I expect we'll be on Zen 7/7c for PS6, so we could even jump PCIe 5.0 and see PCIe 6.0; with conservative clocks that'd put us at 20-24GB/s before compression. Paired with something like 32GB of GDDR7 @ 1.5TB/s and it'd be absurdly fast on the I/O. You could have entire levels/immediate areas loaded into RAM in a split second, resident alongside the framebuffer.
The only issue with this might be the cost of upgradability as PCIe 6.0 drives will likely be necessitated, but then it'll once again be for native titles, plus I believe the PS6 'll likely launch with 4TB of storage; drastically reducing most peoples' requirements for core storage upgrades early in the lifecycle (if at all); when drives those drives will be pricier.
If however we see PCIe 5.0 @ ~12GB/s [and/or GDDR6X with it] then the idea of a small, ultra-fast ReRAM pool that devs can either leave as an Automated L4 Cache or activate as a Programmable Scratchpad sounds cool..
Even with extensive developments I don't see ReRAM being possible as main storage though as they'd have to hit a bare minimum of 2TB for it to be a viable default storage solution in PS6. In addition, it'll limit upgradability options or require proprietary add-ons (which would go against their ethos this gen).
My guess as the most balanced option with two (by-then) maturer standards will be:
4TB PCIe 6.0 @ 22GB/s
32GB GDDR7 @ 1.2-1.5TB/s
Plenty of cache on the CPU/GPU.
I expect we see heavy AI & RT acceleration this time round; requiring lots of bandwidth across the board.
I'd still love to see HBM; they could probably use underclocked HBM3 Gen I/II chips -- which'll be the relative bargain basement option within the HBM space by then -- and still perform great with excellent efficiency. If they take a partial chiplet route they might be more inclined, but I still doubt it.
On the subject of chiplets. A single APU has always made sense up till now as it's a more bespoke design and the chiplet approach made more sense for off-the-shelf dies. But at this point (and with Infinity Fabric as an enabler) a bespoke chiplet design might gain more performance as a result of cost savings vs the efficiencies of a bespoke APU.
Here's a mockup logo set in the mean-time:
(clicky)
Lot of good speculation here on the tech side of things. There are a few I'd add to this:
RAM: 32 GB GDDR7 is most likely, and the bandwidth looks realistic. However, I think (well, hope) Sony go with a GDDR7W variant. Doubles density and bandwidth per chip without doubling the physical size & footprint.
SSD: I can see Sony doing 2 different SKUs with the SSDs, but none having as much as 4 TB of storage. Probably 1 with 2 TB and another with 1 TB or maybe even 512 GB. But with even better compression and decompression, you get more out of that storage, plus users can upgrade to larger capacities if they need to. As for speed, well they will probably provide support for PCIe 6.0 drives, but I can see them including PCIe 5.0 type drives by default instead.
Honestly I'd be more interested in the I/O subsystem for SSD data and data in general, including what way if any they could get the SSD data to act more granular. Maybe replicating the amount of parallel channels with some type of cache buffer system, for example, while using industry-standard m.2 interfaces.
ReRAM:
vivftp
was talking about this earlier, and it would be great to have. But how commercially ready would it be for a mass-produced console launching in 2027 or 2028? I think if Sony can have a fast-enough SSD included by default (say, 12 GB/s or 16 GB/s) and robust enough decompression I/O (around 48 GB/s - 64 GB/s), and a way to replicate channel parallelism access with a cache buffer and localized/PNM processing for it, you really wouldn't need ReRAM at all.
The SSD would be pretty fast, the decompression even faster and the cache much faster than ReRAM. So maybe just 2 MB - 3 MB of a cache buffer operating at a very fast frequency could get the job done, as long as you have fast enough decompression and a fast enough SSD.
HBM: I'd love to see it, too. Actually, I'd love to see some type of HBM-PIM. It would be more energy-efficient than GDDR7. But if GDDR7W is available, it'd be relatively cheap like GDDR7, and be more power-efficient than it, so it'd give some good mix of GDDR7 and say HBM3.
Realistically there's prob a slim chance HBM will be present but it'd be a present surprise.
Chiplets: In some way I think the PS6 is going to be chiplet-based, because the scalability and flexibility that could bring on the production & manufacturing end of things is just too good to pass up, especially if Sony want to continue making variant/peripheral devices adjacent to the console itself in the ecosystem like, say, a next-gen PS Portal. Being able to source all that to the same wafers cuts down on waste from underutilization of any given wafer.
And of course there are the other benefits for energy efficiency. Latency and bandwidth could be issues but sufficient cache would help with that, plus any innovations with cache management of course. PS5 already has some of that like with the cache scrubbers.
Also: neat logo mockups. They remind me a bit of the PS3's logo with the Italic stylization, but with a color scheme more in line with PS2, PS5 & PS5.
Haven't seen much talk about cooling but I'll like to see them invest in some kind of liquid cooling, in conjunction with liquid metal.
Below image for reference, consist of a fan, radiator, pump and reservoir.
Along with quick disconnect tubing for repairability.
Could be interesting if they'd require it. How much do these usually cost for end-customers?
