Discussion RDNA 5 / UDNA (CDNA Next) speculation

[adroc_thurston]

If AMD isn't serious about dGPU then why did they poach all this talent from Intel, NVIDIA, Imagination etc...? All this just for the PS5 and bare minimum dGPU effort.
Might be naive but this massive R&D effort looks like a serious attempt at a Zen-like reset for AMD GPU IP.

They want class-leading GPU IP for their core markets.
They do not want to compete with NV in discrete graphics, it's expensive and pointless.

 

[MrMPFR]

They want class-leading GPU IP for their core markets.
They do not want to compete with NV in discrete graphics, it's expensive and pointless.

So all this for consoles, mobile (including handhelds), and Xbox cloud?
dGPU is just an afterthought that gets scraps of AT0-AT4 and keeps status quo?

 

[adroc_thurston]

So all this for consoles, mobile (including handhelds), and Xbox cloud?

Yeah, markets which they already either own or are favourable in.

dGPU is just an afterthought that gets scraps of AT0-AT4 and keeps status quo?

Yea.
The entire dGFX lineup is scraps from elsewhere.

 

[Thunder 57]

Yeah, markets which they already either own or are favourable in.

Yea.
The entire dGFX lineup is scraps from elsewhere.

OMG enough.

 

[Claudiovict]

If client AT0 doesn't do this, it'll mean the whole RDNA5 architecture is a flop.

fr, 4090 is 40% stronger than 9070 XT, if a new gen with more than double count of CUs doesnt beat it... i even hope it gets well stronger than a 5090

 

[MrMPFR]

fr, 4090 is 40% stronger than 9070 XT, if a new gen with more than double count of CUs doesnt beat it... i even hope it gets well stronger than a 5090

Some napkin math for AT2 and AT0, don't take this too seriously:
- Better scaling efficiency 9060XT -> 9070XT ~91% clockspeed adjusted, modular scheduling can push that much higher, let's raise to a conservative 95%, 2% overall, 1.02
- 64 -> 70 CUs = 9% = 1.09
- 10% higher IPC = 1.1
- 10% higher clocks = 1.1

1.02 x 1.1 x 1.1 = 1.35 = +35%, push clocks a bit more and should land just shy of a 4090. No one else has provided die size, so let's use MLID's 264mm^2 for GMD + add 40-50mm^2 MID. 18GB GDDR7 via 6 x 24Gb 36gbps. If things return to normal AMD shouldn't have that much trouble selling this for $649-699.

As for AT0 even if cut down leaked 154CU MLID config then many ways but ideally 8 SEs, but 7 might happen.
70 -> 154 = 2.2X CUs, but SE frontends might only scale 1.75-2X.
Lets even that out to 2X. Subtract 5% from scaling loss (remember new scheduling is modular and incredibly scalable) and 10% from lower clocks.
- +70% 4090, 5090 is +35%, +26% faster than RTX 5090.
384bit cut down bus should be plenty for gaming.
36GB GDDR7 via 12 x 24Gb 36Gbps GDDR7
- $1399 doable, but only because it's garbage yields

Crackpot AT0 could be 180CUs,
70 -> 180 = 2.57X CUs, frontends 2X
Let's even that out to 2.4X
Scaling loss still 5% and AMD pushes it hard so only 5% lower clocks.
- +116% 4090, 5090 +35%, +60% faster than RTX 5090.
448bit cut down bus should be plenty to feed the beast.
42GB GDDR7 via 14 x 24Gb 36Gbps GDDR7
- Don't see AMD going below $1999.

So yeah it's safe to say the big die will prob be significantly stronger than 5090.

 

[Win2012R2]

They want class-leading GPU IP for their core markets.

Wait, what happened to ""No one at AMD is interested in gaming."?

What are those "core markets", they sound important.

So all this for consoles, mobile (including handhelds), and Xbox cloud?

Yeah, them consoles AKA dedicated gaming machines - which obviously don't count

 

[itsmydamnation]

Wait, what happened to ""No one at AMD is interested in gaming."?

What are those "core markets", they sound important.

Yeah, them consoles AKA dedicated gaming machines - which obviously don't count

to be fair , if you go back a few days , before you started trying to make your point , he said the exact same thing previously.

 

[marees]

maybe there is some hope for Radeon after all 🤔

AMD aims to capture 25% of the GPU market in China - VideoCardz.com​

AMD reportedly reshuffled its China GPU sales team and set a 25% share target for 2026​

Reports from China say AMD has completed a sales structure change for its consumer graphics business in the country. The reports suggest GPU channel oversight is now handled under the same leadership that manages AMD’s CPU sales locally.

According to the reports, AMD’s China graphics card sales director has been reassigned to other responsibilities. The GPU business is said to be managed by the CPU sales director going forward. The same reporting claims AMD plans to run consumer GPU and processor channel work as one combined effort, including coordinated engagement with add-in-board partners.

The reports from Chinese media add that AMD has set an internal goal of reaching 25% discrete GPU market share in China during 2026. Board partners are said to have received higher sales targets as part of that plan.

https://videocardz.com/newz/amd-aims-to-capture-25-of-the-gpu-market-in-china

 

[maddie]

They want class-leading GPU IP for their core markets.
They do not want to compete with NV in discrete graphics, it's expensive and pointless.

That is definitely NOT the same as "not interested in gaming".

 

[adroc_thurston]

That is definitely NOT the same as "not interested in gaming".

Gaming means dGFX gaming. That's where the money's at.

 

[ToTTenTranz]

That is definitely NOT the same as "not interested in gaming".

It's also just not a factual statement.

Both AMD and Nvidia are interested in gaming. Being less focused in gaming doesn't mean they're not interested.

Gaming saved AMD from bankruptcy during Bulldozer/Piledriver/Excavator and brought Nvidia most of their revenue up until 3 years ago. Both AMD and Nvidia know that an AI bubble burst is possible, regardless of what they say in public. Intel is probably going to release a new graphics card today.

They never stopped being interested in gaming.

 

[inquiss]

It's also just not a factual statement.

Both AMD and Nvidia are interested in gaming. Being less focused in gaming doesn't mean they're not interested.

Gaming saved AMD from bankruptcy during Bulldozer/Piledriver/Excavator and brought Nvidia most of their revenue up until 3 years ago. Both AMD and Nvidia know that an AI bubble burst is possible, regardless of what they say in public. Intel is probably going to release a new graphics card today.

They never stopped being interested in gaming.

AMD is not interested in trying to win the dGPU market. They haven't designed a product to compete with Nvidias high end in a way that shows they want to dominate there. They continue to play in the laptop market with APU but they have dropped out of the laptop dGPU market and the high end dGPU. They need to continue to integrate IP to continue to win in console and to grow APUs.

So, they are not interested in devoting resources to usurping Nvidia in either flavour of dGPU (laptop or desktop). You can see that in the choices they have made on what SKUs to move forward with and the cancelling of big Navi.

Could change if they get more mind share through their mid-range options but isn't where the resources are going now. That's it.

 

[adroc_thurston]

Gaming saved AMD from bankruptcy during Bulldozer/Piledriver/Excavator

This is a load of bullshit.
Radeon was opmargin-negative until RDNA1.

 

[marees]

This is a load of bullshit.
Radeon was opmargin-negative until RDNA1.

he means the gaming division (which included a 16 member PS4 team working on jaguar/bobcat apu)

 

[adroc_thurston]

he means the gaming division (which included a 16 member PS4 team working on jaguar/bobcat apu)

Consoles kept the gfx R&D afloat but they didn't give AMD any real money.

 

[ToTTenTranz]

AMD is not interested in trying to win the dGPU market. They haven't designed a product to compete with Nvidias high end in a way that shows they want to dominate there. They continue to play in the laptop market with APU but they have dropped out of the laptop dGPU market and the high end dGPU. They need to continue to integrate IP to continue to win in console and to grow APUs.

So, they are not interested in devoting resources to usurping Nvidia in either flavour of dGPU (laptop or desktop). You can see that in the choices they have made on what SKUs to move forward with and the cancelling of big Navi.

Could change if they get more mind share through their mid-range options but isn't where the resources are going now. That's it.

