Support NeoGAF

[kevboard]

they test Quake 2 RTX, which runs really well, well enough for an official port being viable.

Portal RTX, looks like ass, but only due to the lack of a decent upsampler

Cyberpunk, runs surprisingly good when using the optimisation mod to reduce light bounces to 1. would need a decent upsampler tho as FSR3 looks apparently so bad that they don't even bother benchmarking it.

 

[cormack12]

Is this cos Shu cant say PlayStallion

 

[kevboard]

oh ffs... can a mod fix the title? damn touchscreen keyboards

 

[viveks86]

Surprisingly good performance on Cyberpunk. Thought it won't run at all, even at potato cam resolutions. Bodes well for next gen.

 

[Gaiff]

oh ffs... can a mod fix the title? damn touchscreen keyboards

Done.

 

[Guilty_AI]

Goddamn, i thought we finally had Platypus 5

 

[Kuranghi]

Nah keep the title its very funny. I wish we could leave voice notes on our posts, you'd be getting one of me saying that out loud in various accents for sure.

 

[clarky]

Never knew there was a silent r in Plarstayshion

 

[kevboard]

the mods are comedians now are they? very funny guys!

 

[R6Rider]

 

[RaidenJTr7]

Plarstayshion 5

 

[Imtjnotu]

I will never call it PlayStation ever again

Long live Plarstayshion

 

[Venom Snake]

 

[666]

The new Revelaitons and Arkham Oranges

 

[Gitaroo]

the big thing here is XeSS running on PS5, I thought PS5 doesn't have dp4a support so it's not possible to run any ML upscaler on base PS5. Said so DF themselves back then but this video says otherwise

 

[kevboard]

the big thing here is XeSS running on PS5, I thought PS5 doesn't have dp4a support so it's not possible to run any ML upscaler on base PS5. Said so DF themselves back then but this video says otherwise

it's running extremely poorly because RDNA1 has to emulate it fully in software.

they said in the video that they get the same performance with FSR3 at much higher internal resolutions... it's just that FSR3 is absolutely disgusting at low resolutions, so the even far lower XeSS resolution still looks better.

for example, if you use XeSS quality mode on RDNA1, your performance will actually drop below your native resolution performance. while on RDNA2 it doesn't, although it also won't give you a big performance boost either when compared to FSR3.


in short, RDNA1/Navi 10/the PS5 can run XeSS and FSR4 int8 as the shader model supports it, but they don't have any hardware acceleration for it, which will mean really bad performance.
RDNA2/Series X|S has hardware acceleration for it, although not the best, and there XeSS and FSR4 int8 will run ok-ish

 

[DonF]

the r is silent

 

[Quixz]

Clicked the thread due to the title. I vote it stays

 

[Guardian Monkey]

the r is silent

Plarstayshion

-My preferred pronunciation.

 

[vkbest]

they said in the video that they get the same performance with FSR3 at much higher internal resolutions... it's just that FSR3 is absolutely disgusting at low resolutions, so the even far lower XeSS resolution still looks better.

The video only says FSR3 is faster, but that is true even on PC.

 

[Skifi28]

And I was 100% sure it was a Draugoth thread before clicking.

 

[kevboard]

The video only says FSR3 is faster, but that is true even on PC.

they say in the video that when upsampling to 1080p, XeSS at 348p has worse performance than FSR3 at 540p
this is an internal resolution boost for FSR3 of around 2x while having increased performance. that's a really big difference. on RDNA2 the perfomance difference is far smaller, less than 20% usually.

they used XeSS at that super low resolution because FSR3 just completely shits the bed the moment your internal res dares to be lower than 1080p, due to how bad of an upsampler it is.

 

[Black_Stride]

the big thing here is XeSS running on PS5, I thought PS5 doesn't have dp4a support so it's not possible to run any ML upscaler on base PS5. Said so DF themselves back then but this video says otherwise

Pretty much anything that can do Int 8 can run XeSS.....itll just run like shit.

