kevboard
Member
XeSS is still not open source (Intel promised to eventually make it open source)... if it was I bet we would have already seen it on Series X or S (would probably be a godsent on Series S)
FSR4 working on RDNA2/3 (unofficially)
- Thread starter Zathalus
- Start date
XeSS is still not open source (Intel promised to eventually make it open source)... if it was I bet we would have already seen it on Series X or S (would probably be a godsent on Series S)
rnlval
Member
Strix Point Ryzen AI series APUs have 50 TOPS XDNA2 NPU. Total system: up to 80 TOPS from Ryzen AI 9 HX 370.Given the source apparently leaked it would be possible to port it. The 50TOPs NPUs in 7840U should already be enough (compute wise) to run this. Caveat being the crippling slow memory on the handhelds - but still.
That said - this leaked version still has ways to go quality wise - I tested it a bit today and while it performs better than FSR3 in some scenarios (and even DLSS) it also completely breaks in others. It's with good reasons it hasn't been released in this state.
see above
7840U is Phoenix Point, featuring a 10 TOPS XDNA NPU. Total system: up to 32 TOPS.
8840U is Hawk Point with 16 TOPS XDNA NPU. Total system: up to 38 TOPS.
Total system TOPS can be higher when the TDP limit is higher.
PaintTinJr
Member
They are not making it up IMO, given the technical difference in the way in which an RDNA4 gpus accelerate FSR4 with latency hiding and how the Pro will need to modify the algorithm to work with minimal latency (less than 2ms) accelerated across the 14MBs of CU register memory the Pro has - similar to the way PSSR currently works -but instead generating FSR4 tensors and applying FSR4's inferencing on the tensors rather than PSSR's.
SKYF@ll
Member
2020 GPUs get FSR4 support before PS5 Pro...
This frame rate is not bad for an RX7700XT (70TOPS/INT8).
I think FSR4 (INT8 version) is quite practical with RDNA3.
I hope PSSR2(FSR4) games for PS5 Pro (300TOPS/INT8) will be released sooner than expected.
Possibly because in current state - this version is not what I'd call 'ready to release'.
Last edited:
rnlval
Member
Extra R&D is needed for the FSR4 INT8 version since it's different from the FSR4 FP8 version. Running FSR4 FP8 on RDNA 3's FP16 wasn't optimal when compared to the native INT8 version.
Remember, AMD recently released Radeon RX 7700 16 GB VRAM edition. https://videocardz.com/newz/amd-launches-radeon-rx-7700-with-2560-cores-and-16gb-memory
The Radeon RX 7700 16 GB VRAM edition's price is lower than RX 9060 XT 16 GB VRAM.
Many of AMD's newer chip designs (e.g. Granite Ridge, Turin, Strix Point, Strix Halo, Kraken Point, NAVI 48, NAVI 44) are on a 4 nm process node with certain production capacity. Radeon RX 7700 16 GB VRAM edition uses the older 5 and 6 nm process nodes.
Felessan
Member
They will probably get it around the same time, when INT8 model will be decided to be working and ready to release.
Because standard FSR4 require things (FP8) that RDNA3 doesn't not support. And leaked beta version is a fork in progress to make it work on older cards.
Like DLSS3 was 4000 exclusive and DLSS 3.5 supports older cards - it takes time to backport and adjust some things.
DirtInUrEye
Member
Somehow with laggier inputs.
RobRSG
Member
For non-competitive games, it is perfectly fine for me. Give it a try, probably worth your time
DenchDeckard
Moderated wildly
Just wanna say. FSR 4 is great. I've been so impressed with mine in my front room console replacement pc. Super surprised how well it holds up Vs DLSS. Keep working AMD.
Panajev2001a
GAF's Pleasant Genius
AMD and… Sony/Cerny
. Come on, you have done so many little steps forward… one more 
!
jm89
Member
Not the technical details. I meant Sony reaching out to them. Wccftech tech isn't the type of outlet Sony usually reaches out to.
Last edited:
DenchDeckard
Moderated wildly
AMD and… Sony/Cerny
Honestly its great, just gotta keep pushing. im playing Space Marine 2 at the minute with a couple of mates. Im maxed out settings, everything cranked. 4k FSR4 no frame gen and getting 90 - 110 FPS on my 65 OLED. I have a 4090 upstairs and the experience is more or less identical with VRR.
PaintTinJr
Member
I suspect the credibility of wccftech is too high - and would get cited by others like DF- so Sony would have got in touch just to make sure the disinformation that might disappoint confused Pro consumers(thinking they were getting FSR4 called FSR4) didn't gain traction.
Last edited:
dgrdsv
Member
DLSS3 never was "4000 exclusive", only the FG part was (and still is, now for 40+50).
The problem AMD faces with FSR4 is two fold:
1. They need to port it to the assortment of their previous gen GPUs most of which have varying feature support for 8/16 bit matrix math.
2. They still sell these parts (and even launch new ones on RDNA3! CPUs also have RDNA3/3.5 in them) which means that running a more complex upscaling on them will make them perform worse than they do now with FSR3. This will affect their selling power and may result in a need for another price drop.
jm89
Member
He might burst into flames if he doesAMD and… Sony/Cerny
PaintTinJr
Member
That depends on a few things.
Currently FSR3 needs a higher quality native render to produce higher resolution with higher frame-rates on RDNA2/RDNA3 .
So provided the performance saving(in milliseconds) in dropping to much lower native and the effective gain in upscale quality using FP16 instead of INT8/INT4 in a hybrid FSR4 could offset the multiple fold loss in throughput of FP16 compared to INT8/INT4 for upscaling, then the overall outcome could still be superior compared to FSR3, and this theoretically could be true for the PS5 if the register memory from the 36CUs was adequate.
As an example, say the PS5 with FSR3 runs 60fps with native res of 1920x1080 and takes 2ms. Provided an FSR4 implementation on PS5 using FP16(RPM) can do a FMA per Half Top (so 60 TOPs = 20HTF/s x 3 Ops), then the difference between the Pro targeting 4K60 from native 1080p60 with 300 TOPs(as 2x FMA per INT8/FP8) in 1 millisecond vs 5x less TOPs (300TOPS vs 60TOPS) if the PS5 targeting a quality upscale of 1080p60 from - say - 960x540@60 could free up (5x) 5 milliseconds.
Obviously the end result would need to be a superior fidelity 1080p60 via FSR4 vs 4K60 via FSR3, and I've completely ignored the PS5 Pro custom instructions to halt and switch from rendering to upscaling, meaning the x5 might be 1ms more.
Similarly PCs with RDNA2/RDNA3 GPUS they will need to do something similar with the GPU's LLC being the other factor after the TOPs comparison and still hope on balance the fidelity is superior to FSR3 from lower native res and lower output res.
Edit: maybe I've mixed and matched Ops/sec with Pro vs Ps5 calc comparison and it still needs multiples more, probably needed a drop to 30fps too.
Last edited: