Oh no a positive AMD video.
Also, as I suspected myself (due to no PSSR instability being present from what I can see), FSR4 and PSSR appear to not be related at all, other than both being ML temporal upscalers. So that makes me super curious if the PS5 Pro can support both PSSR and FSR4.
AI is already replacing adamsapple those bastardsHere’s an AI summary:
CES Expectations and AMD's Presentation
- The discussion begins with the presenters, Alex Battalia and his colleague, expressing their high expectations for AMD at CES, anticipating the unveiling of new RDNA 4 GPUs and FSR 4 technologies.
- Despite some journalists being briefed on these products, AMD did not showcase them in a formal presentation due to alleged timing issues.
- Instead, AMD revealed some technologies in a less conventional manner that surprised both presenters.
Visit to the AMD Booth
- The presenters recount their experience at the AMD booth, which was actually a suite in a hotel rather than a traditional booth setup.
- They entered with low expectations as AMD had not provided much information about the RDNA 4 GPUs or FSR 4 technologies prior to the event.
- Upon arrival, they found various displays, including a Black Ops stand, but lacked context about the technologies being showcased.
- During their exploration, they encountered familiar AMD representatives who introduced them to a research project involving FSR 4 technology.
FSR 4 Research Project Overview
- The research project featured two PCs, one running FSR 3.1 and the other presumably showcasing a work-in-progress version of FSR 4, which utilized machine learning-based upscaling.
- They observed gameplay from Ratchet and Clank, comparing the performance of both versions in real-time.
- The presenters noted that the right PC was equipped with an engineering sample of the latest RDNA 4 GPU, which had not been publicly announced.
Image Quality Comparisons
- The presenters conducted a detailed comparison of image quality between FSR 3.1 and the new research version, noting significant improvements in various aspects.
- They highlighted that the research project demonstrated better image quality than FSR 3.1, particularly in handling fine textures like carpet patterns without introducing moiré artifacts.
- The improved image quality was especially noticeable in background objects and effects, with fewer flickering issues compared to FSR 3.1.
Technical Observations and Artifacts
- The presenters noted that while both versions exhibited some artifacts due to the offscreen capture method, the research project appeared cleaner and more stable in person.
- They also commented on the overall sharpness of the image produced by the research project, which was not soft and maintained a high level of detail.
- Specific improvements were noted in particle effects and animation fluidity, particularly concerning confetti and background animations, which had previously suffered from issues in earlier FSR versions.
Conclusions and Future Implications
- The presenters concluded that the new upscaling solution from AMD, although still a research project, showed promising advancements in image quality compared to previous iterations.
- They speculated about the potential relationship between AMD's new upscaler and Sony's PSR technology, noting that while both involve machine learning, they did not observe direct similarities in performance.
- The discussion ended with a light-hearted note about their experience at CES and the anticipation of more announcements and videos related to the event.
AI is already replacing adamsapple those bastards
Who knows? There appears to be little similarity between the observed image we see here with FSR4 and PSSR. Notable, the telltale sign of PSSR in action, the image instability that often looks like fine film grain, is completely absent. Even in the Mark Cerny presentation the ML instructions for the PS5 Pro apparently differ from the WMMA instructions in RDNA, so they diverged there as well. Maybe the separate paths will merge in future, like how CDNA and RDNA is merging into UDNA, which would probably mean UDNA and the PS6 will have dedicated Matrix accelerators.Why would PS make their own if it’s just worse?
They have an amethyst partnership
Who knows? There appears to be little similarity between the observed image we see here with FSR4 and PSSR. Notable, the telltale sign of PSSR in action, the image instability that often looks like fine film grain, is completely absent. Even in the Mark Cerny presentation the ML instructions for the PS5 Pro apparently differ from the WMMA instructions in RDNA, so they diverged there as well. Maybe the separate paths will merge in future, like how CDNA and RDNA is merging into UDNA, which would probably mean UDNA and the PS6 will have dedicated Matrix accelerators.
