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

[Silent_Buddha]

Hell, if Sony do BC like Microsoft has, I'll be buying a PlayStation ASAP. I've got a lot of PS1 and PS2 games I'd love to play again, especially if they got enhanced via the BC process. Can do it with emulators on PC, but results can be hit or miss at times.

Regards,
SB

 

[Gemüsepizza]

probably, but if you have the gpu power to ramp up and they want to push the graphics envelope, they'll naturally push it to 30fps.
I'm mean look at it this way specifically at the GPU.
60 FPS @ 12 TF is = 30 FPS @ 6TF. They do very close to the same amount of work if they are at the same resolution. If PS5 is 60fps mandated, and is 12 TF, it cannot make a generational graphical difference from 1X. You'd barely notice the difference. At 30fps is a different story.

Well, that is if Sony will use Xbox One X settings. The other possibility is, that they will just use PS4 Pro settings, which would probably look like this:

30 FPS @ 4.2 TF = 60 FPS @ ~8.4 TF

And some of the limitations of PS4 Pro games were also a result of the relatively low memory bandwidth. Afair this was the reason why Uncharted 4 is "only" 1440p on PS4 Pro. With 8.4 TF, higher bandwidth and other architectural improvements, I think they could get 4K checkerboarding and 60fps for a game like this.

But obviously there won't be a 8 TF GPU in PS5, probably something like ~12.6 TF. So that means Naughty Dog will have over 4 TF which they can use for better lighting. And if you think about it, that's still a massive amount of GPU power, basically one PS4 Pro. To summarize:

Naughty Dog PS4 / PS4 Pro game:

1080p (2,073,600 pixels) @ 30 fps (PS4)

1440p (3,686,400 pixels) @ 30 fps (PS4 Pro)

Naughty Dog PS5 game:

4K CB (4,147,200 pixels) @ 60fps + 4 TF for improved lighting + 4K textures

This is for a 30fps game. But of course there are already quite a few 60fps games on PS4 like Call of Duty, Wolfenstein and Star Wars: Battlefront (which are among the most popular games). Those will have an additional 8 TF for improvements + 4K textures on PS5. To me this seems like a pretty good generational leap.

And we should not forget that the vast majority of PS4 owners just have a base PS4. This will be a massive jump for them, going from base PS4 to PS5, and they don't even need a 4K TV to see this.

I'm really sorry but 30fps is here to stay. Games that want to be 60fps are 60fps.

I don't really agree with this. We now have more 60fps games than during last-gen. Titles like FIFA, Resident Evil, Metal Gear Solid and Battlefield have been 30 fps during last gen, and they are 60fps in current-gen. We will have much more CPU power for next-gen consoles, which makes 60fps more attractive to devs, because there are less compromisses needed.

Also, many of those 30fps games during this gen are among the most expensive titles. Games like Assassin's Creed or GTA have over 1,000 people working on them. Is it really such a good idea to use all of next-gen's power to only increase graphics? I think there is an economical incentive here to slow down this development. And 60fps is an easy way to do this, while still getting a big improvement for your games. Also, if you look at GTA, it's multiplayer mode is incredibly popular. And there are rumors that multiplayer is coming back to Assassin's Creed. That's why I think that 60fps makes a lot of sense even for such games.

All these wonderful technologies that have been developed to offload work from the CPU to the GPU. If you use them, then you only need a bigger GPU.

Going by this logic, then why not just use a 7nm Jaguar for next-gen?

 

[dten]

If we assume "30 FPS @ 4.2 TF = 60 FPS @ ~8.4 TF" as stated above, we can't just conclude 8.4 TF with more memory bandwidth will get you 4K CB. That's an assumption. Even if the 1440 limit was due to memory constraints, we don't know how much headroom they have by increasing bandwidth. 8.4 TF in of itself needs more bandwidth. Plus the game does run at 60 fps/1080p in multiplayer. Could it have been 60 fps/1080p single player? Maybe. My point is that you can't just assume these kinds of linear calculations. There are many factors in game development. Anyway, if PS5 uses 4K CB very often, there will likely be a lot of "faux 4K" talk next gen. And even then, that's assuming 12 TF.

 

[Gemüsepizza]

If we assume "30 FPS @ 4.2 TF = 60 FPS @ ~8.4 TF" as stated above, we can't just conclude 8.4 TF with more memory bandwidth will get you 4K CB.

Well, I wrote "with [...] higher bandwidth and other architectural improvements". Of course I'm speculating a bit here. 1440p and 4K CB are relatively close when you look at the pixel count, so I think this could be possible. If not, they could still use something like 3616x2034 CB, which is closer to 1440p. My main point was that 60fps could be possible while still offering much better lighting and nice 4K textures. Compared to base PS4, this would be a very big jump imo.

Anyway, if PS5 uses 4K CB very often, there will likely be a lot of "faux 4K" talk next gen. And even then, that's assuming 12 TF.

Not so sure about that. I think the average consumer doesn't really care about "native 4K". Even enthusiasts think that 4K CB and similar techniques are an elegant way to save performance. And I believe if they can pull off 4K CB @ 60fps, most will be ecstatic.

 

[DieH@rd]

When can we expect 7-10nm shrinks of PS4 Slims and PS4 Pros APUs?

How much would PS4 SuperSlim [7nm APU] cost if it had no BD Drive, empty HDD tray and with 64-128gb of slow-ish flash memory?

 

[AlNom]

When can we expect 7-10nm shrinks

10nmFF is more-or-less a node for Apple & other mobile companies as power consumption is of utmost concern there, and they can afford the premiums while also entertaining ~100mm^2 designs to begin with. While it's no slouch (in terms of characteristics) compared to how 20nm planar node also ended up being mobile-only, my impression is that the cost of designing for it is high enough that folks with higher end requirements are skipping it in favour of pouring resources into the following node.

Late 2019 is probably the earliest? Would be cute if they all boosted clocks though.

 

[MrFox]

Would there be a good reason to drop compatibility with some previous vector extensions? Suppose that extension requires a significant amount of silicon, while new better vector extensions are too different to make the old one compatible without adding more circuitry. Otherwise the instruction set would accumulate a lot of bloat if any extension ever made needs to be supported forever. I mean there has to be a point in time where some extensions were a bad idea and get superceded by better ideas.

For a console which is a customized CPU anyway, maybe they might have reasons to cut out some fat.

 

[McHuj]

10nmFF is more-or-less a node for Apple & other mobile companies as power consumption is of utmost concern there, and they can afford the premiums while also entertaining ~100mm^2 designs to begin with.

I thought that was the case as well until Qualcomm released their server chip on a 10nm process. It’s giant and has a TPD of 120w. Whether others will is to be seen, but 10nm can be used for non-noble.

 

[shiznit]

I thought that was the case as well until Qualcomm released their server chip on a 10nm process. It’s giant and has a TPD of 120w. Whether others will is to be seen, but 10nm can be used for non-noble.

Probably can, but is it cheaper? Apple and Qualcomm have much higher margins to cover the 10nm premium (and potentially lower yields vs a mature 16nm).

 

[Blanco]

I thought that was the case as well until Qualcomm released their server chip on a 10nm process. It’s giant and has a TPD of 120w. Whether others will is to be seen, but 10nm can be used for non-noble.

If they stick with AMD, it’s likely they’re not going to be on 10nm since it appears that AMD is skipping 10nm entirely.

I’m on mobile so I won’t look up all the sources but here’s one http://www.tomshardware.com/news/amd-zen-10nm-7nm-intel,32619.html

 

[3dilettante]

Would there be a good reason to drop compatibility with some previous vector extensions? Suppose that extension requires a significant amount of silicon, while new better vector extensions are too different to make the old one compatible without adding more circuitry. Otherwise the instruction set would accumulate a lot of bloat if any extension ever made needs to be supported forever. I mean there has to be a point in time where some extensions were a bad idea and get superceded by better ideas.

For a console which is a customized CPU anyway, maybe they might have reasons to cut out some fat.

Support can be dropped though there's a significant amount of inertia that has to be overcome, and it would generally come at some kind of significant inflection point for the overall ISA and software direction. Instruction sets will consume more encoding space as time goes on, but whether an idea is good doesn't usually matter as much as whether it is used by software that the architecture is trying to remain compatible with.

AMD dropped 3DNow with Bulldozer and Bobcat.
Zen dropped extensions introduced by Bulldozer such as its XOP and profiling instructions.
x86-64 mode overrides a set of older instructions.

Validating or working around cramped encoding space can add complexity (more validation, hardware, modes) and increase the average length of instructions if new prefixes are added.
With microcode or reduced-performance implementations, the actual area cost in a chip can be pretty small for disused instructions. The front end abstracts the internals of the chip from the ISA, so an actual implementation generally doesn't physically reflect the level of complexity of the instructions.
Just dropping instructions a console maker could ignore would seem to be spending money for design, validation, and maintenance for little gain. The decoder might be a little smaller, but it's not like the encoding for everything else would change. Actually taking advantage of the space, either to make code more compact/efficient or adding functions is creating a progressively more costly core, depending on how much of the expensive mainline development cannot be reused.
Such changes would themselves be added to the bucket of support and backwards compatibility worries in the next generation.