 

[MCH2024]

they test Quake 2 RTX, which runs really well, well enough for an official port being viable.

Portal RTX, looks like ass, but only due to the lack of a decent upsampler

Cyberpunk, runs surprisingly good when using the optimisation mod to reduce light bounces to 1. would need a decent upsampler tho as FSR3 looks apparently so bad that they don't even bother benchmarking it.

There can be no decent upscaler on PS5 if it has no Tensor cores or even DP4a.

Also K KeplerL2 , if PS5 didn't have DP4a how come they are running DP4a XeSS on it with decent perf?

 

[kevboard]

Also K KeplerL2 , if PS5 didn't have DP4a how come they are running DP4a XeSS on it with decent perf?

decent performance? they had to run at half the resolution it could run when using FSR3 and still didn't reach the same performance. XeSS on Navi 10 runs emulated and has insane performance costs. it's just that XeSS still looks better than FSR3 even when the internal res is 50% lower lol

 

[vkbest]

they say in the video that when upsampling to 1080p, XeSS at 348p has worse performance than FSR3 at 540p

What second of the video they says that?

 

[Bojji]

There can be no decent upscaler on PS5 if it has no Tensor cores or even DP4a.

Also K KeplerL2 , if PS5 didn't have DP4a how come they are running DP4a XeSS on it with decent perf?

5700XT owners can run XeSS but it performs like crap.

 

[kevboard]

What second of the video they says that?

around the 17min mark. they say they run XeSS at 348p, and it runs worse than FSR3 Performance mode, which would be 540p

 

[winjer]

the big thing here is XeSS running on PS5, I thought PS5 doesn't have dp4a support so it's not possible to run any ML upscaler on base PS5. Said so DF themselves back then but this video says otherwise

DF says a lot of nonsense.
There are 2 ways of running XeSS on a GPU without FP4A support.
One is through Int8 emulation on Linux through FP16. But this is quite heavy. Probably taking twice the upscaling time.
The other way is though the 3rd path that XeSS supports, SM6.4+ This is the one that looks worse and runs worse.

 

[Gitaroo]

DF says a lot of nonsense.
There are 2 ways of running XeSS on a GPU without FP4A support.
One is through Int8 emulation on Linux through FP16. But this is quite heavy. Probably taking twice the upscaling time.
The other way is though the 3rd path that XeSS supports, SM6.4+ This is the one that looks worse and runs worse.

wonder if the performance cost is still worth it if your priority is high fidelity at 30fps.

 

[winjer]

wonder if the performance cost is still worth it if your priority is high fidelity at 30fps.

Very unlikely it's worth it.

 

[KeplerL2]

There can be no decent upscaler on PS5 if it has no Tensor cores or even DP4a.

Also K KeplerL2 , if PS5 didn't have DP4a how come they are running DP4a XeSS on it with decent perf?

XeSS has a fallback path for non-DP4a GPUs

 

[MrCunningham]

It honestly amazes me how far behind so many tech channels are on Linux. Same goes for GamersNexus, even though they dud make some strides. This comes from someone who has been using Linux as their daily driver since 2007 and had the first versions of Proton on the beta builds of Steam back in 2018. The Steamdeck gave them all a taste of what Proton and SteamOS is capable of, with honestly really incredible results. Shame that the Steamdeck is still out of stock. Valve has been kicking ass on Linux for quite a while now, to the point where they made it viable as a gaming platform. Which is a far cry from the days of having to compile binaries to run games like UT2004, and having to carefully work your way around the pitfalls on installing Steam through winetricks and then doing double duties of trying to make said game work within Steam.

It's all compatibility layers that have been propping up Linux as a gaming platform. But at this point, Windows 11 honestly uses a lot of compatibility tricks to run Windows software. The number of native Linux game software is still low. But companies like Feral Interactive who did put a lot of emphasis on Linux software quite a few years ago could never really keep up to date with the windows branches of games when it came to updates. Of course, one of the pitfalls of running games in proton are the deeply rooted anti-cheat toolkits that seem to cause headaches for those who just want to play some games in virtualization.

