Let's discuss the ways the future console architectures could potentially improve in efficiency

[Remij]

This thread can be about any and all potential architectural optimizations you can think of, however I'd like to first try to focus the discussion around IO and the flow of data since there's lots of discussion on this forum about the differences between PC and console architectures such as the PS5 and it's impressive I/O capabilities. Nobody is going to argue that consoles are more efficient than PCs.. they simply are. They're carefully crafted devices which are balanced to provide the best performance for the best value. From the customized components, to the unified memory.. to dedicated silicon meant to ensure all the components can work as efficiently as possible... they also have streamlined OSes/filesystems as well as developer tools. They are meant for gaming.

So today's consoles already exhibit some distinct efficiency gains. They have now have solid state storage, along with unified memory which both the CPU and GPU can access, they have dedicated decompression blocks to offload the CPU in ensuring data gets to memory as efficiently as possible. So my question really comes down to how that flow of data could be improved for future generations? Obviously bandwidths will increase, caches will increase, but my interest in this question comes down to the actual path of data. Is there any obvious ways in which the architecture/physical layout could change to become even more efficient than what it already is? HBM right on the SoC?

As for a different point... what about dedicated system memory and CPU cores for OS/UI function? No contention for resources... nothing slowing down the responsiveness of the device either while gaming or using the dash.

What kinds of tech do you think the next generation of consoles will attempt to employ?

 

[iroboto]

Consoles don’t need to employ the same type of cores used by GPU vendors for tensor processing. It’s important to note that the cores provided by GPUs today are used for training and running.

Consoles only need to run. So I’m not sure how much that could help in getting more from the silicon as a result, but it’s got to be flexible enough to support all the models going forward which is a little tougher.

 

[Nesh]

AI/ML is going to play a HUGE role in visuals. I expect to see exponential difference if things go where I think they are going.
It is going to automate a lot of processes and auto-correct/enhance image to the point where less resources and power will be giving much much better results.

 

[davis.anthony]

They need to have huge onboard cache or move to using a large EDRAM.

Something like PS2's GS but repurposed for tracing rays and data kept in EDRAM or huge cache.

 

[Karamazov]

What would be needed for SSDs to approach the speed of ram ?

 

[MCharta]

They need to have huge onboard cache or move to using a large EDRAM.

Something like PS2's GS but repurposed for tracing rays and data kept in EDRAM or huge cache.

Isn’t this the first thing that is removed/reduced on x86 to reduce costs ?

 

[Shifty Geezer]

What would be needed for SSDs to approach the speed of ram ?

A completely different hardware. Flash won't do it, which is where 3D Xpoint came in as an interim tech. It seems there's no development of interims any more though, so you're looking at existing tech in novel configurations, like a large on chip stacked DRAM caches.

 

[techuse]

Maybe its time for GPUs to move away from pixels processed in quads and finally updating FF rasterizers that can efficiently process much smaller triangles. Also x86 should be ditched in consoles. All the baggage is completely unnecessary and leaves performance on the table. Backwards compatibility isn’t worth it. When are we finally going to get the long promised 3D stacked ram?

 

[BRiT]

Backwards compatibility isn’t worth it.

Strong disagree. I am never leaving behind my existing and always growing digital library of games.

 

[Inuhanyou]

Maybe its time for GPUs to move away from pixels processed in quads and finally updating FF rasterizers that can efficiently process much smaller triangles. Also x86 should be ditched in consoles. All the baggage is completely unnecessary and leaves performance off the table. Backwards compatibility isn’t worth it. When are we finally going to get the long promised 3D stacked ram?

This is the exact opposite of the thinking of the console vendors who depend on their legacy libraries to various degrees. It's becoming more and more important rather than less so.

Nintendo got benefit of a doubt cause wii u was trash. But switch 2 will absolutely fail if it doesn't allow for users to keep all the games they have invested on switch in the next hardware.

Especially now that Nintendo has reached the plateau of their hardware strategy. There is no barrier to bc incoming with some new gimmick from Nintendo. They have decided that switch is what they have been leading up to all along

 

[trinibwoy]

Maybe its time for GPUs to move away from pixels processed in quads and finally updating FF rasterizers that can efficiently process much smaller triangles.

That would be nice. They don’t need to abandon quads yet though. Quad granularity hardware rasterization would still be a huge step up from todays rasterizers that are optimized for triangles 16-32 pixels wide.

