Predict: Next gen console tech (9th iteration and 10th iteration edition) [2014 - 2017]

[Megadrive1988]

IHS said that at launch Sony paid $100 for PS4 APU and $88 for 16 chips of GDDR5 [entire console costed $381], so we will probably see the repeat of that for next console.

For PS5 Sony would be wise to put everything on a single interposer, since that will also simplify the motherboard, power delivery and cooling. We will have 32GB of HBM2 in professional GPU cards in 2016/early 2017, I think consoles will match that only if release schedule calls for early deployment. If PS5 comes in late 2019 or late 2020, we should get more. As for OS ram pool, they should not add new physical chips for it. HBM is plentiful, there is space for both OS and game apps.

Agreed. Not only will there be be 32GB of HBM2 in professional GPUs in 2016 / early 2017, there should probably also be, in 2017, professional APUs with 16 Zen cores, Greenland streaming processor/GPU (Arctic Islands flagship) with HBM2.

I could see AMD having already shown more details to Sony and Microsoft, more extensive roadmap, with that APU being a model from which they could make semicustom versions to meet the desired specs for 9th gen consoles. Maybe 8 or 12 Zen cores and GPU architecture based on Arctic Islands successor.

 

[upnorthsox]

My proposal were based on the fact that you are build a $399 system. True HBM can go to 1 TB/sec and 32 GB, but that will require 4 stacks like Fury. I'm assuming costs for placing 4 stacks are more than 4X higher than just putting 1 stack down on the interposer. I wish I knew more on the economics of this.

HBM2 is planned to go to 8 high stacks so you will be able to get up to 32GB in 2 stacks. This imo is max amount of stacks you should expect on a next gen interposer. Thus, 32GBs is the max amount of ram to expect next gen. This fits nicely as I'd want to know the use case for any more than that.

I was think if it came down to it, they could just use an off the shelf DIMM(s) for the system RAM.

There's less than zero chance of any console using DIMMs.

The interesting possibility for extended storage would be 2 stacks of 3D nand on the interposer. There you could be talking about 128-256GB with ddr level access. That's enough to make an HDD optional, maybe as a usb attached drive and/or have online storage available.

Add a cloud processor (that's what I'm calling it) on the interposer to handle IO requests and external connectivity/processing and you could have a tight little system with a mobo that's smaller than 3x5

 

[fehu]

Agreed. Not only will there be be 32GB of HBM2 in professional GPUs in 2016 / early 2017, there should probably also be, in 2017, professional APUs with 16 Zen cores, Greenland streaming processor/GPU (Arctic Islands flagship) with HBM2.

I could see AMD having already shown more details to Sony and Microsoft, more extensive roadmap, with that APU being a model from which they could make semicustom versions to meet the desired specs for 9th gen consoles. Maybe 8 or 12 Zen cores and GPU architecture based on Arctic Islands successor.

That thing is server class, will cost by its own as two ps5

 

[Globalisateur]

I see everybody thinking about Zen used in next gen... the problem is with Zen AMD aren't really targeting low power / low cost, are they? Which doesn't fit well with consoles requirements.

I see more AMD releasing cheaper and less power hungry tablet family members of Zen that could be used on next gen with a high number of cores like 12 or 16.

 

[fehu]

Zen is roadmapped to scale from tablet to server and replace jaguar

 

[McHuj]

HBM2 is planned to go to 8 high stacks so you will be able to get up to 32GB in 2 stacks. This imo is max amount of stacks you should expect on a next gen interposer. Thus, 32GBs is the max amount of ram to expect next gen. This fits nicely as I'd want to know the use case for any more than that.

You sure? I've only seen this publicly for HBM2:

Indicating a max size of 8GB and 256 GB/sec per stack.

 

[RDGoodla]

As for the rendering resolution, imo 1080p native resolution + optional 4K user interface may become main stream.
1080p rendering resolution is to ensure balance between performance and image quality, and 4K UI/text can make
use of the advantage of 4K TV.

But I still have two questions about this, please help.

First, how does the upscaler work?
We know that 4K resolution is an interger multiple of 1080p. So 1080p images can be resized to 4K (each pixel becomes 2x2 pixels of the same color) without causing any blur.However "resizing" a 1080p image doesn't use the advantage of 4K TV. So most 4K TV still upscale 1080p images to 4K. Generally 4x upscaling may cause severe blur however it seems that "1080p to 4K" upscaling does't blur image very much. So how does the upscaler work when converting a 1080p image to 4K? Just using traditional method of using some smart method?


Second question is how much resource is required to use 4K UI compared with 1080p UI?

 

[Shifty Geezer]

No TV uses a dumb pixel doubling or interpolation to upscale. They all use smarter algorithms.

 

[AlNom]

Second question is how much resource is required to use 4K UI compared with 1080p UI?

Insignificant next to the power of the... oh nevermind.

Larger buffer & 2D assets (memory). If it's 3D (like Dead Space), it'll probably fall under the fillrate/shader & blending category.

 

[MrFox]

You sure? I've only seen this publicly for HBM2:

Indicating a max size of 8GB and 256 GB/sec per stack.

They could make a PS4 Slim with a single stack.

For PS5 I want 6 of those, spare no expense! Toy Story in real time, at last!

 

[mpg1]

As for the rendering resolution, imo 1080p native resolution + optional 4K user interface may become main stream.
1080p rendering resolution is to ensure balance between performance and image quality, and 4K UI/text can make
use of the advantage of 4K TV.

But I still have two questions about this, please help.

First, how does the upscaler work?
We know that 4K resolution is an interger multiple of 1080p. So 1080p images can be resized to 4K (each pixel becomes 2x2 pixels of the same color) without causing any blur.However "resizing" a 1080p image doesn't use the advantage of 4K TV. So most 4K TV still upscale 1080p images to 4K. Generally 4x upscaling may cause severe blur however it seems that "1080p to 4K" upscaling does't blur image very much. So how does the upscaler work when converting a 1080p image to 4K? Just using traditional method of using some smart method?


Second question is how much resource is required to use 4K UI compared with 1080p UI?

4k may be the standard video display output for next gen consoles but games actually being rendered at 4K has even lower odds than games being run at 1080p this gen IMO. Most games will be probably be rendered at 1440p+. Resolution isn't as important as the level of detail and how advanced the graphics engine is. SW: Battlefront at 900p on a 4k tv probably looks better overall than any 3 or 4 year old game running at native 4K.

 

[Scott_Arm]

I'm expecting a lot of clever reprojection and upscale tricks this gen that will carry over to next gen.

 

[upnorthsox]

You sure? I've only seen this publicly for HBM2:



Indicating a max size of 8GB and 256 GB/sec per stack.

Can't find a link now (so take my post as such) but the HBM2 chips are supposed to rise to a 2GB density per chip, with 8 chips per stack that's 16 GB per stack

 

[AlNom]

I'm expecting a lot of clever reprojection and upscale tricks this gen that will carry over to next gen.

Indeed. Assuming we're looking at HBM, MSAA ought to be that much more attractive & the RAM consumption could be less of an issue.

Sort of wonder where conservative rasterization will lead to. You'd almost hope the gen were long enough for devs to squeeze everything out of the GCN architecture before moving to the next.

 

[Newguy]

Looking at the SK Hynix page, "game console" is linked to HBM:

http://i.imgur.com/SHknPYj.png

https://www.skhynix.com/eng/product/dramHBM.jsp

Attempted marketing or something we don't know? Dun dun dunnnnn!

 

[MrFox]

Can't find a link now (so take my post as such) but the HBM2 chips are supposed to rise to a 2GB density per chip, with 8 chips per stack that's 16 GB per stack

First generation only supports 4 per stack (total 1GB) but AMD said they could put two stacks on top of each other for 2GB total. This is still HBM version 1.

It's journaists like ArsTechnica who incorrectly reported this as HBM2, and there was a mix up of what the capacity was per stack versus per die.

HBM2 is definitely limited to 8Gbits per die, 8 dies per stack.

 

[upnorthsox]

First generation only supports 4 per stack (total 1GB) but AMD said they could put two stacks on top of each other for 2GB total. This is still HBM version 1.

It's journaists like ArsTechnica who incorrectly reported this as HBM2, and there was a mix up of what the capacity was per stack versus per die.

HBM2 is definitely limited to 8Gbits per die, 8 dies per stack.

That would explain why I can't find a link now, thanks for the correction.

 

[Globalisateur]

Indeed. Assuming we're looking at HBM, MSAA ought to be that much more attractive & the RAM consumption could be less of an issue.

Sort of wonder where conservative rasterization will lead to. You'd almost hope the gen were long enough for devs to squeeze everything out of the GCN architecture before moving to the next.

Aren't there already console games that bypass the traditional ROPs and use GPGPU instead? For those games, ROPs would be lost APU space that could have being better used with CUs, texture units or bigger caches.

I think I read @sebbbi's post talking about it.

 

[sebbbi]

Is that comfirmed? The quad channel DDR4 alone is 102.4 GB/s bandwidth. That is 3x more than any PC chip with an integrated GPU. To use this bandwidth, AMD would need to have at least 3x bigger GPU in that APU (taking into account delta compression and bigger GPU caches). That GPU would be comparable to next gen consoles. But... they also have an additional 512 GB/s HBM bandwidth for the GPU. To fully utilize that, the integrated GPU part would need to be as big as Fiji.

And in addition to this, they have crammed 32 MB of L3 cache, two fast memory controllers, and 16 Zen cores (that should be bigger than Steamroller cores). I am a bit sceptical.

But if this is our next gen console, I am positively surprised

 

[dogen]

Aren't there already console games that bypass the traditional ROPs and use GPGPU instead? For those games, ROPs would be lost APU space that could have being better used with CUs, texture units or bigger caches.

I think I read @sebbbi's post talking about it.

They bypass ROPs by using compute shaders, but not necessarily with GPGPU(which would imply general purpose CPU type calculations) And I don't think it's possible to not use them entirely.


But anyway, conservative rasterization is a separate thing. Here's a visualization.

I'm not sure how good GCN is at it though, since I don't think it has any hardware acceleration for it(not sure what that would entail though). I guess it would have to be done with a geometry shader or something, maybe limiting how useful it is. Then again, I'm not sure what it's useful for anyway..