The_hunter
Member
Pssr in theory should have better motion clarity over taa.
Last edited:
-
Hey Guest. Check out your NeoGAF Wrapped 2025 results here!
PS5 Pro is getting PSSR 2.0 between January and March 2026
- Thread starter IbizaPocholo
- Start date
- Rumor Hardware
viveks86
Member
In theory, yes. But not in this game. May be the base resolution is too low to get that "better than native" look. Just looking at the screen, you can tell it's softer. And the jaggies are worse in motion, especially on road markings and any mesh/metal structures in the environment. All things that you actually look at for turning/braking reference. Hopefully 2.0 can improve on it. I'd love to turn it all back on as I play exclusively in cockpit and it really could look great with RT, if not for the visual issues.
PaintTinJr
Member
I don't know what car or challenge you are doing, or if it is the 120fps PSSR mode, but that's complete and utter rubbish for the GT F1 X2019 Dragon Tail challenge I was playing tonight with PSSR @ 60fps with the favour RT option - only an hour ago - using cockpit camera.If there's one game that needs a proper visual overhaul on the Pro, it's GT7.
Honestly, it feels like a downgrade from the base PS5 version in almost every way. Half assed ray tracing aside.
PSSR looks okay during gameplay, but once you hit replays it falls apart — terrible shimmering, awful pop-in, and car LODs constantly phasing in and out.
Resolution also looks like it's taken a hit after yesterday's update
Polyphony needs a new updated engine.
Even in the tunnel the imagery looks flawless when paused, which I had to do to check - for an old discussion point with
Zathalus
- because given the amount of involvement in the racing to keep up with the AI that is constantly between 150-230mph in the race I wouldn't have the bandwidth to look for flaws in a real game situation with tires burning up losing traction, traction control off on straights for more speed to offset a fuel map scale of 6 (lean) , so the issues would have to be very visible to draw my attention, which none have.The game looks great with PSSR at 65" in gameplay IMO , so I'm complete confused by anyone actually playing the game is mentioning anything about image issues, outside a 120fps mode my Sony ZD9 doesn't support.
Last edited:
PaintTinJr
Member
Yeah the FSR4 leak is seemingly giving people info to look over , but even after reading this article
Inside AMD’s Accidental FSR 4 Leak: How a 100K-Parameter Network Rewrites Real-Time Upscaling
When AMD mistakenly pushed the FSR 4 source code to its public repo, it exposed more than an unreleased build—it revealed a sharp, disciplined hybrid of hand-tuned graphics engineering and tiny neural networks. This is how AMD is closing the gap with DLSS without surrendering to bloated AI, and...
news.lavx.hu
I can't really put those numbers into a meaningful context to compare to DLSS and PSSR.
1ms runtime on a RX 9070/XT is still a lot of inferencing resource, if that was the design reason for the hardware cutoff on RDNA4, and AFAIK the solution does like a sort of key-frame which are heavy processed, and like intermediate frames that are much cheaper inferenced, making FSR4 look like Frame-Gen fits too well from the get go IMO, unlike say PSSR which uses the same resources per frame AFAIK.
What was interesting from that article was that FSR4 can be tuned per game and debugged, and that the neural net makes suggestions, but a developer can override that suggestion with the base FSR2 methods that FSR4 is built on, allowing it to remain less deep and not get tripped up by basic issues.
Last edited:
Zathalus
Gold Member
I don't know what car or challenge you are doing, or if it is the 120fps PSSR mode, but that's complete and utter rubbish for the GT F1 X2019 Dragon Tail challenge I was playing tonight with PSSR @ 60fps with the favour RT option - only an hour ago - using cockpit camera.
Even in the tunnel the imagery looks flawless when paused, which I had to do to check - for an old discussion point with
The game looks great with PSSR at 65" in gameplay IMO , so I'm complete confused by anyone actually playing the game is mentioning anything about image issues, outside a 120fps mode my Sony ZD9 doesn't support.
"Flawless". That is from a year ago. I'll check again today, but I doubt it has changed.
Yeah the FSR4 leak is seemingly giving people info to look over , but even after reading this article
Inside AMD’s Accidental FSR 4 Leak: How a 100K-Parameter Network Rewrites Real-Time Upscaling
When AMD mistakenly pushed the FSR 4 source code to its public repo, it exposed more than an unreleased build—it revealed a sharp, disciplined hybrid of hand-tuned graphics engineering and tiny neural networks. This is how AMD is closing the gap with DLSS without surrendering to bloated AI, and...
So FSR4 is basically FSR2 replacing some fixed mathematical prediction methods by AI. If Sony is implementing this, I wonder if they will do a software version for PS5 base, so they could share the same code and make the implementation faster from now on
Last edited:
DenchDeckard
Moderated wildly
I hope it ends up good and not a disappointment like the new AMD update.
