Wouldn't Sony using HBM2 make it more difficult for them to scale down component costs over time? HBM2 proliferation in the memory market is nowhere near the level of GDDR6, and if Sony are the only ones really using it at mass volumes it won't be enough to bring costs down over time comparative to GDDR6 which has all three of the big memory manufacturers (Samsung, Micron, SK Hynix) onboard.
Also HBM2 requires thicc buses, but there were other people in this thread saying wider buses make it harder to scale down production components over time due to increasing the size of the silicon area. Sony's strategy with PS4, at least a big part of it, was ensuring they went with components that could easily scale down in cost over time and either had mass industry support or would in a little bit of time. That includes memory. MS, in terms of memory, went with a cheap approach last gen that didn't factor DDR3 being phased out for DDR4, which came to market only a few months after XBO released.
Sony's PS4 strategy revolved around using components and technologies that were either mature and/or would become staples in the consumer market in a small bit of time. There aren't any trends for HBM2 becoming a staple in consumer electronics for the next few years, AFAIK, if ever. There's been a couple of GPU cards with it and Intel did a (very short-lived) line of processors with integrated graphics using HBM, but that's it. I don't want Sony taking potentially another Cell type of move with PS5, that's all.
I was browsing Sony patents regarding memory controllers (only between 2014-2015, they have tons of them) and found several regarding memory controllers and memory controller functions. One describes a custom controller that can manage Ram, Reram and NAND (fits that rumor of nand chips soldered using normal ram as cache) and several regarding systems to improve access time and I/O performance on a custom controller using Reram or NAND as storage.
Could they be really using ReRam together with NAND to make a difference over pcie4 SSDs?
Sounds a bit "CELL"-crazy like.. Is Reram even in mass production?
Easily possible, and it's not on CELL levels of crazy at all. Data center and enterprise markets have been using things like NVDIMM and adjacent tech for fast storage since the start of the decade. There are two types of ReRAM: embedded (the fastest available, but very expensive and capacities much smaller), and storage-type. Companies like Crossbar-Inc provide licenses for their ReRAM IP in both types; understandably the storage-class type is a bit slower but also much more affordable (relatively speaking) and in much larger capacities.
Just going off their specifications, storage-class ReRAM offers page-addressable read, plus byte and bit-addressable writes and true random access. So in that fashion it's very much like DRAM or for better comparison Intel's Optane memory, but has potential to scale down more efficiently in process size. That's also better than NAND, which doesn't have byte (or bit)-addressable write capabilities.
ReRAM of storage-class type can provide speeds of up to 25.6 GB/s; I believe Crossbar's implementation can work over a PCIe bus, but I think it can also be tied to a DRAM controller (such as a DDR3 memory controller) if you'd like to use it closer to DRAM style. The main advantage it'd have over conventional DRAM though is in being non-volatile, so data is stored even when powered off, and it has much better P/E cycle ratings than almost all NAND memory types besides SLC (where I believe it ties, but it could be slightly better than that), and lower power consumption for writes than NAND across the board.
The only potential issue would be price; Intel sells Optane Persistent Memory (DRAM-style Optane memory) for server markets who already pay a premium anyway and therefore isn't the best comparison, but 128GB of Optane PM costs $842, or $6.57 per GB. Again though, that is sold to server and data center markets at profit, and we can assume Intel is getting at least a 2x profit for each GB, so production cost per GB is likely at most $3.20.
If you go by that amount, Sony would have to shell out $210.50 for 64GB of ReRAM. However, we can assume that thanks to the volume they'd be using it in and (potentially) doing at least some of the manufacturing in-house, ReRAM per GB could probably only cost them closer to $1.87 to manufacture (assuming a 3.5x reduction on the per GB cost of Intel's DC Optane Persistent Memory). At that price, 64GB could only cost them about $120 to manufacture. At 96GB, that cost would rise to $180.42.
Assuming they'd be selling the PS5 at $499 and at a slight loss, that could easily fit within their BOM. A MSRP $399 PS5 would be pushing their bleed on each system sold quite higher (assuming they'd be selling at a loss), though, and that'd likely push them down closer to 64GB of ReRAM (IF they're using ReRAM at all; I'm just trying to figure where this benefits them pricing-wise).
For the performance a relatively fast pool of 96GB ReRAM cache could bring, in addition to 16-20GB of GDDR6 (I
highly doubt they're using HBM2, for various reasons), that could be delivered pretty well in a reasonable BOM and push near that 16x memory increase some people seem to be obsessed about in order to define a true next-gen leap
The other question though is...is Microsoft doing anything similar? We honestly don't know, but I don't see why they couldn't work something out with Intel and their Optane Memory, since that is already seeing use in the PC space. It would also give Intel another partner to manufacture for, potentially key considering Micron ended their development on 3D Xpoint (the PCM memory analogous to Intel's Optane, as they essentially co-developed the tech together and spun it off into their own respective product lines) earlier this year (if not back in late 2018).
Personally, I hope they're BOTH taking this type of approach, because that will open up some great opportunities for next-gen gaming development and also raises the baseline for console-level gaming on PC that much more (meaning the higher-end PC specs will get utilized and pushed further themselves). It's a win-win for everybody.