Discussion RDNA 5 / UDNA (CDNA Next) speculation

[marees]

Which it will be en masse for PS6

PS6 would be RDNA 5+ tho

The Xbox is rumored to be identical to the upcoming 10070 xt

 

[soresu]

Which it will be en masse for PS6

Unless you mean the SDK for the platform there's no guarantee of that.

For example, PS4 Pro had rapid packed math, but I'd wager not that many games actually put it to use until UE4 or Unity implemented it "off the shelf" as it were.

 

[Win2012R2]

For example, PS4 Pro had rapid packed math

Compared to quality upscaling that is becoming totally necessary that feature had zero use, from memory Far Cry 5 claimed to use it, clearly sponsored attempt, plus PS4 Pro had very small installment base in the first place to bother.

I've just played Mafia Old Country on 9800X3D + 4090 at 4k - used quality level DLSS upscaling (so 1440p), total rubbish perf (and this isn't even open world), given profiliration of Unreal Engine 5 games in the next 5 years they will all need aggressive upscaler on consoles, almost nobody will be rendering at native 4k on PS6, will be lucky if it's 1080p.

Unless you mean the SDK for the platform there's no guarantee of that.

Well quality upscaling will certainly be default feature in SDK for PS6, by that time the dust should settle nicely for this "ai" upscaling malarkey and stable solution will be avaliable.

PS6 would be RDNA 5+ tho

Same thing really though? I hope Sony makes AMD add dedicated hardware decompressor

 

[Magras00]

Compared to quality upscaling that is becoming totally necessary that feature had zero use, from memory Far Cry 5 claimed to use it, clearly sponsored attempt, plus PS4 Pro had very small installment base in the first place to bother.

I've just played Mafia Old Country on 9800X3D + 4090 at 4k - used quality level DLSS upscaling (so 1440p), total rubbish perf (and this isn't even open world), given profiliration of Unreal Engine 5 games in the next 5 years they will all need aggressive upscaler on consoles, almost nobody will be rendering at native 4k on PS6, will be lucky if it's 1080p.


Well quality upscaling will certainly be default feature in SDK for PS6, by that time the dust should settle nicely for this "ai" upscaling malarkey and stable solution will be avaliable.


Same thing really though? I hope Sony makes AMD add dedicated hardware decompressor

The raster purists will have a meltdown nextgen. Things will only get more AI and upscaled.

DGF is comfirmed in future HW (RDNA 5) as per AMD's introductory GPUOpen Blog. Photo for RDNA 5 HW DGF compressor in my prev patent post (see the imgur pictures) if you want to know how they'll do it.

 

[soresu]

The raster purists will have a meltdown nextgen. Things will only get more AI and upscaled.

tbf most of that AI is there because the hardware still isn't for RT/PT.

That being said, I feel like upscaling of all things is the worst because it has far more potential to distort the original artistic intent of the game/mod devs.

nVidia love to tout this ridiculous standard of "better than native rendering" for DLSS - which is just another way of saying it's not rendering as the game devs intended it to.

By next gen I think with all the various stuff like NRC/extensions + GATE + ReSTIRapalooza + path guiding we should be at a point where hi end to at least upper mid end GPUs shouldn't need upscaling just to keep it playable even with path tracing.

Definitely by next next gen unless nVidia get desperate enough to prod game devs into spamming volumetrics just to add moar compute complexity because they can.

 

[adroc_thurston]

The raster purists will have a meltdown nextgen

You're literally getting another massive shader core update focused on ISO-clock perf.

tbf most of that AI is there because the hardware still isn't for RT/PT

It will never be 'ready' since it's a workload you don't wanna do anywhere but CPUs.

By next gen I think with all the various stuff like NRC/extensions + GATE + ReSTIRapalooza

All that rewarmed garbage only to get blurry 0.5spp visuals.
Nifty.

 

[soresu]

All that rewarmed garbage only to get blurry 0.5spp visuals.

Reduce spatial or temporal resolution then.

It will never be 'ready' since it's a workload you don't wanna do anywhere but CPUs.

Plenty of GPU backends for production offline renderers say otherwise.

 

[adroc_thurston]

Reduce spatial or temporal resolution then.

I don't think people want *more* blur.

Plenty of GPU backends for production offline renderers say otherwise.

