That's literally what a port is though. Taking the target platform's hardware limitations and goals (such as your examples) is part of design considerations for a good port. It's certainly not copy-paste and clearly a complex endeavor. It's a joint effort by both Sony and AMD, so I'm not taking any credit away from Sony in calling it a port. It is a port of a joint Sony-AMD collaboration to a specific hardware target
Hardware limitations and architecture. Pro is a very unique configuration of RDNA 2 and 3 (and may be 4?). FSR 4 currently doesn't work on anything below RDNA4 due to dependencies related to FP8 computation. Cerny seems to be confident that they can make it eventually work, since Pro has the required hardware grunt, so they may be able to overcome that dependency.
No. I'm not talking about codebase. I'm talking about the final result. Like taking a PS5 game and porting it to PC and adding PC specific features. That's a port. FSR4 now exists on AMD RDNA4. Cerny essentially wants that trained neural network (with the same inputs and hopefully similar output) to work on a specific target hardware which it wasn't originally built/optimized for. Ergo, a port.
RDNA 4 and the hardware in PS5 Pro are completely separate designs. The goal of Project Amethyst is to create a more ideal hardware architecture for machine learning.
With Sony being involved with the development of fsr4 and most likely cerny, probably wasn't something they would have hesitated admitting.
They already started. It just takes time.
True. But at the minimum, FSR 4 is not entirely Cerny's baby. The original CNN seemed to be. So it's just nice when he is able to call a spade a spade and not shy away from praising an alternative. Seen enough geniuses who go into some weird denial or start spinning the truth to make it seem like it was all their own work all along, or they just go silent altogether and pretend there is nothing better out there. Cerny has always come across as not one of them. I found it worth pointing that out is all.
I never claimed it was the same right now, but with the FSR 4 upscaling network being reworked to replace the current PSSR network, it's clear that in future they will be extremely similar. The network model is the thing that requires extensive training so duplicating the work between AMD and Sony makes little sense, the collaboration project is meant to combine the resources going forward, be it hardware or software. I'd expect the upscaling part of PSSR/FSR 4 to be near identical on the PS6, with PSSR likely maybe having some custom Sony work done on the inputs and outputs.
wtf happened to this thread?
Cerny also just seems like an honest guy. He doesn't bullshit you or sell you grand promises. He's very forward and admits what needs work and what is working. At least that's my feeling. He's a game developer/coder/architect, not a PR guy like Jensen or Phil Spencer.
We can agree on that. I asked myself
For the PS5 Pro, I'd assume so, but the ML capabilities of the PS6 should be plenty, even for full transformer based upscaling and ray reconstruction. It is going to be based on UDNA and should be significantly more capable than the Pro, or even the 9070 XT.
RDNA5 have Matrix Cores. That also going reduce FSR4 and future versions perf hit
The fastest previous gen AMD GPUs are about 1/4 of the TOPs of RDNA4 releases, well below the PS5 Pro.
The funny thing is - his last talk confirmed this, he just didn't explicitly call out sparsity. And for some reason - especially this decade, any statement from Sony is always assumed to be the worst of if not explicitly verified by (at least)2 Cernys.
From Cerny.
I don’t get it. Isn’t the 9070 XT something like 1.500 TOPs?
Maybe in 2 years there will be a (good) PC GPU sold at MSRP. But I seriously doubt it.
That's Int4.
I have the faintest idea that we are talking about chronic DF reflexes (when it comes to Sony) here. Cerny truly seems to be a zen guy to still willing communicate with them despite his words being always received with mountains of doubts as if he has a history of being untrustworthy.
Agreed. Pro is already in the same league as the current gen AMD GPUs for AI so the goal seems feasible. Getting this to work on older AMD cards seems like a whole different can of worms and hard for anyone sitting on the outside to determine if the juice is worth the squeeze.
In fairness - this predates Cerny's role in system design - Sony's always done dev-first all the way back to original PS1. And a lot of what made PS2 - special - was thanks to this approach (though admittedly, many titles also suffered because of it).
This is a bit more questionable though. No console to date has done this (none of the Microsoft hw either), among other reasons because it handicaps one of the main advantages developers have on consoles. But ok - if Sony actually opened up the refresh rates that displays support, I'd be willing to concede this could be a win.
This can't be system forced - titles that were tested against one version of PSSR can't be revalidated if system forces an update. However, I could see a rationale to allow users to select PSSR revision themselves - it's very PC and not at all console like, but then that ship sailed when the first Pro console launched, and we have HDR and other system settings now.
And 291 for RX 9070 i guess, which is funnily enough less than PS5 PRO.That’s INT-4 asBojji replied above. INT-8 without sparsity using very simple math is a quarter of that number.
RTX 4090: 660
RTX 3090: 320
RX 9070 XT: 390
PS5 Pro: 300
RTX 3080: 238
INT-8 TOPs without sparsity.
