What would those memory changes be? HBM is out of the question, and GDDR7 just makes sense to me.
HBM could be in-play depending on affordability of the cheaper HBM3 variant. If not, then options like the eventual GDDR7W (variant of GDDR6W) could be possible if it's available by that time.
Really, the bigger advances would be implementation of PIM (Processing-In-Memory) or PNM (Processing-Near-Memory); the latter is something PS5's SSD I/O subsystem already implements a form of with the coprocessors, cache coherency engines etc. coupled with the flash memory controller and DDR4 memory. It's not hard to imagine PS6 will implement a more advanced version of that focused not just on general storage data I/O but data particular to AI & machine learning, probably with a further implementation for metadata.
For me the timelines were separate and then crossed paths.
- AMD was developing an AI solution for its FSR upscaling. RDNA 3 incorporated some matrix accelerator instructions that are important for achieving machine learning upscaling. This to me indicates that there was already an intention to do this probably in FSR3.
- Sony was developing an AI solution for the PS5 Pro, likely based on RDNA 3.
- Their paths crossed when Sony noticed that AMD's solution was better, despite being computationally heavier. So they joined forces to share models to improve both companies' solutions.
If AMD's approach was better why would SIE take the lead in announcing and putting force Project Amethyst to the public? I think that's something the side with a more desirable implementation would've done, don't you?
Truthfully I don't think either implementation at least up to a certain point was outright "better" than the other; they had different implementations for different purposes. AMD's likely being more scalable (since it's focused on PC), SIE's being more efficient for PS5 and planned PS5 Pro & PS6 design specs. Each with strengths and weaknesses, but not possible to actually specify one as objectively "better" than the other IMHO.
All that really matters now is, both companies are in partnership for Project Amethyst and there's going to be a lot of collaboration going forward with them WRT AI, ML and other related tech. I suspects SIE will try focusing on aspects specific to gaming while AMD focus on aspects related to data centers, cloud computing etc. but both sides sharing from a common R&D pool they each are actively feeding into.
To accelerate PSSR, the PS5 Pro uses WMMA instructions derived from RDNA 3. Cerny calls the PS5 Pro GPU RDNA 2.X, as it contains RDNA 2 for raster, WMMA instructions from RDNA 3, and RT implementations from RDNA 4.
It doesn't make sense to me that AMD didn't think of a machine learning solution for its upscaling when including hardware for this in its GPUs.
For me, AMD has always intended to do something with machine learning. GPU Open has had a lot of articles for a long time about different ways to use Radeon for AI, there is even an SDK, Radeon ML. Sony may have planned PSSR in 2021, but it obviously already had in mind the hardware on which it would base the AI training, otherwise, where would it do it? It would be inefficient to do it on different hardware, it goes beyond the purpose of a console.
In my opinion, the paths were somewhat separate until AMD's research yielded better results, which made Sony adapt what they had already been developing to unify efforts. In the interview with Cerny, this is clear to me, but it seems to me people are trying to build a story where Sony is responsible for everything that AMD brings.
I don't think anyone is actually saying Sony (SIE, technically) are wholly responsible for everything AMD is bringing to the table. In fact I've seen no one other than yourself implying that. However, it's also worth saying that the inverse isn't true, either, i.e Sony for whatever reason dumped all of their progress to adapt what AMD was doing because somehow AMD had all the answers for AI-driven upscaling for a PlayStation platform better than the company that's been making PlayStations for three decades :/.
So what I was saying above, that both parties having their own approaches beneficial to their own end goals, with primary intentions for different markets (console gaming for SIE, data centers/cloud compute etc. for AMD). Then at some point, they saw each other's individual results, liked each other's individual results and decides
"hey, let's work together since we've got some crossover mutual goals (gaming specifically: console for SIE, PC for AMD
)", and boom: Project Amethyst is born.
Generally speaking, there's no reason to think both sides didn't enter the partnership as equals, versus thinking one side "needed" the other's progress because their own results were proving inadequate for intended purposes. But that's just my take on it, I guess :/
