Predict: Next gen console tech (10th generation edition) [2028+]

[cheapchips]

Sony: Mobile PS5/PS5 Pro Switch, essentially. Keep backwards compatibility for both implicit for a long while, PS5 base BC in handheld and PS5 Pro BC in docked. People want handheld devices, and care less and less about each console generations "upgrades". Sure you can have upgrades for the future, 18-24gb of ram, a GPU with high end matrix multiplication and formats, maybe even a dedicated high end NPU. Note: on current devices these are for constant low end AI calculations while GPU is for intensive workloads. However NPUs are more power efficient here, and for a mobile console a large one that can do constant intensive work might make sense power wise.

Plus you get to take the controllers off the side and turn the screen/processing into a VR headset for cheap. Maybe Apple Vision comes along and makes AR/VR super popular in a year or two (2025, 26?). Sure this makes the base unit more expensive, you need a 4-5k screen and extra cameras and watnot. But this way you can sell a VR kit not for $549 but $199, way cheaper. And at $699 total you're probably selling a high end AR/VR headset as cheap or cheaper than anyone else is.

Portable PS5 seems like a topic of it's own really. A shrunk PS5 SoC is a bad chip for such a device. They need to up their BC game!

Portable devices slotting into a headstrap is a terrible VR experience. The weight distribution is all wrong. It's a step backwards even against the bargain basement efforts they've offered so far. By the time we get to 2028+, thin lenses and micro oled should be cheap enough that everyone else is offering thin and light devices. Sony can't come along with some that looks like this:



They'd be better off their portable PS5 being a stand alone HMD, with it operating as a wireless HMD for PS6.

 

[Tkumpathenurple]

For the PS6, I predict memory and storage will be the main tweaks, and that persistence will be the next generation selling point.

The SSD won't see as much developmental focus because that work was done in ample fashion for the PS5. Faster SSD's will come, but an improved write speed is all that we're likely to notice as users, as the SSD will only be used for storage rather than direct streaming (when it comes to PS6 games, that is.)

There will be ReRAM as the system's "main" storage, allowing for persistent worlds and instant resume even from complete power off. Games will copy onto it in the 5 seconds it takes to transfer ~50GB's by that time. It will be upgradeable like the Xbox Series SSD.

We'll see 32GB of 32gbps GDDR7 on a 256 bit bus for 1024GB/s bandwidth, coupled with a single, small stack of HBM. Why the HBM? For the poetry of it. The PS2's EDRAM allowed it to punch above its weight for years, and the X360's gave it free AA for the entire generation. Bring it back in this age of bandwidth hungry BVH's and AI.

 

[Globalisateur]

Not sure what Sony and MS are doing exactly, but I know what I'd advise them to do:

Sony: Mobile PS5/PS5 Pro Switch, essentially. Keep backwards compatibility for both implicit for a long while, PS5 base BC in handheld and PS5 Pro BC in docked. People want handheld devices, and care less and less about each console generations "upgrades". Sure you can have upgrades for the future, 18-24gb of ram, a GPU with high end matrix multiplication and formats, maybe even a dedicated high end NPU. Note: on current devices these are for constant low end AI calculations while GPU is for intensive workloads. However NPUs are more power efficient here, and for a mobile console a large one that can do constant intensive work might make sense power wise.

Plus you get to take the controllers off the side and turn the screen/processing into a VR headset for cheap. Maybe Apple Vision comes along and makes AR/VR super popular in a year or two (2025, 26?). Sure this makes the base unit more expensive, you need a 4-5k screen and extra cameras and watnot. But this way you can sell a VR kit not for $549 but $199, way cheaper. And at $699 total you're probably selling a high end AR/VR headset as cheap or cheaper than anyone else is.

MS: You've got the money, split the market. For low income/mobile wanting market segment sell a mobile console at $349-399 (whatever you can get it down to). It needs to have more ram than the Series S, devs want to drop that asap. But otherwise it just needs to be something BC with the Series S that devs would rather target while mainting BC with Series X/PS5. 1080p 8" screen, only need a few upgrades other than 16gb of ram minimum that will be cheap by 2027/8. Basically a target for games to hit at 1080p 30fps.