 

[AlNom]

I was mulling over the size of Vega vs Fury again - so we've heard that part of the transistor bloat is from designing for increased clocks. How about the L1 amount as well? They mention that the CUs are now setup in groups of three now instead of up to 4. For Fury, they had 4 sets of L1 for the 4 quads of CUs per SE (4+4+4+4 = 16CUs). For Vega, they'd have 6 sets of L1 (3+3+3+3+2+2 = 16CUs) per SE.

Soo.... +50% L1 in comparing between those CU configs?

Might have to factor that in for the approximating sizes next time too.

 

[DieH@rd]

https://twitter.com/x/status/937931138152120320

A very interesting notion - PS5 Portable.

Of course, I'm not delusional that Sony will actually attempt to miniaturize PS5 hardware... but what about this.... 10 years after launch of PS4, they release portable version of PS4 it and call it PS5 Portable [which is good name for marketing purposes]. If Sony can push devs to create steady stream of "PS4 tier" games in the future that would be a double win [offering fresh content for both gazilion of PS4 users and new PS5 Portable owners], and if they cannot, they still have a large PS4 library that they can extract online sales on this new portable platform, which is another win.

[ofc, most likely PS5 Portable listing by GS is just a analysts wet dream]

 

[AlNom]

Wouldn't that be more confusing marketing-wise? PS5Portable, but actually a PS4 portable in terms of software library?

PS4 Go

(Forego!)

 

[Magnum_Force]

My ideal for next gen would be a low tdp base unit which would double as a portable, but could also dock into a gfx card caddy.

You could do it now with COTS equipment:

Base unit (non mobile): Ryzen 2500u with 8gig ram, no battery or screen.

Base unit (mobile): same as above but with additional battery and screen. The screen and battery would simply slot into the non mobile base unit making it mobile.

Dock: Vega 56 with psu over a hot pluggable pcie 3.0 x8 connection.

That's over 10 Tflops with ROV, CR, etc goodness. Portable makes the Switch look like an Antique and can play all the same games as the current gen with very minimal changes.

XFR means that when the base unit is in a non mobile configuration it could clock the 2500u sky high as long as the cooling was robust enough.

 

[DieH@rd]

Wouldn't that be more confusing marketing-wise? PS5Portable, but actually a PS4 portable in terms of software library?

PS4 Go

(Forego!)

Well, they could also go with PS5 Portable that is built on same [but downlcocked] CPU like full PS5 and much weaker GPU [less and downclocked CUs]. Small 20W APU+RAM package that mimicks big console with its core components. But IMO, having PS4 compatibility is a must. [and brute-forcing 720p to reduce rendering load]

 

[mpg1]

I don't know why they'd call a portable PS4...PS5. Just call it PlayStation Go, PS GO, etc..

 

[3dilettante]

I was mulling over the size of Vega vs Fury again - so we've heard that part of the transistor bloat is from designing for increased clocks. How about the L1 amount as well? They mention that the CUs are now setup in groups of three now instead of up to 4. For Fury, they had 4 sets of L1 for the 4 quads of CUs per SE (4+4+4+4 = 16CUs). For Vega, they'd have 6 sets of L1 (3+3+3+3+2+2 = 16CUs) per SE.

Soo.... +50% L1 in comparing between those CU configs?

Might have to factor that in for the approximating sizes next time too.

Per earlier GCN slides, the L1 instruction cache is 32KB. 32786 bytes x 8 bits x 6 transistors bit = 1572864.
If there was a move to 8T cells or additional ports it would go from ~1.5 million transistors to ~2 million in the arrays, but I'm not sure if higher porting or stability/RAS considerations factor into a GPU L1.
There would be more shared front ends with unknown transistor counts. 24 L1 caches nets 48M transistors.
The scalar data cache would be .75 to 1 million, although those could change how many CUs shared them across implementations.
Logic for accessing them would add to the count, but it seems like these caches are operating several orders of magnitude down from the ~4 billion transistor increase that was supposedly dominated by wire delay and clock optimization.

 

[Magnum_Force]

Just thinking about this, and Xbox portable might be an opportunity for Microsoft to move to an arena where PlayStation is absent. The Ryzen 2500u/ 2700u could be used in a portable "xbox one" which would from the portable component of a next gen system, using the gpu dock example I gave above. Benefits of this would be much lower R&D for Microsoft, as well as a streamlined manufacturing as AMD and Microsoft would be using the same chips (maybe even the same fully manufactured Gfx cards)

 

[AlNom]

Well, it might be a neat scenario as a Surface off-shoot.

"Surface Xbox/Surface Gamer"