This is a massive convolution of two separate facts.

Not trying to "win the dGPU market" doesn't mean they stopped caring about gaming. Had they stopped caring about gaming they'd be announcing they'll leave the client dGPU market altogether and gradually tone down all driver development and gamedev relations, which they're not. Consoles are gaming as well. AMD didn't drop out of laptop gaming either, they're trying to transition laptop gaming into APUs and you can see that very clearly with the Medusa range.
They're also coming up with an AT0 client dGPU and that definitely means they're interested in gaming as well.


This "AMD isn't interested in gaming" theme is a hoax. It's pure conjecture / wet dreams / made-up nonsense from a couple of people who have consistently been proven wrong time and again, especially in recent months/years. Yes, there's a focus on AI and that's where most of their resources and top minds are going towards at the moment, but there's no plan to leave gaming because they're not interested in it. Everything else is noise.

This is a load of bullshit.
Radeon was opmargin-negative until RDNA1.

Nonsense.

https://www.sec.gov/Archives/edgar/data/2488/000119312514057240/d674550d10k.htm

In 2012 and 2013, Radeon (Graphics and Visual Solutions) was the only division making a profit.

 

[adroc_thurston]

In 2012 and 2013, Radeon (Graphics and Visual Solutions) was the only division making a profit.

Radeon != graphics.
The actual gaming (what NV labes as such, dGFX) sales were opmargin-negative until RDNA1.
Mind you, that was opinc with no roadmap, just infinite rebrands. Kind of funny.

AMD didn't drop out of laptop gaming either, they're trying to transition laptop gaming into APUs and you can see that very clearly with the Medusa range.

"they didn't drop outta gaming laptops, they're just recreating them from the first principles" sure buddy, sure.

They're also coming up with an AT0 client dGPU and that definitely means they're interested in gaming as well.

You gotta toss die candidate rejects somewhere.

This "AMD isn't interested in gaming" theme is a hoax. It's pure conjecture / wet dreams / made-up nonsense from a couple of people who have consistently been proven wrong time and again, especially in recent months/years. Yes, there's a focus on AI and that's where most of their resources and top minds are going towards at the moment, but there's no plan to leave gaming because they're not interested in it. Everything else is noise.

whoa man that's a lot of hopes and dreams.
you ain't getting anything though. just die rejects from other markets.

 

[ToTTenTranz]

Radeon != graphics.

During the 15h CPU era (2011-2016), AMD didn't launch a single client or server product out of Graphics and Visual Solutions that wasn't made up of Terascale and GCN chips, developed by the Radeon team.
There was not a single AMD-powered console that didn't use Terascale or GCN architecture, developed by the Radeon team.

During AMD's dark days of Bulldozer, Radeon == graphics.

The actual gaming (what NV labes as such, dGFX) sales were opmargin-negative until RDNA1.

Made-up nonsense quickly disproven just by looking at quarterly reports.

You gotta toss die candidate rejects somewhere.

They could toss candidate rejects into lower-end SKUs for the server market that would still have >2x higher markup than any gaming SKU they're putting out there.
Yet, they're still choosing to make a fun gaming dGPU. Because they're interested in gaming.

you ain't getting anything though. just die rejects from other markets.

You mean those other markets that are still having to pay for die area dedicated to ROPs, raytracing units and a bunch of other stuff that's useless for anything that isn't gaming?

whoa man that's a lot of hopes and dreams.

It's the other way around. Someone needs to cope a bit harder with the fact that AMD is interested in gaming.

 

[adroc_thurston]

During the 15h CPU era (2011-2016), AMD didn't launch a single client or server product out of Graphics and Visual Solutions that wasn't made up of Terascale and GCN chips, developed by the Radeon team.

They didn't launch any GCN updates either.
The roadmap was fed into the woodchipper.

Made-up nonsense quickly disproven just by looking at quarterly reports.

Show me detailed opex breakdowns between client and gfx.

They could toss candidate rejects into lower-end SKUs for the server market that would still have >2x higher markup than any gaming SKU they're putting out there.
Yet, they're still choosing to make a fun gaming dGPU. Because they're interested in gaming.