And then for the high end one you target 4k+, 120hz? Near future 3d? (Lightfields! Think if holograms were real). Sure it's a home console but you can charge $599 for this easy. 15gbps(+?) SSD; true games find it hard to use more than 8 cores, but super high clocks and a large cache. Go all out, get as much of a chiplet based GPU arch as you can and target 5080 levels of performance, or 5090 even. Anything the mobile one runs at 1080p 30 this can run at 4k60 upframed to 4k120.


Either way any mobile console will need a few things ones today don't have, which is a cheap but useful 5g modem. An increasing number of people are mobile only, their console needs to accomodate that without the need for wifi.

Sony has that already covered with their Portal. I am expecting to improve and tweak their Portal in the future.

For the PS6, I predict memory and storage will be the main tweaks, and that persistence will be the next generation selling point.

The SSD won't see as much developmental focus because that work was done in ample fashion for the PS5. Faster SSD's will come, but an improved write speed is all that we're likely to notice as users, as the SSD will only be used for storage rather than direct streaming (when it comes to PS6 games, that is.)

There will be ReRAM as the system's "main" storage, allowing for persistent worlds and instant resume even from complete power off. Games will copy onto it in the 5 seconds it takes to transfer ~50GB's by that time. It will be upgradeable like the Xbox Series SSD.

We'll see 32GB of 32gbps GDDR7 on a 256 bit bus for 1024GB/s bandwidth, coupled with a single, small stack of HBM. Why the HBM? For the poetry of it. The PS2's EDRAM allowed it to punch above its weight for years, and the X360's gave it free AA for the entire generation. Bring it back in this age of bandwidth hungry BVH's and AI.

SSD is the most boring aspect of PS5 and it's going to be the same on PS6. This is the decompressors and caches that do the most important work and very likely increase PS5 performance in many areas. Here a short excerpt of Nixxes interview about how helpful the decompressors are on PS5. Before people think PS5 decompressors are just a conspiracy theory.

The PS4 CPU cores were not so stellar and the PS5 and the PCs were far more powerful. With the PS5, that gap has certainly gotten smaller. And there's still quite a few things on the PC where there's more overhead, like the APIs have more overhead, we don't have the decompressor for example, we don't have hardware doing decompression for us as we're streaming in content - that gets left to the CPU. So we certainly have more CPU challenges to go around even when we're doing the same things.

Without those decompressors, I doubt Spider-man would run above 60fps (or even stable 60fps) on PS5 with its puny underclocked mobile Zen 2 CPU. The stable 60fps mode is probably there because of those I/O decompressors. This is why I say Cerny will likely want to improve the performance of the I/O further because they likely increase the performance in (the few) games heavily using them for just-in-time streaming. And we know Cerny is specifically looking at Insomniac open-world games for his console designs. Bluepoint also used just-in-time streaming in their launch game.

 

[function]

I don't think 3nm provides enough perf and power scaling to give us a good generational bump.

I agree if we're just talking about using evolved versions of the core technology currently found in RDNA (in a similar way to how RDNA followed GCN).

There are technologies that are game changing and that benefit from specialised processing units like tensor cores though, and these have the ability to bring about generational changes using less silicon than simply pushing for a wider version of the GPUs that consoles currently have.

I think it's these technologies that will really define next generation. Cyberpunk on consoles versus path traced on high end PC with ML upscaling and high quality frame generation is already a huge step up.

Upscaling, frame generation, high performance ray/triangle testing, better ways of handling and updating BVH trees, more efficient handling of micro-polygons and super high poly meshes ... there's a lot there to go at even if you were only at N3.

 

[McHuj]

We'll see 32GB of 32gbps GDDR7 on a 256 bit bus for 1024GB/s bandwidth, coupled with a single, small stack of HBM. Why the HBM? For the poetry of it. The PS2's EDRAM allowed it to punch above its weight for years, and the X360's gave it free AA for the entire generation. Bring it back in this age of bandwidth hungry BVH's and AI.

I would love to see HBM in a console, but after almost a decade the cost still has not come down sufficiently for it to be even used in enthusiast GPU's. I think there is some good news in that there are a few startups out there that are working on substrates for chip stacking to reduce cost of a chiplet design including HBM on an organic interposer vs a silicon one. I just don't think the tech will be mainstream by the time new consoles are coming.

However, hypothetically if someone did make a console with HBM, you'd have it very likely backed by LPDDR not GDDR7, they wouldn't put two power hungry and expensive technologies on the same die. I think a 16 GB HBM stack with 1 TB+ bandwidth and 32-64 GB LPDDR would make a pretty compelling console.

