Discussion RDNA 5 / UDNA (CDNA Next) speculation

[SolidQ]

FSR Diamond

Wonder if RDNA4 will support improved FSR, except MFG

 

[adroc_thurston]

Wonder if RDNA4 will support improved FSR, except MFG

no lol

 

[SolidQ]

no lol

Nvidia support older cards with Transformer. or it's FP4 FSR?

 

[adroc_thurston]

Nvidia support older cards with Transformer. or it's FP4 FSR?

Nvidia ships the exact same SM for 8 years.
AMD doesn't do that.

 

[SolidQ]

AMD doesn't do that.

and FSR Diamond will be supported via Opticscaler, like FSR4 on RDNA2/3 😉

 

[iLLusiveMan]

FSR Zirconia for RDNA4 😀

 

[adroc_thurston]

and FSR Diamond will be supported via Opticscaler, like FSR4 on RDNA2/3 😉

Nope, it will just not run there.
You need at least 1250 feature level for FSR5.

 

[SolidQ]

AMD about announcing, but nothing add to the games, except FSR4 upscaler. No Radiance cache, no redstone, etc.
Maybe FSR Diamond will last longer, because Xbox next/PS6.

 

[adroc_thurston]

AMD about announcing, but nothing add to the games, except FSR4 upscaler. No Radiance cache, no redstone, etc.

It's not in games because no one buys or cares about Radeon. It's a dead brand.

Maybe FSR Diamond will last longer, because Xbox next/PS6.

Yes, but again, gfx13-only.

 

[dangerman1337]

next generation of DirectX

Defintely seems DX13 is for next-gen GPUs at this point (will Xe3P support it along with SM 7.0?).

4 GB is more like 2028.

The MLID RDNA 5 slide had 4GB SKUs in 1H of 2027 for professional/datacenter. So a late 2027 launch could easily see 4GB modules IMV when production switches/scales to 4 from 3/2GB modules.

 

[adroc_thurston]

(will Xe3P support it along with SM 7.0?).

Yes.
BMG too iirc.

 

[MrMPFR]

I'll post this in the Console thread, too, but I saw this on Xitter:
View attachment 139764

As expected extensive DX_next and RDNA 5 co-design. This is Mantle 2.0: Focus on work graphs, ML and next gen ray tracing pipeline.
DXR 2.0, Cooperative vectors, and SM 7.0. Feature set eclipsing 50 series extremely likely.

Any idea of what GPU directed work graph execution means?

FSR Diamond
- Neural Arrays + nextgen ML pipeline (a lot of stuff from Gfx12.5), complete clean slate = no BWC for RDNA 4 without major porting work (not gonna happen)
- RDNA 4 won't get MFG
- A serious attempt at ray denoising, not Jokestone Ray regen.

Nextgen direct storage, gdeflate replaced by zstd, neural texture compression. Likely we'll see other forms of neural compression added over time.

Very vague but still a bit surprised we get this info this early.

 

[adroc_thurston]

A serious attempt at ray denoising, not Jokestone Ray regen.

You can make it as serious as if you want if you're willing to nuke your frametimes.

 

[1250]

Nope, it will just not run there.
You need at least 1250 feature level for FSR5.

What about Blackwell? Fsr5?

 

[adroc_thurston]

What about Blackwell? Fsr5?

GeForce doesn't run proprietary FSRs.

 

[1250]

I don't know what it is, but there seems to have been a big change in 1250. 소리 ㅋㅋ

 

[1250]

Is there a reason it is not from gfx950?

 

[adroc_thurston]

Is there a reason it is not from gfx950?

Yes, they converged client and DC roadmaps.

 

[1250]

GeForce doesn't run proprietary FSRs.

What I'm curious about is the difference between Blackwell and Rubin (rdna5 and 4 are hopeless)

 

[SolidQ]

Nvidia ships the exact same SM for 8 years.

Rubin same SM's or new?

 

[adroc_thurston]

What I'm curious about is the difference between Blackwell and Rubin (rdna5 and 4 are hopeless)

Idk.
RDNA5 is very nice.

Rubin same SM's or new?

Idk.

 

[1250]

RDNA5 is very nice

1250 too?

 

[adroc_thurston]

1250 too?

Yeah.
AMD makes great GFX IP and then proceeds to do nothing interesting with it.

 

[1250]

I'm just complaining, but Koreans ignore amd. They think Lisa Su is a hbm beggar. For example, they say they are visiting Samsung foundry because tsmc does not accept mi500. Imao

 

[adroc_thurston]

For example, they say they are visiting Samsung foundry because tsmc does not accept mi500

This is so adorable in context of AMD being the first N2 customer (period).