That's not an excuse for running pointer chasing competitions on hardware least optimized for gmem latency.

 

[Magras00]

You're literally getting another massive shader core update focused on ISO-clock perf.

Was not referring to RDNA 5 but how devs will use PS6 and project amethyst. FSR4 already very good so don't think devs will target 1440p to 4K at 60FPS. 1080p to UHD more likely, maybe even sub full high definition depending on how the upscaler evolves the rest of this decade. More raster and better upscaling will just allow devs to push things even further.

Your comment about samples per pixel is triggering spam filter but I'll still try to reply: All that neural rendering and path tracing optimization fine will enable devs to do more with less. NRC, GATE, Intel's Histogram Stratification for Spatio-Temporal Reservoir Sampling technique, NIS, path guiding etc...

I really don't care how devs achieve their goals as long as the tech underlying it all is solid and it'll be nextgen. We're in nascent stage rn.

 

[Win2012R2]

it'll be nextgen. We're in nascent stage rn.

Nah, it will be cross gen with PS5 until at least early 2030-s

 

[basix]

Many of the new techniques can be used on PS5 as well. Most of the new tech in e.g. UE5 runs fine on a base PS5. Much of the research stuff of Intel and AMD runs quite well on RDNA2 (much is compute shader based).

The only exception is stuff with very high HW demands for RT and WMMA. But many of the technological advancements do not completely rely on that or can be stripped down. Or sometimes a "last gen technology" like FSR2/3 or TSR is available as replacement (altough with reduced quality).

Let's make an example with e.g. UE5: RT + Temporal Upsampling + RR + FG

• PS5 uses standard HWRT and TSR
  • 60fps with 720...900p -> 1440p (TSR) -> Spatial Upscaling to 4K with HWRT (TW4 Showcase Specs)
• PS6 uses HWRT with higher fidelity, Ray Reconstruction and FSR4/5
  • 60fps 1080p -> 4K (FSR4+) with Path Tracing (higher fidelity and resolution compared to basic HWRT) and Ray Reconstruction
  • Frame Generation to 120fps
  • Generally more details (more objects, grass, people, range of view, ...)

The base design requirements for the game are the same. It just looks much better on PS6.

An interesting topic would be Neural Rendering on RDNA2. Neural Radiance Caching, Neural Texture Compression, ...

• Unclear, how well this would run without WMMA acceleration
• Could have effects on developer workflows and art creation (e.g. Neural Texture Compression)

 

[Win2012R2]

UE5 runs fine on a base PS5

UE5 does not run fine on anything right now...

 

[marees]

UE5 does not run fine on anything right now...

Senua's sega Hellblade 2 was the showcase game for Xbox

Are you referring to the Nvidia RT branch ?

 

[basix]

UE5 does not run fine on anything right now...

Yeah, that is right. See the latest Mafia game release. But UE 5.6 is a big step into the right direction. And what has been announced for UE 5.7 goes into the same direction.

I mean, I was speaking of a time when PS6 gets released. Until then there should be many improvements to UE5. Especially out-of-the-box performance wise. And with that all in place I do not see too many or even blocking restrictions, which would hold back technical progress on PS6. PC as as a PS6 sidekick will also help. Everything from RDNA4 and Lovelace onwards will support most of the important features.

That it took 5 years from original UE5 show case (May 2020) to UE 5.6 (June 2025) / 5.7 (end of 2025) release to more or less complete the compute based rasterization approach with performant HWRT on top is a pity. But even with the resources of Epic it seemed to be a difficult task. Could Epic improve performance earlier? I don't know. All release notes from 5.0 to 5.6 state performance improvements of various sorts.

 

[adroc_thurston]

Was not referring to RDNA 5 but how devs will use PS6 and project amethyst.

They're gonna use it like you would use any RDNA5. Aka more oomph per shader core.

All that neural rendering and path tracing optimization fine will enable devs to do more with less.

You're not doing more with less, we're trying to eke the bare minimum here since the hardware can't run RTRT at all.

Everyone's just bolting on crutches since the inherent assumption of RTRT being fast enough eventually was just outright not true. It'll never be fast enough.

 

[soresu]

Everyone's just bolting on crutches since the inherent assumption of RTRT being fast enough eventually was just outright not true. It'll never be fast enough.

Not for full screen rendering anyway.