I agree if we're just talking about using evolved versions of the core technology currently found in RDNA (in a similar way to how RDNA followed GCN).

There are technologies that are game changing and that benefit from specialised processing units like tensor cores though, and these have the ability to bring about generational changes using less silicon than simply pushing for a wider version of the GPUs that consoles currently have.

I think it's these technologies that will really define next generation. Cyberpunk on consoles versus path traced on high end PC with ML upscaling and high quality frame generation is already a huge step up.

Upscaling, frame generation, high performance ray/triangle testing, better ways of handling and updating BVH trees, more efficient handling of micro-polygons and super high poly meshes ... there's a lot there to go at even if you were only at N3.

I agree. I really hope next gen goes all in on raytracing and for that they will definitely need to some dedicated hardware to squeeze out performance.

I think AI-based denoisers and upscaling like DLSS is the correct approach, so I hope we see something similar from AMD. I wouldn't be surprised if we saw some early attempts in the PS5Pro. It would be good for the PS5 Pro to be a kind of testbed the next gen.

 

[fehu]

AMD is planning to use infinity cache* on the IO die and in the future on gpus too. I have no idea what size must it be to be useful, but an alternative to HBM does exist for the feverish minds of console geeks.

* I meant stacked cache

 

[Frenetic Pony]

Sony has that already covered with their Portal. I am expecting to improve and tweak their Portal in the future.

Sadly PS Portal is a failure on every level. You also can't travel with it, and it's questionable how useful even nearby streaming (same city) will be.

The assumption that all console generations must simply upgrade the graphics/ram/cpu should be questioned, and at this point I'd say it's a no.

PS5 can run Star Citizen, an upgraded one could run it at high settings on a handheld, full raytraced GI and all at 1080p (60fps even?). Star Citizen is the ultimate dream game target short of "literally the Matrix", so if a handheld can run it and look good doing so there's little point in a dedicated stationary console that's faster. As we're not getting to "literally the Matrix" by 2028 I don't see the point of a stationary console, unless you're going with one that's a massive expensive upgrade over a base mobile console that can run Star Citzen well already.

 

[vjPiedPiper]

I wonder if there is some trade off, of total die area vs using the absolute latest or latest+1 process,
where it is actually cheaper to use more silicon, in an older process vs less silicon in the latest process.

Eg. The cost of 600mm on 5nm dies, vs 350mm of 2nm dies.
If AMD can get their APU's to be scale across chiplets properly,
we might even see a next-gen console with 3 x 200mm 5mn dies.
instead of a single 350mm @ 2nm die.
It's more silicon in total, but you'd surely save some $$ by not using the absolute latest process.

Multiple smaller dies would be great for yields,
your not going to be competing for the absolute latest process so you can probably going to be able to pump out massive numbers even at launch.
cooling is probably easier with a larger surface area.

Eg. An 8/16 core CPU chiplet + 2 GPU chiplets, possibly even sitting on a HBM or X3D cache style chunk of memory.

To stay inline with the actual topic..
I predict the PS6 will be 1 CPU die of 16 Zen7c cores, 2 x custom RDNA 6 GPU cores, all on a 256Mb X3D cache layer acting as a unified system cache, connected to 32Gb main GDDR.
2TB SSD.

although really I just want a MI300A

 

[Nesh]

There is something that I find troubling with next gen, and I sense we may either see more diminishing returns or the industry crashing.

Being involved with the creation of game assets and seeing the craftsmanship of some AAA games, I don't thik it's sustainable to be pushing the envelope.

CG movies are non interactive, they last a few hours, the creators have full control of every scene to make it as realistic as possible. There are cheats everywhere. It needs craftsmanship but the amount of people required is less and there are disciplines that are absent.

But games? The optimization required to make the assets functional in real time graphics, the work that has to be put to make everything work consistently during every interaction, the amount of assets required to be fine tuned and detailed, the amount of content, etc to create games that last for hours while playing well and keep the player engaged, connected to the story and characters is way too much.

The artists are hitting their ceiling of expertise. Going beyond that requires crazy amounts of manpower and effort. The costs of creating even more complex natural worlds that are getting closer and closer to CG movies are unfathomable.

 

[Davros]

Predict: Next gen console tech

Making Toast obviously....

 

[Frenetic Pony]

Making Toast obviously....

