Sony PlayStation 5 Pro

Looking at the bandwidth bottlenecks, expect next gen PS6 to have quite a lot more than 32GB of memory and more than 1TB/s of bandwidth if its not going to bottleneck acceleration for Raytracing and more so for AI/ML. Memory architecture seems to be the biggest bottleneck on the PS5 pro for ML acceleration for graphics workloads.
Most consoles were more or less bandwidth limited and especially ps4pro was. They did same "mistake" with ps5pro so why do we think now they will dramaticly increase cost of console and make it bandwidth killer. I mean I realy would like to see it but in the end it has to be mainstream machine.
 
It's nice to finally have some of the longstanding uncertainties cleared up by Cerny. The RT speed up is specifically for ray calculations and the AI is done on the shader core. They passed on the dual issue as well. Still no mention of whether or not there is support for DX12U feature equivalency. I'm also still curious to know if those features would have any real world value over the existing set.
 
Last edited:
Most consoles were more or less bandwidth limited and especially ps4pro was. They did same "mistake" with ps5pro so why do we think now they will dramaticly increase cost of console and make it bandwidth killer. I mean I realy would like to see it but in the end it has to be mainstream machine.
You have a good point about the cost, but let me explain my point. 1.) iirc the Xbox One did not in fact have mem bandwidth issues nor has the base PS5/XSX. Memory wasnt an issue this gen except now with PSSR.

2.) The cost of the SSD(which is also part of the memory hierarchy iirc) took up alot of the cost this gen. Developer feedback was to provide sata ssds and more RAM and they went with full on NVMe which was good. As well a doubling of RAM was enough for this gen but next gen you will have 3 major accelerators, one for Rasterization, Raytracing and another for AI/ML. With this in mind, Cerny echoed some things I've mentioned here for months if not years now; The biggest leap next gen will be related to the memory subsystem/memory hierarchy and acceleration for AI/ML and RayTracing. The large initial cost of the console will go towards these and this cost particularly the memory will go down significantly as the gen progresses. Sony will sell the console at a loss for the first few months to a year max like it did this gen and the last gen. doubling of RAM and adding some accelerators will not be enough. They'll have to more than double the RAM and increase the memory bandwidth to support the new accelerators. With something like "GDDR7X" this may give the generational leap needed at a low cost over an 8 year lifetime.

As Cerny noted, PSSR targets 300TOPs but is bandwidth limited in large part because the PS5 pro has to be hw/sw compatible with the base PS5. So they couldnt do much with the memory subsystem especially since they need 4 years to work on the machine. So now with the current gen PS5 and PS5 pro having high disk throughput, the next emphasis will be towards placing a significant upgrade on the system memory and particularly the memory bandwidth. With that they can get more out of a large initial investment on the accelerators. I saw some of the things they are trying to do with the WGPs to increase memory bandwidth and these may not be necessary/present on the PS6 with a proper memory subsystem in terms of system memory and the processor caches.

So in summary a large initial investment in upgrading the memory subsystem/hierarchy with more than double the RAM and over 1TB memory bandwidth as well as large die areas for the GPU, possibly separate die area for an NPU will produce a profitable system over the 8 year lifetime of the console and deliver a generational leap to consumers as well as ease development for Sony's HW engineers and Software developers. Perfect mainstream machine.
 
You have a good point about the cost, but let me explain my point. 1.) iirc the Xbox One did not in fact have mem bandwidth issues nor has the base PS5/XSX. Memory wasnt an issue this gen except now with PSSR.

2.) The cost of the SSD(which is also part of the memory hierarchy iirc) took up alot of the cost this gen. Developer feedback was to provide sata ssds and more RAM and they went with full on NVMe which was good. As well a doubling of RAM was enough for this gen but next gen you will have 3 major accelerators, one for Rasterization, Raytracing and another for AI/ML. With this in mind, Cerny echoed some things I've mentioned here for months if not years now; The biggest leap next gen will be related to the memory subsystem/memory hierarchy and acceleration for AI/ML and RayTracing. The large initial cost of the console will go towards these and this cost particularly the memory will go down significantly as the gen progresses. Sony will sell the console at a loss for the first few months to a year max like it did this gen and the last gen. doubling of RAM and adding some accelerators will not be enough. They'll have to more than double the RAM and increase the memory bandwidth to support the new accelerators. With something like "GDDR7X" this may give the generational leap needed at a low cost over an 8 year lifetime.