RSMan
Member
I don't know what car or challenge you are doing, or if it is the 120fps PSSR mode, but that's complete and utter rubbish for the GT F1 X2019 Dragon Tail challenge I was playing tonight with PSSR @ 60fps with the favour RT option - only an hour ago - using cockpit camera.
Even in the tunnel the imagery looks flawless when paused, which I had to do to check - for an old discussion point with
The game looks great with PSSR at 65" in gameplay IMO , so I'm complete confused by anyone actually playing the game is mentioning anything about image issues, outside a 120fps mode my Sony ZD9 doesn't support.
It definitely has image quality issues. I'm glad you don't notice them, but calling it flawless? No chance, man.
I'm playing on a 75-inch TV, so maybe the issues im seeing are more obvious, but when it comes to PSSR and ray tracing, there are some serious issues that still need addressing.
Like Zathalus pointed out. Ray tracing and PSSR need work...
Last edited:
winjer
Gold Member
Yeah the FSR4 leak is seemingly giving people info to look over , but even after reading this article
Inside AMD’s Accidental FSR 4 Leak: How a 100K-Parameter Network Rewrites Real-Time Upscaling
When AMD mistakenly pushed the FSR 4 source code to its public repo, it exposed more than an unreleased build—it revealed a sharp, disciplined hybrid of hand-tuned graphics engineering and tiny neural networks. This is how AMD is closing the gap with DLSS without surrendering to bloated AI, and...
I can't really put those numbers into a meaningful context to compare to DLSS and PSSR.
1ms runtime on a RX 9070/XT is still a lot of inferencing resource, if that was the design reason for the hardware cutoff on RDNA4, and AFAIK the solution does like a sort of key-frame which are heavy processed, and like intermediate frames that are much cheaper inferenced, making FSR4 look like Frame-Gen fits too well from the get go IMO, unlike say PSSR which uses the same resources per frame AFAIK.
What was interesting from that article was that FSR4 can be tuned per game and debugged, and that the neural net makes suggestions, but a developer can override that suggestion with the base FSR2 methods that FSR4 is built on, allowing it to remain less deep and not get tripped up by basic issues.
We only know about the amount of parameters that FSR4 uses, because of the leak. Neither nvidia, nor Intel have disclosed how many parameters their model use.
But estimates place them in the range of a handful of millions, for DLSS3 and XeSS. And maybe 1 or 2 dozen million for DLSS4. But there is no concrete evidence on this.
The thing to consider is that a smaller model, doesn't mean fewer things to process at runtime.
But it means it can better fit in the caches. Consider that most of the ML processing with FSR4 will happen in the LDS and L2 caches. Tough, FSR4 will still have to work on the full resolution frame of the game.
The other thing to consider is that a U-Net architecture is a Convolutional Network.
AMD said that it is a hybrid of a Transformer and CNN model, but this probably means that very few things are done using a Transformer model.
It's impressive that AMD managed to make such a small CNN model look as good as this is.
PaintTinJr
Member
What? Are you joking? So not something observable when racing? Why is that important? I've never considered stopping in a tunnel and then using the camera like that on a static vehicle.
How does that make PSSR not flawless for the purpose of this "racing simulator" where you are expected to be racing? It is the equivalent of bounds checking a game and finding iffy shaded polygons and arguing the game graphics aren't flawless.
Fire up dragon's trail course event with X2019 and capture real racing footage of playing that and showing flaws, and then I'll see what you are getting at, but from a PSSR for gamers perspective - that are actual gaming, - I have no idea what the so called issues are in GT7 cockpit camera mode I should be seeing when I'm not using every last millisecond while racing to do more to go faster.
Last edited:
Senua
Member
Holy cope
PaintTinJr
Member
Are you still a gamer?
Senua
Member
More cope?
Zathalus
Gold Member
Not observable? It was immediately observable the first time I drove through a tunnel. The noise is horrendous. Especially in cockpit view. Its distractingly bad, hence I simply do not use PSSR in GT7.
It's not some subtle blink and you will miss it effect.
PaintTinJr
Member
You need a video of that then. With a GT7 made car - the X2019 - I'm not seeing it on Dragon's trail in Croatia. So given the amount of carried over assets you either need to find the issues on the setup I've described or show your specific setup so I can verify what you are claiming, and then check other vehicles, or the same vehicle on other tracks, to check it isn't some combination setup when the game has a massive amount of permutations and assets at different quality levels.
Last edited:
PaintTinJr
Member
Yeah, the guy that loves the game because of its game play loop - since the beginning - needs cope for someone making disparaging remarks about the graphics with PSSR, that even if were true would still look a million times more detailed than GT1? That's your thought? Okay.