That's current gen you out of touch phillistine

Next gen is a wafflemaker

 

[Lurkmass]

Upscaling, frame generation, high performance ray/triangle testing, better ways of handling and updating BVH trees, more efficient handling of micro-polygons and super high poly meshes ... there's a lot there to go at even if you were only at N3.

That sounds like the exact opposite of your proposition to "less silicon" ...

 

[function]

That sounds like the exact opposite of your proposition to "less silicon" ...

"Less silicon" wasn't my proposition. What I actually said was:

"There are technologies that are game changing and that benefit from specialised processing units like tensor cores though, and these have the ability to bring about generational changes using less silicon than simply pushing for a wider version of the GPUs that consoles currently have."

It wasn't a long post, and I thought I was being super clear about this.

 

[davis.anthony]

"Less silicon" wasn't my proposition. What I actually said was:

"There are technologies that are game changing and that benefit from specialised processing units like tensor cores though, and these have the ability to bring about generational changes using less silicon than simply pushing for a wider version of the GPUs that consoles currently have."

It wasn't a long post, and I thought I was being super clear about this.

You raise a good point and think it's important that to realise that sometimes, it's better to dedicate a small amount of transistors to a specific function, rather than chucking more GPU compute at the problem and try to brute force efficiency.

Sony's I/O decompression block is prime example of this, they could have used GPU compute to handle it but I imagine moving it over to a fixed function block had way more benefits.

 

[Lurkmass]

"Less silicon" wasn't my proposition. What I actually said was:

"There are technologies that are game changing and that benefit from specialised processing units like tensor cores though, and these have the ability to bring about generational changes using less silicon than simply pushing for a wider version of the GPUs that consoles currently have."

It wasn't a long post, and I thought I was being super clear about this.

Either way, it's doubtful that implementing more specialized hardware will result in more gains for lower complexity as opposed to straightforward scaling ...

We aren't even certain yet if that they are or will remain "game changing" in the future. Some vendors would prefer not to be burned again by another crazed fad for more specialized hardware (MSAA, tessellation, dynamic states, shader objects, VRS, etc) that turns out to be irrelevant later on ...

 

[Shifty Geezer]

That's why the custom hardware should be something proven as something devs will use, such as ML upscaling and frame interpolation, and BVH management, and IO acceleration. Look at mobile chips and their many functional units - you get the right ones, you save a truck load of general purpose silicon to do the same work.

 

[neckthrough]

Either way, it's doubtful that implementing more specialized hardware will result in more gains for lower complexity as opposed to straightforward scaling ...

No, there is no doubt that fixed function hardware is more efficient. Efficiency isn't the problem. Their "fixed" nature is. So you want to be judicious about what idiom the fixed hardware is accelerating.

We aren't even certain yet if that they are or will remain "game changing" in the future. Some vendors would prefer not to be burned again by another crazed fad for more specialized hardware (MSAA, tessellation, dynamic states, shader objects, VRS, etc) that turns out to be irrelevant later on ...

Correct, we aren't certain. Some things stick. Others don't. Choosing the idiom carefully based on domain expertise goes a long way. And they don't need to last forever. You mentioned MSAA as a negative example, but MSAA actually gave us excellent service for a decade, if not more. Baking that was 100% worthwhile for a few GPU generations, even if it eventually fell out of favor. Matrix accelerators have already paid for themselves and will do so for the foreseeable future. So what if they become redundant after 10 more years? Judging every in-vogue idiom with the "crazed-fad" lens is extremely short sighted. Some ideas are bound to fail. That's ok. I'd say it's ok for 9 out of 10 ideas to fail if it gives us that 1 moonshot that pushes the field forward substantially.

 

[Lurkmass]

That's why the custom hardware should be something proven as something devs will use, such as ML upscaling and frame interpolation, and BVH management, and IO acceleration. Look at mobile chips and their many functional units - you get the right ones, you save a truck load of general purpose silicon to do the same work.

If developers are okay with the idea of the fragmented nature and black box approach to temporal filtering then maybe it's fine to look into an ML upscaling solution supposing there's no political breakdown between console vendors and whomever gains control over modifying the neural network. It's questionable if we really need a hardware implementation of frame interpolation since a competitive software based solution could potentially reach maturity ...