Given the AMD nature of the PS5 Pro, It is neat to see how well things do run on a Steam compatibility layer. This is not even optimized for the custom PS% GPU. The Mesa opensource drivers handle things really well.

 

[PaintTinJr]

...


in short, RDNA1/Navi 10/the PS5 can run XeSS and FSR4 int8 as the shader model supports it, but they don't have any hardware acceleration for it, which will mean really bad performance.
RDNA2/Series X|S has hardware acceleration for it, although not the best, and there XeSS and FSR4 int8 will run ok-ish

The video is completely misleading on this front with regards to performance IMO. Any native solution for an upscaler on PS5 by Sony is going to async the inferencing, use specific AMD ISA instructions that are atleast a x2 if not x4 factor improvement, and balance the memory use in a way that this generic PC running so much indirectly gets nowhere close to.

If anything the Quake 2 PT experience with its tiny memory use, cheap base raster workload probably reflects how custom dev efforts on PS% would perform using lower native with higher RT on AAA games, and then use PSSR2/FSR4.1.

Although to be fair, this DF video trashing the possibility of the PS5 using FSR4.1/PSSR2 and talking rubbish about the Pro needing to run at 540p (when a Steam deck uses less than 1TOP for FSR4 320p->720p) is actually the kick up the arse PlayStation has been missing this generation.

If this is really how toothless the PS5 is going to finish the gen, then Sony need to start work on a AI HAT (20 TOPS) pass through module that sits in the NVME expansion bay with a secondary NVME SSD IMO.

 

[Parazels]

We desperately need arcade ports with path tracing. Why do developers miss the opportunity?

 

[Tajaz2426]

Started my day with this thread, did not disappoint.

Good job, men of the world.

 

[winjer]

The video is completely misleading on this front with regards to performance IMO. Any native solution for an upscaler on PS5 by Sony is going to async the inferencing, use specific AMD ISA instructions that are atleast a x2 if not x4 factor improvement, and balance the memory use in a way that this generic PC running so much indirectly gets nowhere close to.

If anything the Quake 2 PT experience with its tiny memory use, cheap base raster workload probably reflects how custom dev efforts on PS% would perform using lower native with higher RT on AAA games, and then use PSSR2/FSR4.1.

Although to be fair, this DF video trashing the possibility of the PS5 using FSR4.1/PSSR2 and talking rubbish about the Pro needing to run at 540p (when a Steam deck uses less than 1TOP for FSR4 320p->720p) is actually the kick up the arse PlayStation has been missing this generation.

If this is really how toothless the PS5 is going to finish the gen, then Sony need to start work on a AI HAT (20 TOPS) pass through module that sits in the NVME expansion bay with a secondary NVME SSD IMO.

Yes, DF lost a lot of performance by using XeSS. They could probably run CP2077 and Portal at a significantly higher resolution and then have FSR3.1 do the upscaling.

 

[Guilty_AI]

We desperately need arcade ports with path tracing. Why do developers miss the opportunity?

Cause most will look weird. Works with quake 2 because its baked lighting already accounted for bounces and other real-world properties, but games designed around simple lighting tech just look weird with it.

 

[Parazels]

Cause most will look weird. Works with quake 2 because its baked lighting already accounted for bounces and other real-world properties, but games designed around simple lighting tech just look weird with it.

 

[Guilty_AI]

Lighting and darkness work is one of the most important things to work on when wanting 3D games to look good. To just slap ray tracing into a game that wasn't made with it in mind without any further consideration creates scenes that don't really look right.

 

[Zathalus]

Yes, DF lost a lot of performance by using XeSS. They could probably run CP2077 and Portal at a significantly higher resolution and then have FSR3.1 do the upscaling.

The tried CP2077 with FSR and performance was better, and the internal resolution was higher, but it looked way worse.