 

[Seanspeed]

I'm thinking that by the time next gen consoles are meant to be released(I think 2028 at earliest, and ideally longer) advanced packaging is definitely going to be quite a bit farther along. I think one of the most obvious things they can do, and which we're already seemingly going to see with CDNA3, is a large L3/IO chip stacked underneath the compute chip. This has two huge advantages:

1) It allows you to take off the area-inefficient L3 from the compute chip, letting you pack in way more actual compute into the same space.

and

2) Allows you to pack in a ton of L3 since a lower layer doesn't need to worry about 'covering cores' or anything so temperature sensitive. With a dense structure, you could pack in like 192MB of L3 or more to share with the CPU and GPU.

This obviously comes with the extra cost of two dies + packaging(though you can use older process for L3 die to reduce the pain some), but honestly, the limits of what you can do with modestly-sized monolithic SoC's, especially given the lack of scaling of SRAM and all that, are going to face heavy diminishing returns in the future. And yes, prices will reflect the extra costs. $600 minimum. Just gonna have to be that way.

 

[Inuhanyou]

I am probably the only one who thinks this, but I really wish all these cool technologies were fully optimized and researched before launching. I don't need a PS5 pro, I need a PS6 that has as many of the bells and whistles as possible put in. Put all the budget going to a pro to that and use the rest to make the current PS5 super power efficient and small and quiet.

Is it impossible in the next 5 years to ask for a sleek ps5 form factor?!

 

[techuse]

Strong disagree. I am never leaving behind my existing and always growing digital library of games.

At what % of performance increase would this change for you?

 

[BRiT]

At what % of performance increase would this change for you?

Doesn't matter to me. A sustainable digital library that carries forward is what I am in the market for and is the only thing I will purchase going forward. They can change the hardware as much as they want so long as it plays my library.

 

[Seanspeed]

Is it impossible in the next 5 years to ask for a sleek ps5 form factor?!

A proper PS5 Slim makes a lot more sense to me in the near future than a PS5 Pro.

There's notable cost savings that could be made, and if they just keep the price the same(which they will if sales keep up), they can keep all those cost savings to themselves.

But I'm with you - I cant bear the idea of buying the PS5 as it is currently. It's a monstrosity and I cant even imagine where I'd put it. I need a smaller revision - one that preferably is much less visually obtrusive.

 

[Inuhanyou]

A proper PS5 Slim makes a lot more sense to me in the near future than a PS5 Pro.

There's notable cost savings that could be made, and if they just keep the price the same(which they will if sales keep up), they can keep all those cost savings to themselves.

But I'm with you - I cant bear the idea of buying the PS5 as it is currently. It's a monstrosity and I cant even imagine where I'd put it. I need a smaller revision - one that preferably is much less visually obtrusive.

I'm saving up for a series S right now because it will be my microsoft exclusives machine along with my library of og Xbox and 360 games, since I've decided to replace my 360 with a newer machine.

But I have always planned for PS5 to be my main machine for Sony games and third parties, so I'm actually intentionally waiting for PS5 to get a revision before I buy one.

But I hope it's soon cause I am just about fed up with my jet engine launch PS4. It's served me well over the many years but I want to be future proofed for the next decade.

I am actually very impressed with how compact and quiet I've heard the series S is relative to its performance.

 

[fehu]

Is it impossible in the next 5 years to ask for a sleek ps5 form factor?!

At this rate, AMD in some year will have a PS5 comparable apu sold in a notebook

 

[fehu]

A sustainable digital library that carries forward

Plus, on new hardware it can perform even better

 

[milk]

My guesses of possible avenues of development:

Drastic increases in mem badwith, perhaps through massive caches, or HBM style stacked memory, or a huge package like apples M1.

Writable SSD for games with large modifiable persistant worlds.

Separate SOC for OS and system apps, woth a separate small and cheap pool of memory and access paths and time-slice reservations into storage so running games are 100% unnafected by background downloads, instalation, patching, file transfers, etc.

More considerable speed increase in SSD top speed for short bursts, to be used exclusively for game switching or for level load. Sustained speed can increase moderatly from where it is with current new-gen.

HW rasterization will probably be completely overhauled. There is no way all previous assumptions arent being re-considered by HW vendors after nanite went with purely compute-based SW raster and it was faster (most of the time)

Perhaps other fixed function units and parts of the arctecture will be further generalized too. Texture units? Texture compression? Hi-Z? Blending?