Better acceleration structure management/rebuild looks attractive to developers but it doesn't inspire confidence that the closest vendor exhibits several major flaws in their current implementation so they'd have to potentially implement even more specialized hardware to fix these flaws on top of their already complex hardware design. Geometry shaders were also once attractive to graphics programmers but most hardware vendors absolutely hate implementing them going by their recommendation to avoid their usage altogether ...

It's not necessarily wrong to take the more general purpose approach when that has been the clear trend for over two decades now ...

No, there is no doubt that fixed function hardware is more efficient. Efficiency isn't the problem. Their "fixed" nature is. So you want to be judicious about what idiom the fixed hardware is accelerating.

Depends if they're actually been used or not ...

Correct, we aren't certain. Some things stick. Others don't. Choosing the idiom carefully based on domain expertise goes a long way. And they don't need to last forever. You mentioned MSAA as a negative example, but MSAA actually gave us excellent service for a decade, if not more. Baking that was 100% worthwhile for a few GPU generations, even if it eventually fell out of favor. Matrix accelerators have already paid for themselves and will do so for the foreseeable future. So what if they become redundant after 10 more years? Judging every in-vogue idiom with the "crazed-fad" lens is extremely short sighted. Some ideas are bound to fail. That's ok. I'd say it's ok for 9 out of 10 ideas to fail if it gives us that 1 moonshot that pushes the field forward substantially.

From what I've seen, most hardware vendors will just remove unused features from their future designs and if the said feature in question is an API requirement they'll just "emulate" the said feature but all of a sudden you've got another potential API performance minefield on your hand ...

If some hardware vendors have enough foresight to get an idea how feature xyz will pan out, most of the time they'll avoid committing the same mistakes as before to avoid wasting their time and resources. Implementing a bunch of redundant hardware features is a sure-fire way to head into obsolescence ...

 

[Seanspeed]

2nm won't see the light of day before 2026 or 2027 on iphone.
Sounds complicated, not to mention the high cost of the new node.
Maybe samsung 3nm Gaa could be significantly faster than 3nm tsmc finfet and cheaper.

2+ years before being seen in GPU's or a console APU doesn't seem unreasonable. I know the economics become harder still by then, but I think they're gonna need to spend more on something to get a legit next-gen leap and consumers are gonna have to pay for it. So it's either 2nm or they go with some large cache chip stacked underneath or something that's going to add to silicon/complexity costs no matter what.

I dont think a single basic 300-350mm² die on 3nm is gonna offer them the ability to really push things enough.

I should also say I'm a big advocate of pushing the next gen consoles further out than even 2028, which would make 2nm an easier justification. And if it is going to be 2028, I'd rather see a $600 proper next gen console rather than a $500 underwhelming one.

 

[neckthrough]

If developers are okay with the idea of the fragmented nature and black box approach to temporal filtering then maybe it's fine to look into an ML upscaling solution supposing there's no political breakdown between console vendors and whomever gains control over modifying the neural network. It's questionable if we really need a hardware implementation of frame interpolation since a competitive software based solution could potentially reach maturity ...

We are arguing about fixed-function vs. programmable hardware. You are conflating hardware fixed function blocks with proprietary software algorithms that make use of those boxes. The hardware boxes that ML upscaling uses are generic matrix crunchers. In fact this is one of the best examples of a solid idiom to harden in hardware. Similar story with frame interpolation -- there is no "hardware implementation" of frame generation. The hardware box is the optical flow engine, which is a fairly generic idiom, although arguably less useful than matrix crunchers as of today. Like I said, some things will stick, others won't. I just can't grok your argument about "control over modifying the neural network" -- that has nothing to do with what we are discussing here.

It's not necessarily wrong to take the more general purpose approach when that has been the clear trend for over two decades now ...

As I've stated in the past, I am a huge champion of general purpose computing. But general purpose computing was amazing due to the exponential pace of Moore's Law. That's dead, son. So we have to accept some specialization.

From what I've seen, most hardware vendors will just remove unused features from their future designs and if the said feature in question is an API requirement they'll just "emulate" the said feature but all of a sudden you've got another potential API performance minefield on your hand ...

If some hardware vendors have enough foresight to get an idea how feature xyz will pan out, most of the time they'll avoid committing the same mistakes as before to avoid wasting their time and resources. Implementing a bunch of redundant hardware features is a sure-fire way to head into obsolescence ...

This is very defeatist. Nobody has a crystal ball. But if you don't take informed risks, the world will move on without you. Make mistakes. Fall down, pick yourself up.