No, Just that RDNA4 as a whole is way more faster in AI than what the ps5 pro has.Are there costs involved in licensing FSR4 tech?
It could also be a newer revision based off the same foundation of tech.Who knows? There appears to be little similarity between the observed image we see here with FSR4 and PSSR. Notable, the telltale sign of PSSR in action, the image instability that often looks like fine film grain, is completely absent. Even in the Mark Cerny presentation the ML instructions for the PS5 Pro apparently differ from the WMMA instructions in RDNA, so they diverged there as well. Maybe the separate paths will merge in future, like how CDNA and RDNA is merging into UDNA, which would probably mean UDNA and the PS6 will have dedicated Matrix accelerators.
All the previous FSR implementations have been open source and no cost was needed to implement it. Unsure for FSR4. I'd imagine it is a simple explanation, PS5 Pro and RDNA 4 had similar designs but diverged in certain aspects and the PSSR/FSR4 teams were separate and working on different timescales. PSSR needed to be ready 4 months ago, FSR4 is still not out yet. Or Sony just wanted full control over their own upscaling solution and doesn't want it to be open source?Are there costs involved in licensing FSR4 tech?
No, Just that RDNA4 as a whole is way more faster in AI than what the ps5 pro has.
They really invested on that area on this gen.
Why would PS make their own if it’s just worse?
They have an amethyst partnership
Is it supposed to be impressive? Seriously I don’t know shit about AI !I don’t see how it’s faster, ps5 pro has 300 tops
If it's true that PSSR and FSR 4 are different technologies then that would raise a lot of interesting questions as to why both pursued different technologies even though they're in a technical partnership for ML development.
I t may also explain several of Cerny's remarks about the Pro's ML hardware being "custom".
It's pretty good in terms of the number itself, especially relative to the compute performance. The 9070 XT from AMD should have about 390 TOPs and a 4070 Super has about 290 TOPs (both INT 8). But each reach that number differently, WMMA for RDNA, Tensor cores for Nvidia, and PS5 Pro uses custom ML extensions. So It's not as simple as just comparing the number and calling it a day.Is it supposed to be impressive? Seriously I don’t know shit about AI !
Thx!!!It's pretty good in terms of the number itself, especially relative to the compute performance. The 9070 XT from AMD should have about 390 TOPs and a 4070 Super has about 290 TOPs (both INT 8). But each reach that number differently, WMMA for RDNA, Tensor cores for Nvidia, and PS5 Pro uses custom ML extensions. So It's not as simple as just comparing the number and calling it a day.
Looks like a PLAYSTATION3 ad
Im thinking of going AMD for my kid GPU
I can't find the quote now but I think Cerny said Project Amethyst started fairly recently. So most likely FSR 4 and PSSR were developed separately and then at the end of PSSR's development, the two companies decided to come together for future projects relating to UDNA.If it's true that PSSR and FSR 4 are different technologies then that would raise a lot of interesting questions as to why both pursued different technologies even though they're in a technical partnership for ML development.
I t may also explain several of Cerny's remarks about the Pro's ML hardware being "custom".
It’s a different metric. 300 from the Pro is regular INT8, so for the 5090 that would be 1676. Which is probably a bit overkill just for DLSS.Thx!!!
So I guess this bullshit? Or PR!
It is pretty great if you want to run local models on a GPU, alongside wide bandwidth and 32GB VRAM.It’s a different metric. 300 from the Pro is regular INT8, so for the 5090 that would be 1676. Which is probably a bit overkill just for DLSS.
Oh yeah, for personal AI stuff it’s great.It is pretty great if you want to run local models on a GPU, alongside wide bandwidth and 32GB VRAM.
PSSR will also be improved through updates in a few months.My God, PS5 Pro with the worst AI upscaling technology available and that being its selling point for $699, being in raw power just a vitaminised PS5. In a few months any mid/low end graphics card will get better results than PS5 Pro, sad.
"You don't need AI for good upscaling, FSR2/3 is open and it will improve itself soon to be better than DLSS and can be run on every hardware, dedicated hardware like tensor cores are classic Nvidia scam"