PS5 Pro is getting a big upgrade in 2026 — I asked Mark Cerny what’s coming, and why AMD’s future PC GPUs feel more 'PlayStation' than ever
Posted by PorchettaM@reddit | hardware | View on Reddit | 67 comments
NeroClaudius199907@reddit
Nvidia got ai warchest for r&D
amd got xbox & sony
Intel got?
exomachina@reddit
Better chip yields than everyone else.
Exist50@reddit
Intel? What?
andoke@reddit
Laptops
BarKnight@reddit
They still have around 75% of the CPU market and foundries.
BoreJam@reddit
Mainly because of laptops and work systems. They're dwindling in gaming but truth is gaming only a portion of the PC market.
shoneysbreakfast@reddit
Cerny said all of this exact same stuff back in March in a long interview with Digital Foundry and everyone got mixed up then and I’m seeing the exact same thing happen again.
After PSSR was finished in 2023, Sony and AMD started a collaboration called Project Amethyst and the first product of it was FSR4. The PS5 Pro is NOT getting FSR4, it’s getting a new version of PSSR that uses work from that collaboration, a reimplementation of the FSR4 training and model using a different algorithm. This is because FSR4 relies on hardware features of RDNA4 which Pro does not have, it has a custom Sony developed ML solution instead. Cerny says that PSSR2 or whatever they end up calling it will be just as efficient as FSR4 and will net the same results, it’s just not straight up FSR4.
Their short term plans for Project Amethyst are to collab together on neural networks and training for graphics which resulted in FSR4 and the upcoming PSSR2, and their longer term goal is to collab on future hardware to make it better at ML and RT. Sony had to roll their own ML/RT solutions for PS5 Pro and both companies realized it’s better if they work together, which is why Cerny is working on RDNA5.
The idea is to work together on these things so they aren’t separately working on the same problems and coming up with redundant solutions. They aren’t in competition with each other so more convergence between AMD’s GPUs and Playstation in both hardware and software makes a lot of sense and will make both better.
bubblesort33@reddit
If it has a quarter the ML performance as a 9070, and gets "the same results" as FSR4, that would be pretty weird. Maybe it gets the same results at 4x the compute time, or it'll have worse results.
shoneysbreakfast@reddit
We'll see how it plays out. Cerny is a pretty straight shooter though and I don't really have any reason to not believe him, he's typically very open about limitations of hardware/software he's been involved with.
bubblesort33@reddit
Marketing speak. You can't get the exact same results with a fraction of the ML performance. Either the frame time cost will be a lot more, or it'll be different in a meaningful way.
shoneysbreakfast@reddit
The RX 9060 XT does 410 8-bit TOPS and supports FSR4 (which uses FP8) and the PS5 Pro does 300 8-bit TOPS so the gap isn't as big as you think.
Cerny isn't a marketing guy and has a history of being honest about the capabilities of products he's involved with, he isn't just talking out of his ass.
Strazdas1@reddit
based on his previuos talks about PSSR it seems it is taking a very different approach in hardware use to FSR.
Vince789@reddit
I wonder if AMD could make "FSR3.5" based on PSSR2?
PSSR2 would be great for RDNA3 GPUs/RDNA3.5 iGPUs
shoneysbreakfast@reddit
FSR4 and PSSR2 both use ML hardware that doesn't exist on RDNA3.
Someone forced FSR4 to run on RDNA3 on Linux and the result is that it's just not practical and would take AMD a ton of effort for a very minimal benefit. Lots of information in this article.
https://www.xda-developers.com/running-fsr-4-rdna-3/
The same would apply to PSSR2, it would take a ton of work re-re-working it for an entirely different architecture and there wouldn't be a whole lot of benefit.
Jsilverstreak@reddit
It was still better performing than native. With how bad some built in taa is ill gladly take fsr4 at minimal performance gain.
shoneysbreakfast@reddit
In the three games they tested:
Ratchet and Clank was 47fps native 4K and FSR4 was 61fps in Performance (1080p internally), that's pretty underwhelming boost considering you're rendering a quarter of the pixels and that's the best case scenario
TLOU1 was 43fps native 4K and 47fps FSR4 Performance.
Horizon Zero Dawn was 83fps native 4K and 67fps in FSR4 Performance, plus the image quality was broken
So in three games you have one modest boost, one teeny tiny boost and one big regression.
This is the crux of the problem, AMD would have to go through every single game and thoroughly test FSR4 on RDNA3 to see if it a) worked at all and b) even be beneficial. Without the hardware acceleration of RDNA4 at best you're getting a very modest performance uplift in comparison to other upscalers and validating it all is a ton of work. The juice isn't worth the squeeze if you're the one paying for the effort and it's not going to move GPUs anyway.