 

[winjer]

The tried CP2077 with FSR and performance was better, and the internal resolution was higher, but it looked way worse.

Let me guess, with the FSR2 that CP2077 implemented and completely fucked up?

 

[Zathalus]

Let me guess, with the FSR2 that CP2077 implemented and completely fucked up?

I'd imagine FSR3 as that is what the game has.

 

[winjer]

I'd imagine FSR3 as that is what the game has.

FSR3.0 is just FSR2.2 + FG
The game eventually implemented FSR3.1 and somehow it looks even worse than FSR2 that it had before.
But use Optiscaler and it looks a lot better. Mind you it's still FSR3.1, but it's miles ahead of the fuck up that CDPR did with their implementation.

 

[PaintTinJr]

The tried CP2077 with FSR and performance was better, and the internal resolution was higher, but it looked way worse.

Oliver went on and on about TOPS of the various RDNA cards and the Pro, and even throwing shade with XsS, completely failing to realise just how light FSR4 is on the SteamDeck for a upscale close enough to the level he was suggesting for the Pro, which would be the level the PS5 would need, and probably cost less than 4 theoretical HFLOPs of the PS5's 20HFLOPs the analysis is perfect for getting PlayStation to actually put in some SDK work this gen for the OG PS5.

 

[Zathalus]

Oliver went on and on about TOPS of the various RDNA cards and the Pro, and even throwing shade with XsS, completely failing to realise just how light FSR4 is on the SteamDeck for a upscale close enough to the level he was suggesting for the Pro, which would be the level the PS5 would need, and probably cost less than 4 theoretical HFLOPs of the PS5's 20HFLOPs the analysis is perfect for getting PlayStation to actually put in some SDK work this gen for the OG PS5.

Where was he talking about TOPs in this video? The only comparison I can find is talking about the 9060 XT and that was just in the context of RT performance being comparable to the PS5 Pro.

 

[PaintTinJr]

Where was he talking about TOPs in this video? The only comparison I can find is talking about the 9060 XT and that was just in the context of RT performance being comparable to the PS5 Pro.

It is from around 24mins 34mins IIRC

 

[Wolzard]

Given the AMD nature of the PS5 Pro, It is neat to see how well things do run on a Steam compatibility layer. This is not even optimized for the custom PS% GPU. The Mesa opensource drivers handle things really well.

Patches have been made for the PS5 GPU in the Mesa driver. It's not like it's anything out of this world, it's an AMD GPU (GFX1013) like any other, just more limited.

Mesa & AMDGPU Linux Driver See Patches For The Sony PS5 GPU - Phoronix

www.phoronix.com

PS5 games aren't different from PC games either, they run on an AMD computer with FreeBSD on top. What you see as advantages or customizations is simply the benefit of developing exclusively for one platform, you can dedicate more development time to covering gaps.

 

[winjer]

It is from around 24mins 34mins IIRC

That part really showed how little DF understands about tech and upscaling.

 

[Zathalus]

It is from around 24mins 34mins IIRC

I'm not seeing anything substantially wrong with what he said. He rightfully points out that the Pro has way better ML capabilities than even the 7900 XTX and that the leaked version of FSR4 was pretty heavy on RDNA3 and very heavy indeed on RDNA2, all quite true. You bring up the Steam Deck, but the Deck operates at such low resolution and FPS that it's pretty much irrelevant as a benchmark for RDNA2 GPUs and consoles. For example a 6750 XT at 1440 FSR4 Performance (so 720p) only gets a performance uplift of 10-30 percent. Better than nothing sure, but hardly ideal, especially since 720p->1440p with FSR 4.1 is not quite up to par.

How this relates to path tracing, they also said Path Tracing with PSSR2 would probably work pretty well on the Pro, firstly because the Pro is significantly faster than the PS5, and that PSSR2 is a very good upscaler, much better than what they were testing with on the base PS5.

That part really showed how little DF understands about tech and upscaling.

What exactly did they say that was in error?