16-32 mb of shared cache.
10CUs@1 Ghz,
256bit bus
8 GB of DDR3/4
totale die-size = 250 mm^2, TDP 125-150W.

I want to give a chance to DDR4.

Very accurate but very incomplete specs. Not sure about the 1.0 ghz cus, that's very hot for a console.

 

[kots]

Okay, a reliable rumor is that Durango GPU has at least 3 sections of custom external hardware to help in gaming tasks. One said to be inside the GPU, and that "you can guess some of them easily" or something like that. Well, I cant, lol.

I don't get it ... why go all this trouble to assist an abomination of a gpu and don't just get a better gpu in the first place .