For eye tracked foveated rendering in NED's there's some small chance if the renderer does multiple SPP for the fovea and increasingly sparse 1 spp for the periphery.

It's still crutches technically, but given the visual system isn't even "full resolution" beyond the fovea it is at least anatomically justifiable crutches.

 

[adroc_thurston]

For eye tracked foveated rendering

How the hell are you gonna eye-track outside of VR.

 

[soresu]

How the hell are you gonna eye-track outside of VR.

Read that whole line again.

 

[soresu]

Also you can eye track outside of VR, it's just useless for foveated rendering.

IIRC some 3D volumetric display designs work by gaze tracking to fake volumetric rendering with just stereoscopic imagery shifted according to viewer position.

 

[Magras00]

They're gonna use it like you would use any RDNA5. Aka more oomph per shader core.

Yeah but they're not going to use extra raw raster power that to increase internal res on PS6 (baseline for all games post crossgen). FSR4 will only get better. Transformer SR is nowhere near tapped out. End result for devs = crank graphics up to 11 + keep internal res at or below 1080p.

Everyone's just bolting on crutches since the inherent assumption of RTRT being fast enough eventually was just outright not true. It'll never be fast enough.

If you mean RTRT without tricks then yes. Neural rendering, ML SR + RR, and various optimizations for PT will enable close enough graphics. That will still be so ahead of PS5 probe based RTGI that it'll be distinctly nextgen.
Crutches and fake graphics only going to get more widely used nextgen.

 

[adroc_thurston]

Transformer SR is nowhere near tapped out.

yeah it is lmao.
Unless you have some frametime to spare.

Neural rendering, ML SR + RR, and various optimizations for PT will enable close enough graphics

More noise and more blur? sure why not.

That will still be so ahead of PS5 probe based RTGI that it'll be distinctly nextgen.

I always wanted 'real-time' lighting implementations to have 240+ frames of accumulation lag.

 

[soresu]

Transformer SR is nowhere near tapped out

#1. FSR4 neural super resolution doesn't use a transformer model.

It uses a CNN I believe.

#2. Transformers are not devoid of compute/memory cost relative to CNNs.

That will still be so ahead of PS5 probe based RTGI that it'll be distinctly nextgen.

No reason any non NR gfx software techniques can't be employed on PS5/Pro too.

PS5 OG is limited by RDNA2's far inferior matrix + ML ops handling relative to RDNA4 in PS5 Pro, so any NR techniques employed are likely to be hidebound to the latter console, increasing its attractiveness prior to PS6's release, so long as game devs actually take advantage of the hw feature set.

In the case of something like Doom The Dark Ages path tracing mode it would basically be a necessity to run it reasonably well, likely with FSR Redstone's full capabilities.

 

[Josh128]

Y'all seen this? Rubin GPU arc is getting a delay due to a supposed "reworking" due to expected competition from AMD MI 400/450 series. Sounds like some rubbish to me, but maybe theres something to it.

NVIDIA’s Rubin AI Architecture Is Reportedly Delayed, as the Firm Has Been Influenced by AMD and Is Now Redesigning to Compete with the Instinct MI450 Lineup

NVIDIA's next-gen Rubin lineup might be delayed by at least a few months, as Team Green is redesigning it to give AMD tougher competition.

wccftech.com

 

[SolidQ]

FSR4 neural super resolution doesn't use a transformer model.

Hybrid CNN/Trans

Rubin GPU arc is getting a delay due to a supposed "reworking" due to expected competition from AMD MI 400/450 series.

Interesting. Maybe Huang also gonna rework RTX 6xxx to Compete with RDNA5, but anyway need comfirmation

 

[Win2012R2]

But UE 5.6 is a big step into the right direction. And what has been announced for UE 5.7 goes into the same direction.

The (un)real problem is that apparently it's not easy to upgrade the engine even within the same major version, and the issue isn't about already released games that won't bother upgrading due to extra cost, but the fact that it takes 5-7 years to get big games done these days, so games that are coming out now were using 5.3-4 as baseline at best.

Tried Mafia Old Country using 4090 yesterday, runs like **** at 4k using DLSS quality upscale, which frankly should not be necessary on 4090, which is what - at least x5 than PS5?

Are you referring to the Nvidia RT branch ?

Nah, anything RT is expected to run bad even on 4090.