Strazdas1@reddit
To be fair, ill take FSR4 quality over "native" even at identical framerate because FSR4 just like DLSS 2/3/4 is simply a better AA than native TAA implementations.
Vince789@reddit
Sorry, you missed my point, I'm not asking for FSR4 on RDNA3
Sony claims to have done all that work to get PSSR2's ML running on RDNA2
Thus PSSR2 should run on RDNA3/RDNA3.5 if AMD can license Sony's work with PSSR2
Would Sony be willing to license their PSSR2 work so AMD can make "FSR3.5"?
shoneysbreakfast@reddit
The PS5 Pro doesn’t use off the shelf RDNA2, it has some features from RDNA3 and some from RDNA4 and has a unique custom ML solution that PSSR relies on that doesn’t exist on RDNA2/3 GPUs. It’s a custom hybrid RDNA setup.
Vince789@reddit
Oh true, thanks, I forgot about the customizations in the PS5 Pro
Sony probably ported the FSR4/PSSR2 tech to their unique custom ML solution
I guess that means RDNA2 probably isn't capable of PSSR2
ItsMeSlinky@reddit
PSSR is Sony-proprietary; it’s not going to come to PC Radeon GPUs
Runnin_Mike@reddit
But if it's a collaborative effort it could give AMD ideas too. I don't think they were necessarily saying it would be straight PSSR just like how pssr2 isn't fsr4. There's degrees of context here that I feel like you are intentionally missing lol.
anival024@reddit
PSSR and FSR are fundamentally the same.
Sony and AMD worked on it together. PSSR is just the PlayStation branded version of it that fits within the constraints of the console's hardware. On PC, it's less restricted and will get updates more frequently and is branded as FSR (4 and up).
shoneysbreakfast@reddit
PSSR1 is a Sony specific ML based upscaler that uses the Sony developed ML hardware on the PS5 Pro and has nothing to do with FSR1-3.
The collab started after PSSR was developed.
PorchettaM@reddit (OP)
Summary of an interview with Mark Cerny and AMD execs. Main insights:
Jayram2000@reddit
Super interesting that Cerny and his team are working on direct IP for the next radeon gaming archtecture. Their partnership is way closer than I imagined
Darkknight1939@reddit
It's out of necessity. AMD on its own has consistently demonstrated they cannot compete with Nvidia.
Sony has an obvious vested interest in making AMD GPU IP more performant and having closer feature parity to Nvidia. They likely played a very big role in FSR4.
Jayram2000@reddit
AMD hasn't fully caught Nvidia, but saying they cannot compete seems a bit extreme. RDNA2 was highly competitive with Ampere and RDNA4 is comparably performant while using last gen memory and similar silicon sizes. I think if they continue to improve at the pace they have been, and Nvidia continues to degrade at the pace they've been, we'll say parity within the next two gens.
Strazdas1@reddit
RDNA 2 was not competetive. It was dead on arrival. RDNA 3 was as well. It took until RDNA 4 for them to finally implement hardware that Nvidia did in 2018.
InformalEngine4972@reddit
Ampere Aged way better though. Full dlss 4 support. Can still run most games with rt on… don’t even have to think about that on rdna2
Darkknight1939@reddit
RDNA2 products were on a massive node advantage versus Ampere products (and still lost overall in performance.) TSMC 7NM vs Samsung 8NM. Ampere had that much of an advantage as an actual architecture.
RDNA4 is definitely a big step in the right direction, but they're still behind versus DLSS (FSR4 is still a huge improvement). I'm excited to see what they do with Sony for UDNA. Especially as a bazzite user.
Jayram2000@reddit
Was TSMC that huge of a difference? I knew Samsung was worse but didn't know it to be massively worse
NeroClaudius199907@reddit
Samsung 8nm \~61mm2
Tsmc 10nm \~53mm2
Tsmc 7nm \~91mm2
InformalEngine4972@reddit
8 nm Samsung is basically 10 nm tsmc
Cheerful_Champion@reddit
Samsung 8nm is Samsung 10nm++. Transistor density is just one of characteristics of node, but Samsung 8nm had 64.40 MTr/mm2 and TSMC N7 had 91.2 MTr/mm2
DavidsSymphony@reddit
It was nowhere near competitive when you take into consideration DLSS2 and ray tracing features. And that's just in 2020. 5 years later, the 3080 is a miles better GPU to have than a 6800XT thanks to DLSS alone.
imaginary_num6er@reddit
With PS5 using RDNA2 and PS5 Pro using RDNA4, AMD does not need to worry about FSR4 being unlocked for PS5 since FSR4 will not work on RDNA2
Strazdas1@reddit
PS5 Pro is not using RDNA 4.