No you can't, server can take a very limited SKU volume. Especially cloud gaming where you have strict instance ratios.

You mean those other markets that are still having to pay for die area dedicated to ROPs, raytracing units and a bunch of other stuff that's useless for anything that isn't gaming?

Those are cheap.

Someone needs to cope a bit harder with the fact that AMD is interested in gaming.

if they're interested in gaming, show me a dedicated halo part aimed at winning dGFX race.
Just to lay a claim to the actual gfx TAM, not the zeromargin ghetto AMD is stuck in.

Oh, but you can't. Because no one at AMD is allowed to build discrete graphics aimed at competing with Nvidia.
They're not even allowed to have their own roadmaps anymore; folded wholly under client.

 

[ToTTenTranz]

They didn't launch any GCN updates either.
The roadmap was fed into the woodchipper.

GCN had 4 major revisions, dude. It goes up all the way to GCN 5 on Vega, and I'm not even counting on the minor GCN 5.1 revision for Vega 7nm that later evolved into CDNA.

Show me detailed opex breakdowns between client and gfx.

Doesn't matter when they're all using the exact same chips and architecture developed by the Radeon team and they're the only division posting profits.

No you can't, server can take a very limited SKU volume.

I guess this picture is a lie, then.

Especially cloud gaming where you have strict instance ratios.

You mean AMD is interested in gaming after all?

Those are cheap.

9% of die area on raster frontends and display PHYs alone isn't cheap.

if they're interested in gaming, show me a dedicated halo part aimed at winning dGFX race.

The overwhelming majority of gaming doesn't happen in halo parts. This is a flawed argument.
AMD can be interested in gaming (which they are) without launching a low volume halo part.

Because no one at AMD is allowed to build discrete graphics aimed at competing with Nvidia.

What are the RX 9070 and 9060 competing with, then? High-end coffee beans and t-shirts?

 

[soresu]

What are the RX 9070 and 9060 competing with, then? High-end coffee beans and t-shirts?

I think he meant hi end SKUs.

Which is a fair assessment given 7900 flopped as a competitor to nVidia's contemporary hi end, and they didn't even bother with 8x or 9x SKUs for the 9000 series despite clearly having a competitive µArch this time (probably their biggest face palm moment since Vega 10 tbh).

 

[adroc_thurston]

GCN had 4 major revisions, dude

Calling any of that a major revision is lol.
NV had 3 major SM overhauls across 2012-2017 relative.

Doesn't matter when they're all using the exact same chips and architecture developed by the Radeon team and they're the only division posting profits.

It's very easy to post opinc when you demolish your opex into nothigness.
Can you post 2013 gfx opex?

I guess this picture is a lie, then.

??
NV DC income is all from bespoke DC parts. Same as AMD.

You mean AMD is interested in gaming after all?

It's VDI; very different die design considerations there.

9% of die area on raster frontends and display PHYs alone isn't cheap.

Display shoreline you're gonna have on most non-HBM parts anyway due to the sheer GDDR edge requirement.

The overwhelming majority of gaming doesn't happen in halo parts. This is a flawed argument.

You can never gain any share or command any price without winning at the top.

AMD can be interested in gaming (which they are) without launching a low volume halo part.

They can be interested in gaming without building any discrete parts at all.
Which RDNA5 has none of!

What are the RX 9070 and 9060 competing with, then?

Themselves, mostly.
They're not gaining AMD any mss.

 

[adroc_thurston]

I think he meant hi end SKUs.

Oh no those are also flops.
Healthy margins but dismal volumes.

Which is a fair assessment given 7900 flopped as a competitor to nVidia's contemporary hi end, and they didn't even bother with 8x or 9x SKUs for the 9000 series despite clearly having a competitive µArch this time (probably their biggest face palm moment since Vega 10 tbh).

Navi31 was horribly undersized relative to ad102.
They're just not allowed to build big.
NV can ship a 750mm^2 client die. AMD guys have to blink and surrender since Lisa won't ever sign off a client program this expensive.

 

[ToTTenTranz]

I think he meant hi end SKUs.

No, he means gaming in general. Which is factually wrong.