Custom RDNA2 Differences Between XSX and PS5

[invictis]

Both the XSX and PS5 GPUs are custom RDNA2 cards, or some may say RDNA 1.5, but there is distinct differences with how Sony and MS went about putting their GPUs together. Some parts of both GPUs are similar to RDNA1, while others are RDNA2, and other parts are unique to that GPU.

I thought it would be good to discuss these differences and what effect that will have on the potential of each console.

For instance, while not GPU related, the PS5 CPU has cut down FP pipes—from 256-bit down to 128-bit.

Locuza on Twitter has outlined alot of these differences like the ROPs for instance.

"The new Render Backend+ is a lot smaller. Instead of 4 Color ROPs + 16 Z/Stencil ROPs, the new RB+ has 8 Color ROPs + 16 Z/Stencil ROPs. Xbox Series/RDNA2 GPUs have half the amount of Z/Stencil ROPs per SE vs. PS5/RDNA1. PS5 pays a lot more for the Render Backend, area wise."

So it appears the PS5 is using RDNA1 ROPs, while the XSX is using RDNA2 ROPs.
What difference can we expect to see because of this?
From what I understand one of the changes to the RDNA2 ROPs was the inclusion of VRS, which further proves that the PS5 does not have VRS.

Alot of these differences hint towards the PS5 starting its development before the XSX, which is backed up by GutHub.

There is also differences with the Render Frontend and Backend, and the layout of the WPG on the PS5 is the same as RDNA1.

What other differences is there with these Frankinstein GPUs, and will it matter in games? Is RDNA2 better allround than RDNA1, or not?

 

[BRiT]

Other thread touched on this a lot @ https://forum.beyond3d.com/threads/...echnical-spin-post-launch-2021-xbsx-ps5.62271

 

[Tabris]

For instance, while not GPU related, the PS5 CPU has cut down FP pipes—from 256-bit down to 128-bit.

https://twitter.com/i/web/status/1364232411123494913

 

[iroboto]

We've now seen 6700XT vs 5700XT normalized. And the results are nearly identical. In this case, there's no difference between a RDNA 1 and RDNA 2 CU aside from feature set and the ability to clock higher, a least with respect to the benchmarks shown so far.

If one would want to make an argument about the differences, I would only look at the feature set differential. Everything else is likely a wash or in this case, identical.

All in all, the only differences between the two performance wise will come down to how games are made in the future 4-7 years from now. There's nothing really to predict at this moment, it takes a long time for game engines to change, and COVID didn't help. I suspect this generation will last longer than normal as well because of the setbacks from COVID. A next gen game is coming, but it won't be for a couple of years.

I can't tell you it's going to be about VRS, or ML, or RT or whatever. There are a lot of great features introduced in FL12_1 to 12_3 that largely were overshadowed in favour of larger features like RT and DLSS etc.

 

[Deleted member 13524]

Alot of these differences hint towards the PS5 starting its development before the XSX, which is backed up by GutHub.

Starting Development != Ending development. The SeriesX actually released a couple of days before the PS5.

The only fact you can take from the github leaks is that the Oberon project obviously wasn't finished by that doc's late 2018 date. The final GPU clocks were significantly higher and the referred Navi 10 Lite (probably an overclocked RX5700 at the time) didn't even have any RT capability.

For instance, while not GPU related, the PS5 CPU has cut down FP pipes—from 256-bit down to 128-bit.

The PS5's FP co-processors support 256bit, as confirmed by Cerny on the road to PS5 video.

The latest theories on the smaller FPU sizes point to AMD shortening that part of the pipeline. On one hand there's some area savings, but on the other it should prevent the FPU from continuously running FP256 operations at max clocks due to a risk of hotspot overheating. It would go hand-in-hand with Cerny saying it's not great for performance to constantly FP256 ops and warned about it causing the CPU to downclock.
Which is a non-issue for almost all games because AFAIK they don't run AVX2 / AVX256 ops in any meaningful way. The PC port of Horizon Zero Dawn used 128bit AVX1 and even that created a number of problems.

Is RDNA2 better allround than RDNA1, or not?

Clock-for-clock, on rasterization they're similar. Using an average of several games, at 1080p RDNA1 wins by 4% and at 1440p RDNA2 wins by 2%.
RDNA2's main advantages are the massive clock bumps, Infinity Cache allowing for lower external bandwidth (in the case of the PC parts) and obviously RT capabilities.