Last edited:
Senua
Member
What are you blabbering about? It's not flawless, none of these upscalers are. Zathalus points out flaw, you then cope by setting inane qualifiers
PaintTinJr
Member
It would seem so, and for PSSR2 using the FSR4 model, I wonder if the work PlayStation needed to do, wasn't just about fitting the solution to the Pro hardware and possibly switching out the FSR2 component with their own Santa Monica studio non-ml ai scalar they reference as their image quality baseline in the GOW Ragnarök paper when using ML AI for texture inferencing.
Modifying their linear upscaler and integrating FSR4's model, and then fitting it to tile on the PS5 Pro's tiled 14MB of WGP memory. followed by testing and tweaking on games, would certainly be a valid explanation for why it is taking so long.
winjer
Gold Member
Both the Nvidia and AMD Tensor cores are a part of the SM or CU. And they both share some resources with the rest of the units.
But AMD's solution in RDNA4 has the Tensor Units sharing more resources than nvidia.
The same probably happens with the Pro. But RDNA4 has the advantage of having more advanced frontend to cope with the workloads.
The Pro still has the same RDNA1/2 shader units, caches, etc.
Exactly which is why I laugh a little when people talk about 300TOPS on this fixed hardware. It's a theoretical max and there are far more actual bottlenecks to consider when talking about its use in actual games.
Would that be worth the computational cost? I doubt it personally. Otherwise I suspect somebody would have done this for both XSX and PS5 already. I could be wrong though and somebody comes up with a software solution that has a low performance cost on the actual game on these base systems.
winjer
Gold Member
Consider that the resources shared are things like caches, workgroup scheduling and some datapaths.
Of course those 300 TOPs will never be achieved in reality. But if programmers manage to balance other operations withing each workgroup, there should be enough TOPs for upscaling.
One of the problems with upscaling is that although RDNA is a tile based renderer, to upscale it has to process the full frame.
So despite the ML algorithm can reside in the caches and not have to go to memory, when processing the full frame, be it for depth, color, motion vectors, it has to copy back and forth from the GPU memory and this takes time.
Zathalus
Gold Member
Highly reflective car like the SLR Mercedes 09. Trial mountain. That was what caused the really bad noise you can see in those videos. It's what made me put the game down in disgust.
But, thank God, that have mostly fixed it. You can still certainly see noise in the reflections, and on the highly metallic surfaces in the dash, but that intense noise is no longer present. The combination was PSSR+RT+120hz off. But now it is significantly improved. Maybe PSSR2 would eliminate the rest.
It's certainly enough for upscaling because it obviously upscales but as you say it's a balancing act of trying to make the resources/time cost of PSSR as small as possible on that fixed hardware so that the frametime hit is low and it doesn't hinder other parts of the game/engine. Same reason we don't see a 60% performance boosts with every Pro game despite the tflops difference. Add to that the differences in architecture and it's just not a very good metric to compare different systems upscaling ability.
YeulEmeralda
Linux User
It is true that upscaling is not perfect but if you want to play at 4k resolution they're pretty much mandatory.
According the link the other user shared. FSR4 it's basically FSR2 replacing some fixed algorithms by a small model. Sony could do their own implementation of FSR2 for PS5 base and share the same API so developers can implement both at the same time
PaintTinJr
Member
I think it is also important that the upscaler is relatively latency free - meaning just the immediate 2 frame history before current, for cache/bandwidth reasons - so it is at the end of the renderer, rather than on PC where FG and MFG are pretty much inseparable frameworks from DLSS and FSR4 where histories can span far more frames(as many as 30 AFAIK) because there is no guarantee of making a performance threshold(30, 60, 90, 120) regardless of your hardware, meaning you can do accumulated lighter frames in bandwidth bottlenecks and deeper heavy inference frames when you have excess compute, and then when spanned across more history one heavy frame with more lighter frames can eliminate more noise at the expense of needing hardware with bigger caches,, and particularly using FG/MFG gives the illusion of more frame-rate too, which doesn't really fit with consoles hardware or software IMO.
Last edited:
DLSS is using mainly L1 cache from SM, the 5060ti have around 4MB vs 15MB on Pro. Sure have a bigger L2 memory, buts it's much slower
winjer
Gold Member
Where did you see that nvidia runs off the L1 cache? And how do you fit a 20 million parameter model onto a 128kb cache.
I thought Pro has 256KB per WGP? Does it have more?
The weights of the model is not saved on L1, you don't have enough cache for that, it's mainly the input image (tiles), motion vectors, etc, the model is going to be million of operations over that data in a short space of time, so you need your data in the fastest memory you can have. Also, I would be surprised if DLSS is not tiled too
Last edited:
I'm pretty sure Cerny said in their Pro presentation they have 512KB per WGP, around of 15MB in total
winjer
Gold Member
Do you even know what the weights are?