As Cerny noted, PSSR targets 300TOPs but is bandwidth limited in large part because the PS5 pro has to be hw/sw compatible with the base PS5. So they couldnt do much with the memory subsystem especially since they need 4 years to work on the machine. So now with the current gen PS5 and PS5 pro having high disk throughput, the next emphasis will be towards placing a significant upgrade on the system memory and particularly the memory bandwidth. With that they can get more out of a large initial investment on the accelerators. I saw some of the things they are trying to do with the WGPs to increase memory bandwidth and these may not be necessary/present on the PS6 with a proper memory subsystem in terms of system memory and the processor caches.

So in summary a large initial investment in upgrading the memory subsystem/hierarchy with more than double the RAM and over 1TB memory bandwidth as well as large die areas for the GPU, possibly separate die area for an NPU will produce a profitable system over the 8 year lifetime of the console and deliver a generational leap to consumers as well as ease development for Sony's HW engineers and Software developers. Perfect mainstream machine.

The wants hundered of Terrabytes of memory bandwidth GDDR7 is not the solution. They need to improve what they do with PS5 Pro and WGP.
 
Most consoles were more or less bandwidth limited and especially ps4pro was. They did same "mistake" with ps5pro so why do we think now they will dramaticly increase cost of console and make it bandwidth killer. I mean I realy would like to see it but in the end it has to be mainstream machine.
Was it a mistake or simply a matter of, we’re selling a $400 machine, so there are limits to what we can deliver?
 
Last edited:
Hi does anybody have a PS5 Pro?
I have one and compared to my launch PS5 I can hear the fan spinning. It is not loud but compared to PS5 it is. With PS5 often I did not know it it was on or on standby so I would be surprised to see the white (non standby) light. With PS5 Pro I can hear without telling.

If anybody has a PS5 pro, can you tell how yours is?
 
But I would be surprised to learn if this is not in RDNA4 as it is.
Yes, but then isn't that what Amethyst is about? What kind of collaboration would it be if AMD doesn't take anything from Sony's efforts? I think it easy to imagine AMD is doing all the work and Sony are just riding their coattails, but taking Cerny at face value, and he's earned my respect in his presentations always being no-nonsense so I do, SIE's work was their own, training their NNs, deciding how to get the performance and particular RAM performance from the silicon, using the workgroup caches, and asking this from AMD. AMD may have come to the same conclusions as, faced with the same problems and limitations, it all makes sense, but that doesn't detract from SIE's work and findings and contributions to AMD's hardware. So yeah, RDNA4 is likely going to look a lot like PS5 Pro has been described. I expect AMD is notably enriched by Sony's first-party studios providing incredibly close data on game performance and training materials, almost like their own First Party Studios to derive key insights into their considerations and models.
 
Yes, but then isn't that what Amethyst is about? What kind of collaboration would it be if AMD doesn't take anything from Sony's efforts? I think it easy to imagine AMD is doing all the work and Sony are just riding their coattails, but taking Cerny at face value, and he's earned my respect in his presentations always being no-nonsense so I do, SIE's work was their own, training their NNs, deciding how to get the performance and particular RAM performance from the silicon, using the workgroup caches, and asking this from AMD. AMD may have come to the same conclusions as, faced with the same problems and limitations, it all makes sense, but that doesn't detract from SIE's work and findings and contributions to AMD's hardware. So yeah, RDNA4 is likely going to look a lot like PS5 Pro has been described. I expect AMD is notably enriched by Sony's first-party studios providing incredibly close data on game performance and training materials, almost like their own First Party Studios to derive key insights into their considerations and models.
That’s a bit of a separate discussion for me I guess. I was actually thinking that 5pro and RDNA4 should be similar in their ML properties. If all RDNA4 is the WMMA enhancements improved over RDNA3 I would be surprised I guess.

It does seem like this solution is more complete. So I’m expecting it to be there.

I’m not sure what their collaboration is or means to be honest. It could very well be a large spectrum of just marketing fluff, to full on developing things for amd. Likely somewhere in the middle of that, but I don’t really have a strong feel of where that is.

The geometry engine sounded like a big deal, but it really seemed to be an offshoot of the primitive engine. They talked about ID Buffer should be showing up in AMD products as a result of their work, but it never did.

I feel like a big part of “semi-custom” is to allow the platform to market AMDs stuff as their own within limits. It’s not like MS isn’t developing custom stuff with AMD, they are there helping to define the direction with DirectX for instance.

I dunno. Yea. I really don’t know how much of that is marketing, and if there is really going on here more than their standard semi custom relationship.
 
Both Forbidden West and Zero Dawn look so much better. Weird though as I was always impressed by the image delivery but it's definitely a step up in both games on the Pro.
 
Back
Top