Yaknow, there's more I can ramble on regarding possible future of consoles, but I don't have the time to put it within this post. Sometime later, I might reply to this and go stream of consciousness on the topic. Although it'd be more related to the big consoles than the Switch.
And ramble time it is.
First off, if you want to focus more on the Switch and save some time, feel free to skip ahead to after this post.
Reminder: I'm just a rando outsider layman. Enjoy what I'm saying here for fun, but don't take it as anything more serious than shooting the breeze while we're waiting for more Drake-related stuff. Maybe I'll get a bunch of things wrong; corrections are always welcome.
Alright, the big consoles moving forward...
Starting at the top, what are the constraints of a device like the PS5 and Xbox Series X? (the following will actually sound quite familiar... Switch isn't the only system that has to deal with these things of course)
They're basically computers in a box that you place somewhere and then hook up to a display. Such a box can only be so big, before they kind of start becoming a pain to figure out where exactly you're gonna put it, right? From that, there's a limit to how much power it can draw, because that power gets converted to heat. And that heat needs to be removed efficiently and 'quietly' (as decided by... whoever's in charge here). Also, ideally, such a device doesn't noticeably heat up a room while running,
unlike some high end desktop PC setups. Anyway, there's some power draw constraint.
Also, these are supposed to be mass market devices. Multiple tens of millions are expected to be sold over their lifetime, right? Ergo, their retail price can only be so high to remain affordable to a wide enough audience. And in turn, that infers a constraint on how costly can the device be to produce in order to not financially wreck yourself.
There is also an
expectation constraint. This applies moreso to Playstation than the Xbox, and that's mainly due to branding. Over the years, what has the general audience been trained to expect from The Next Console Generation? People expect to be WOWED to get motivated to spend their money. And first impressions are the best way to attempt that, right? So visuals have been the traditional initial selling point. And I think that I recall reading Crusters mention in some post about the diminishing returns on this sort of thing; that more and more effort has to be put in order to impress people. So that establishes some performance floor needed to hit for a PS6 or PS7. A 'Pro' also has its own floor to hit, even if its lower. A Series XX/SS is a bit murkier; Microsoft gave themselves some leeway on what direction to go in.
So there's only so much energy a system can use, it can only be so costly, and it needs to offer at least so and so.
From a PS6, you're probably expecting the equivalent output of at least several times of what the PS5 can achieve, right?
How do you improve visual capabilities? It's some combination of raw grunt (shader cores and clocks) and features/techniques which amplify said raw power.
How do you improve raw graphics processing power? More shader cores and/or higher clocks. More cores = more Area = more expensive chip. Higher clockers = more energy/Power.
Remember that the PS5 is on the N7 family of nodes. Now, AMD does have RDNA3 on N5. N7->N5 advertises -30% power draw and 1.84x logic density. There's not enough raw power increase there while still staying within the same power usage. What about techniques? As far as I'm aware, RDNA3 doesn't use dedicated hardware for accelerating matrix math (ie no tensor core equivalent). So nothing like Nvidia's hardware accelerated DLSS. Forget it, N5 isn't enough to offer a PS5 Pro or PS6.
What about N3? RDNA4's expected to be on that. N5->N3 should be another -~30% power draw and... I wanna guess ~1.6x logic density? Compounding together N7->N5->N3, you get what, ~-51% power draw? You can roughly double the shaders, keep the same clocks, and come out close enough to even in power. But double the raw grunt combined with architectural improvements on its own still isn't at that 'next generation' level. Probably not even a Pro, depending on what a Pro is trying to achieve. Although if AMD can offer some hardware accelerated temporal upscaler (ie a DLSS competitor that's an all around improvement over FSR 2), even v1 of such a thing could be enough for a 'Pro'. Of course, I haven't taken into consideration the cost of such a chip; both to design and to manufacture. I dunno; a 'Pro' might not even be economically viable?
And then further into the future... N2, with manufacturing projected to start in late 2025 and product in 2026. N3->N2 currently advertises another -25-30% power draw and a chip density increase of '>1.1x'. 'Chip density' in TSMC terms should be a mix of 50% logic, 30% SRAM, and 20% analog. Ergo, I'm guessing at least 1.2x logic density, maybe up to 1.3x? Uh, that's not good. That really screws with the area (or transistor)/$ proposition. Setting aside the $$$ for a sec, would a PS6 be viable on this node, as far as power and performance goes? Maybe, but it cannot be only through raw grunt + architecture improvements. A leap in features/techniques will be necessary here, IMO. Also, add in a couple of years for cost to depreciate. Maybe want the node after N2 to start up so Apple can move over to that to free up capacity. We're somewhere in that 2028-2030 window here. There'd still be sticker shock.
...which is just as well. Digression here, but reminder for the readers:
Nvidia's been investing heavily in AI research for a long time and is now multiple iterations in with DLSS and tensor cores. To contrast, AMD... has not. Because keep in mind, it wasn't that long ago that AMD was in dire straights. Zen's release in 2017 was a hail mary. Given that the computing world has shifted to emphasize machine learning, I have zero doubt that AMD started invested in that area as soon as financially possible, contrary to all their current public bluster about general shaders being sufficient. Maybe it'll take until the late 2020's to see the fruit of their efforts here? And yes, I'm a believer in dedicated hardware accelerated, ML powered temporal upscaling over generalist shader powered meatbag tuned.
That was all focused on the graphics side of things, but what about the CPU side? There was a huge leap from PS4 to PS5 thanks to combination of a massive increase in IPC (going from the Jaguar cores to Zen 2) and more than doubling the frequency. So there's more than a few times increase in CPU power there. That's not going to repeat again with a PS6 by the year 2030. At the very least,
we're not doubling frequency again. We're not going from mid 3 ghz to 7 ghz. I'd expect maybe a quarter increase or so to the low-mid 4 ghz
at best, if the sweet spot on the power-frequency curve keeps creeping upward. You can't push CPU frequency too hard, because the energy spent there could've gone towards the GPU instead. You probably also don't want set too high a requirement to maximize yields (what's acceptable from the manufactured dies), which dips into the $$$ side of things. I wouldn't expect a similarly significant leap in IPC. Once you get past the jump from Jaguar to Zen, it's more incremental. AMD did great work with Zen to Zen 2 then to Zen 3, but Zen 4 doesn't seem to be all that great on the IPC side of things given the amount of time. It gives the impression that the lowest hanging fruit's already been picked. Anyway, what are the ways to increase IPC of a CPU core? Better rearrangement of transistors, increasing transistor budget, improved
branch prediction, and... what else? Ehh, yea, there's cranking up cache, but that's area/$$$. AMD seems to like their 'mid' sized jack of all trades type approach to the Zen cores, so I don't expect a whole lot of transistor budget expansion for IPC increasing.
What about more cores? I'm not really expecting such, as I'm not sure if that's worthwhile. That's an increase in manufacturing difficulty. Also, you'd probably to have to lower the max all-core clock to not creep into the GPU power budget. Plus,
Amdahl's law. It's the computing world's version of 'a fleet moves at the speed of its slowest ship'. The impression I have is that for a lot of currently existing game design, they're not super parallelizable. That there'll at least be a main thread(s) that just can't be further broken up into small chunks and so single thread grunt is still necessary. But I might also be completely speaking out of my ass here. Actual devs, correct me on this.
Another digression: Hmm, design-wise, can we go over 8 cores in a cluster anyway (to avoid increased latency from inter cluster communication)... with Zen 3, AMD shifted to core complexes of 8 cores, with the interconnect officially described as a bidirectional ring. But
Dr. Ian Cuttress suspects that it's a bit more than that. I suppose it's possible with a bisected ring?
Hot take: I would expect the relative difference in CPU grunt (in percentage terms) between PS5 and a PS6 to not be all far off from say... the relative difference between my most optimistic target for Drake and PS5. Target! I said target, not prediction! Alternatively, it'd probably be similar to say... the relative difference between the PS4 and my semi-optimistic target for Drake.
...incidentally, you know what else sucks? Memory. Bosintang mentioned
the Von Neumann model. CPU and memory are separate. Instruction and data are fetched from memory and transferred to CPU through a shared bus. There's a performance bottleneck because you can't do both at the same time. There's also a noticeable energy cost to move those bits back and forth between CPU and memory. I'm assuming that's a matter of energy needed to propagate a signal across so and so distance. And thus, RAM to CPU is so and so much energy per bit, transferring back/to cache is an order of magnitude less expensive because cache's so much closer, and transferring back/to registers is another order of magnitude less because they're, uh, right there.
Quick example of the energy needed when utilizing RAM: IIRC, our napkin math projections for (docked) Drake ended up somewhere between... 3 and 4 watts (for both ~102 GB/s from 128-bit LPDDR5 and ~136 GB/s from 128-bit LPDDR5X... and an improved node over what's used for LPDDR5). That's a noticeable chunk if you're planning for somewhere in the ballpark of 15 watts. I think that for undocked, we tended towards lowering RAM speed to get power draw down to somewhere between 2 and 3 watts. Which still isn't ideal when one's trying to squeeze everything in within single digit watts.
But enough about Drake in this post in a thread about Drake & other future Nintendo hardware; back to future non-Nintendo hardware rambling! Remember that LPDDR is more efficient per bit than DDR or GDDR. And the consoles need more bandwidth than what DDR or LPDDR can provide. Wiki says that the PS5 uses 256-bit GDDR6 for a total of 448 GB/s. That's at least 4x the energy draw of what we'd expect from Drake before taking account the difference in efficiency between GDDR and LPDDR. So at least 14 watts? That's not nothing in the context of a ~200 watt budget. And it's only going to go up from there for a PS6. Oh sure, there's HBM if you really want the energy efficiency, but that's just too expensive for a consumer grade product. Anyway, rising energy requirement for RAM eats into what's available for the GPU.
Ah, yes, the $$$ side. GDDR's more expensive than LPDDR. It's a more specialized product, I assume? It's mainly the consoles and consumer grade graphics cards that use GDDR, right? I've the impression that the datacenter stuff use HBM. And the rest of the computing world's DDR or LPDDR. And LPDDR certainly enjoys the economy of scale due to being used in mobile devices. Gah, pricing for a PS6 doesn't sound fun.
Btw, if anybody concludes that there's a distinct possibility that in the 2030's, a theoretical 'traditional' style PS7 might not be feasible (as far as satisfying all three of performance/energy/price restraints), I may be inclined to agree!