Strazdas1@reddit
FSR4 cannot come to PS5 Pro. It does not have the hardware necessary. This literally cannot be done.
Healthy-Doughnut4939@reddit
Dedicated fixed function hardware for BVH traversal, storage and stack management
Features copied from Blackwell
Changes to the cache hierarchy. AMD could implement a huge 256kb L1/SLM like Nvidia or Intel while eliminating the mid WGP level L1 and LDS or they could dramatically increase LDS storage to 160kb or larger like in CDNA 4.
AMD's caching system has lower latencies at all cache levels compared to Nvidia and Intel but more cache levels increase complexity as more tags checks and snoop filters need to be implemented to maintain coherency.
CDNA 4.0 features like faster fp4 and fp4 =fp6 speed
Larger GPU implementations with larger L3 MALL cache and a wider memory bus.
anival024@reddit
That will depend on the game can run the "full-fat version". If not, you get PSSR, which is fundamentally the same thing just cut down to work with the hardware's spec and all game workloads.
Spiral1407@reddit
So what was the point of pssr then?
Chocotaco24-7@reddit
I think Cerny said he thought it would have taken way longer to develop the algorithm for fsr4 but it only took the AMD team 9 months.
cloud_t@reddit
Still seems a bit long for what effectively is optimizing an existing algorithm for a specific hardware layout.
conquer69@reddit
Seems very messy. Couldn't they hold the PS5 Pro for another 6 months? Maybe AMD didn't tell them?
itsjust_khris@reddit
Holding for 6 months could've been very costly depending on where in the process of launch they already were.
raknikmik@reddit
Would be even closer to PS6 than it already is.
Earthborn92@reddit
He thought it would take longer than it really did.
kenman345@reddit
Read up a few comments, it seems Sony thought it would take AMD much longer to implement FSR4 (like years) but AMD did it in 9 months
pmth@reddit
Probably took some time to get FSR4 working 100% on RDNA 2.5
InformalEngine4972@reddit
Ps5 is rdna 1.5 , not 2.5 :)
luiz_leite@reddit
That's very nice to hear. AMD still has to catch up to NVIDIA software-wise, and after hearing about stuff like Project Amethyst and FSR Redstone I feel like they're commited to do so.
mxlevolent@reddit
I know that these hardware teams are full of intelligent, competent people, but hearing that Mark Cerny is seemingly heavily involved in the next AMD architecture seems great - because that is a very intelligent man, there.
Healthy-Doughnut4939@reddit
"Big chunks of RDNA 5, or whatever AMD ends up calling it, are coming out of engineering I am doing on the project," said Cerny. "And again, this is coming out of trying to move things forward. There are no restrictions on the way any of it can be used."
UDNA is looking like it's going to be a massive uarch rework.
I would expect:
Dedicated fixed function hardware for BVH traversal, storage and stack management
Features copied from Blackwell
Changes to the cache hierarchy
CDNA 4.0 features
Larger GPU implementations with larger L3 and wider memory bus.
ResponsibleJudge3172@reddit
256kb L1 is a datacenter only thing for Nvidia. They still have 128kb L1 for client
hackenclaw@reddit
still not worth for how much they are charging.
PS5 pro is twice the price of PS5.
exomachina@reddit
"And we're in the process of implementing it on PS5 and it will release next year on PS5 Pro."
Does this mean it's also coming to base PS5 as well eventually?
Ok-Beyond-201@reddit
A AMD-Sony collab to improve FSR and the RDNA/UDNA architecture?
Seems promising. Watch out NVIDIA.
neutralityparty@reddit
Interesting I wasn't expecting this close collaboration between Sony and AMD
Sacredfice@reddit
They have been working together for over a decade already tho.
Raikaru@reddit
Sure but now Sony is straight up helping to build AMD’s architectures
goodbadidontknow@reddit
What is wrong with helping AMD take on the two other behemoths? I say go Sony!
Darkknight1939@reddit
It seems like they've been directly contributing since at least RDNA2.
EloquentPinguin@reddit
Since the PS4 Sony used customized AMD IP in CPU and GPU, especially for the GPU this means they also have custom driver stacks which can correctly and appropriatly use the modified versions of the AMD GPU IP.
It might not have been so much work on the surface level, but internally the connection has been very strong for some years now.
neutralityparty@reddit
Yeah but it's still going directly towards competition (PC). Although I guess Sony is cool now with PC they even made spiderman available for it
Snapdragon_865@reddit
PCs are in a different market. Wasn't the case 10 years ago though
EnigmaSpore@reddit
dude, Sonys engineering team doesnt f around. they're like id software but for hardware. they know what makes games go and really do push for the little things that matter.