Discussion RDNA 5 / UDNA (CDNA Next) speculation

[luro]

They intend to drop support of rdna2 while those GPUs still on shelf and all the rdna lines that should support ROCm still has serious issues running it

It's actually a lot worse than that.

AMD dropped support for RDNA3/3.5 which is an architecture coming in future products like Gorgon Halo.

At the moment, there's hardly any reason to trust AMD's GPU support at all.
Buyers of recent RDNA4 GPUs are likely to regret their decision in a couple of years when RDNA5 is out there.

Their latest driver literaly has a bunch of fixes for RNDA2 and RDNA3. I dont know how this is not supporting.

 

[ToTTenTranz]

I don't know about ROCm, but at least the driver thing AMD somewhat backtracked from was "only" about RDNA1+2, and more about Day 1 support + optimizations for new games, rather than total droppage.

RDNA3 is in "maintenance mode" just as much as RDNA1+2. FSR4 and Redstone were exclusive to RDNA4 and there's zero sign of this changing.

Their latest driver literaly has a bunch of fixes for RNDA2 and RDNA3. I dont know how this is not supporting.

Yes, fixes means "maintenance mode". Support should mean game optimization and support for newly launched software features.

 

[gdansk]

The RTX 4080 is now faster than a 7900xtx and the gap between a 7900 xtx vs 4090 is nearly 39. percent(where it used to be 29%). The 4070 regular is now beating a 7800xt where it previously lost.

The test suite changed. Even something as simple as using a different, faster CPU will favor Nvidia's CPU heavy drivers.

And other test suites disagree with that assessment.

Go spread your multivariable schlockbenches/FUD in another (non-vendor) thread. RDNA3.x isn't in maintenance mode despite not having all features. Much like Ampere isn't in maintenance mode despite not having MFG.

 

[Win2012R2]

Yes, fixes means "maintenance mode". Support should mean game optimization and support for newly launched software features.

It's unrealistic to expect any of the following:
1) new software features like FSR4 that rely on new hardware added
2) doing anything other than maintenance mode for product lined that EOL-ed
3) doing game optimisation in drivers - drivers should have never done that, it's the job for devs to optimise their games that they sell: driver should just be driver that complies with the spec and does it as fast as possible.

 

[gdansk]

it means AMD is just doing the bare minimum to ensure compatibility and stability

No, it doesn't. Your multivariable tests show nothing of the sort.

You can easily look at AMD commits to llvm forks and linux in the last year to show it isn't in maintenance mode. They had made changes to their compilers that we can see in public for older architectures. Your worthless theory crafting ignoring architectural differences, testing setup variation and game selection is falsified easily by looking at real, shipping code which improves performance on e.g. gfx11.

 

[marees]

RDNA 5 (AT2 — 10070xt) & Xbox Magnus ready to go by Q3 2027 — an MLID source who has never been wrong

 

[basix]

The test suite changed. Even something as simple as using a different, faster CPU will favor Nvidia's CPU heavy drivers.

And other test suites disagree with that assessment.

If you look at the raytracing test suite of Computerbase, the 7900XTX gained ground against the 4080. We now can argue that Nvidia has worse support for raytracing games

 

[DAPUNISHER]

Go spread your multivariable schlockbenches/FUD in another (non-vendor) thread.

Precisely. The deleted posts would have been fine in any vendors thread, except for AMD.

Apologies for the interruption.

 

[ToTTenTranz]

It's unrealistic to expect any of the following:
1) new software features like FSR4 that rely on new hardware added

This goes out the window the moment you know AMD has a working version of FSR4 using INT8 that is a lot better than FSR3.x and runs fine on RDNA3 and even RDNA2.

2) doing anything other than maintenance mode for product lined that EOL-ed

FSR1 was officially supported by all GCN1 GPUs (released 8 years prior). FSR3 is officially supported by Vega GPUs (released 6 years prior) for upscaling.
FSR4 isn't supported by RDNA3.x GPUs that released 2 years prior. Some of the RDNA3.5 iGPUs released last month and some of them are still yet to release.

This level of planned obsolescence from AMD, considering the fact that the hardware is capable, is unprecedented.

3) doing game optimisation in drivers - drivers should have never done that, it's the job for devs to optimise their games that they sell: driver should just be driver that complies with the spec and does it as fast as possible.

Wanting this to happen doesn't make it happen, though.

 

[jpiniero]

This goes out the window the moment you know AMD has a working version of FSR4 using INT8 that is a lot better than FSR3.x and runs fine on RDNA3 and even RDNA2.

That's a lot slower that AMD decided wasn't worth it.

 

[adroc_thurston]

That's a lot slower that AMD decided wasn't worth it.

New DLSS update is also horribly slow on non-FP8-capable GPUs.
It's whatever at this point.

 

[luro]

Yes, fixes means "maintenance mode". Support should mean game optimization and support for newly launched software features.

No, means they are still supporting it. Its just that they are not prioritizing them anymore. Oh, and Nvidia does the same, they just dont't make it clear.

Edit: Won't discuss semantics. If we are going that way, Nvidia also dropped support for older generations by not providing FG.

 

[tsamolotoff]

even RDNA2.