I don't remember that. I remember it being doubled to 256KB on the Pro. Could be remembering wrong though.
I think he does. Those AI model weights are split into subsets, specifically active layers or tiles that fit into L1.
To compute large Matrix Multiplication (GEMM) operations efficiently, the matrices are broken down into smaller, manageable "tiles" that can fit into the high-speed L1 cache. By storing the current tiles of matrices and activation data in L1, the processor (or Streaming Multiprocessor/SM in GPUs) avoids stalling the compute units. Tiling allows for temporal locality, meaning the same data (tile) can be used for multiple operations before it is evicted, reducing total memory bandwidth consumption.
If I understand it correctly this can reduce latency overall in the memory hierarchy.
winjer
Gold Member
The weights are not some independent part of the model.
What will happen is that the model will be in the L2, that has enough space for it, then instructions are sent down the cache levels until the registers and then executed.
And what is the point of having the weights permanently on the L1 cache? They are just the final part of the decision process, on whether a neuron is activated or not.
All the matrices calculated before, with a ton of FMAs are more important for the performance.
Last edited:
PaintTinJr
Member
I bought the vehicle
and did a quick video - driving dynamics of the car are rubbish, so it is a slow lap fighting the crappy car in every breaking and steering section - handbrake was needed a lot.But I did see a few flaws on the first lap for 1-2seconds, where I think I saw a rogue aliased edge all along one of the windscreen wipers, but after about 7 laps, I couldn't see anything wrong, as I was too busy focused on the driving, and the lighting in the tunnel reminds me of the Oxford to Heathrow journey on a bus at night when the motorway streetlights are every 5-10m metres and the light changes so rapidly in the vehicle. So I'm happy for you to tell me what I should be seeing that I'm not.
I'm not sure I understand what you mean. Do they don't need to be? Are you saying the problem cannot be tiled?
Zathalus
Gold Member
I bought the vehicle
But I did see a few flaws on the first lap for 1-2seconds, where I think I saw a rogue aliased edge all along one of the windscreen wipers, but after about 7 laps, I couldn't see anything wrong, as I was too busy focused on the driving, and the lighting in the tunnel reminds me of the Oxford to Heathrow journey on a bus at night when the motorway streetlights are every 5-10m metres and the light changes so rapidly in the vehicle. So I'm happy for you to tell me what I should be seeing that I'm not.
Yeah, it used to be much worse as you can see from the capture I did, the whole dash was this noisy mess. But that was over a year ago. It's mostly stable now, only some slight instability on the metal dash vents and the section where the rearview mirror hangs from. But mostly unnoticeable in regular play. This car is a worse case scenario though, with the large interior and highly reflective exterior.
winjer
Gold Member
The weights are too much big to get in L2 or L1, they are mainly on VRAM and are partially streamed to L2 and L1 when they are needed for the specific tile.
winjer
Gold Member
Do you even have a source for that?
PaintTinJr
Member
I'm not seeing any instability in the dash vents, just rapidly changing light on the vents from the highspeed of going through the tunnel passing 2 lights every second and a half. You'll need to capture a few screenshots from my video and annotate what I should be looking at, as It looks completely natural at normal speed IMO.
Large weight and activation matrices are broken down into smaller, manageable tiles that fit within the limited, high-speed cache. It works mostly for transformer I believe.
All sources say 256KB L1 per WGP:
DF Weekly: New PS5 Pro GPU details emerge - including a 2.35GHz max boost clock
Plus hardware VRS and mesh shaders
www.digitalfoundry.net
I guess these were wrong.
Last edited:
winjer
Gold Member
Can you provide the source that confirms that is what DLSS is doing? Or PSSR?
And remember that the models for these upscalers are MB in size, not Gigabytes.
Last edited:
Dorfdad
Gold Member
Anyone in these forms is well enough informed to understanding the PSSR 1.0 was a field test. It was released developers played with it gamers got some improvements and Sony collected data. 2.0 will be or should be vastly improved, but it will follow suite. Ps6 is where PSSR will truly shine. I'm hoping 2.0 bridges the gap and allows me to play most of my library at 4k/120 without bullshit raytracing.
I don't have one that says DLSS specifically is doing this. DLSS inner workings are not that open, but vkbest made the assumption it is:
I'm just saying the entire AI model weights don't need to fit into L1 with this method. They can be split into subsets (tiles, activation layers) and reused repeatedly.
winjer
Gold Member
But I never said that they fit in the L1, That was vkbest .
What I said is that DLSS is probably using the same locality strategy as most programs do, where they lower the level, as they are needed.
And mind you, he quoted me while I was talking about FSR4, the one upscaler that we do know about.
winjer
Gold Member
Ok
Read the thread.
Similar threads
- Locked