If we imagine that AMD actually cared about GPUs, there was a rationale for AMD not to launch FSR4 for older GPUs performance wise. FSR4 takes a bit of time even at 1080p with full blown N21, more than 1ms and that rises to 4ms ish at 4K. This means the cards that would 'benefit' (I'm against all forms of TAA cancer, so i don't care if it works or not) from this upsmear technology the most (like 6600xt) would also be severely capped by max fps they can provide with FSR4 enabled (multiply my numbers by 2-3x). This would naturally be meticulously 'researched' by nV foundry and its likes, so that would be a bad press for AMD (and there were enough of it because of the MSRP price tricks and launch rebates for the vendors)



1 / (3.8ms) = 263 fps, which is actually below what my monitor can do (if i was an avid fan of what TAA does with image and movement clarity, i'd be terribly disappointed)

 

[Win2012R2]

This level of planned obsolescence from AMD, considering the fact that the hardware is capable, is unprecedented.

We live in unprecedented times of AI bubble: any "AI" GPU engineer will be working on far higher margin stuff now than supporting products that got EOLed and never meant to run this stuff in the first place.

Once things settle with (hopefully!) RDNA5 in regards to fast execution of "AI" stuff then from that point on there is a decent hope that this hardware will be supported better.

FSR4 takes a bit of time even at 1080p with full blown N21, more than 1ms and that rises to 4ms ish at 4K.

Yeah, if AMD releases this then Nvidia shills will say it's slower crap, and recommend to buy used Ampere instead, there is no winning here for AMD.

If we are going that way, Nvidia also dropped support for older generations by not providing FG.

FG requires (for good perf) dedicated hardware that was first introduced in Ada.

 

[coercitiv]

Yeah, if AMD releases this then Nvidia shills will say it's slower crap, and recommend to buy used Ampere instead, there is no winning here for AMD.

This argument is the weakest of them all: Nvidia shills grilling AMD for NOT releasing this will have an even greater effect, for an even longer period of time.

For example look, at this negative coverage of FS4 INT8 on RX 6700 from Digital Foundry:

 

[Joe NYC]

RDNA 5 (AT2 — 10070xt) & Xbox Magnus ready to go by Q3 2027 — an MLID source who has never been wrong

I wonder what it means for timing of RDNA5 APUs (Medusa Premium, Medusa Halo) and dGPUs.

 

[Heartbreaker]

For example look, at this negative coverage of FS4 INT8 on RX 6700 from Digital Foundry:

Is that supposed to be sarcasm? Because I'm not seeing the "negative coverage". They are just showing the limits of going back too far with FSR 4 on older HW.

They did the same thing show DLSS 4.5 on RTX 3070, basically saying it just didn't cut on older HW.

 

[coercitiv]

Is that supposed to be sarcasm?

Yes.

 

[marees]

I wonder what it means for timing of RDNA5 APUs (Medusa Premium, Medusa Halo) and dGPUs.

medusa premium needs a new socket FP10 shared with medusa point 1 (128bit lpddr5x). probably CES 2028 reveal. GPU 10050xt somewhere around june/july 2028

the fly in the ointment for medusa premium is the unified memory. it can easily take 64gb or more. attractive for super budget AI use cases. expect prices to remain high until late 2028 or 2029

 

[marees]

I wonder what it means for timing of RDNA5 APUs (Medusa Premium, Medusa Halo) and dGPUs.

by precedence of 9070xt, AMD will wait 3 months after nvidia to release the magnus AT2 as a stand alone GPU

 

[marees]

I wonder what it means for timing of RDNA5 APUs (Medusa Premium, Medusa Halo) and dGPUs.

forget about medusa halo / AT3. It is lpddr6. prices will remain high until 2030

 

[marees]

I wonder what it means for timing of RDNA5 APUs (Medusa Premium, Medusa Halo) and dGPUs.

magnus itself is semi-custom. by that definition Microsoft is the sole entity that can buy this chip from AMD

Medusa premium is a bog standard laptop apu on a bog standard socket using a bog standard memory. it is the z3 extreme for PC handhelds. expect a slight premium like the ASUS / Lenovo handhelds.

also it can be used as a steam machine like NUC (basically the defacto series S2 — but officially there is no series S2). will sell around the price of the steam machine

 

[Joe NYC]

medusa premium needs a new socket FP10 shared with medusa point 1 (128bit lpddr5x). probably CES 2028 reveal. GPU 10050xt somewhere around june/july 2028

Medusa Premium goes into the old socket FP10 using 128 bit LPDDR5. Only Medusa Halo gets a new socket (384 bit LPDDR6).

At some point, AMD will introduce 192 bit socket for LPDDR6, but that's not on any roadmaps so far.

the fly in the ointment for medusa premium is the unified memory. it can easily take 64gb or more. attractive for super budget AI use cases. expect prices to remain high until late 2028 or 2029

Medusa Premium would likely be the low end of that use case (local / client AI). With Strix Halo still possibly outperforming it, and having 2x memory.

And then, if Medusa Halo is imminent (right behind Medusa Premium), it will get more attention.

 

[Joe NYC]

magnus itself is semi-custom. by that definition Microsoft is the sole entity that can buy this chip from AMD

I wonder if there is some deal for AMD to be able to resell it, and in turn AMD offering cheaper price to Microsoft.

Since Microsoft wants more XBox like devices out there, it could be a win-win.

 

[adroc_thurston]

it is the z3 extreme for PC handhelds

Nope.
Handheld line is dead.