Discussion RDNA 5 / UDNA (CDNA Next) speculation

[ToTTenTranz]

I think he meant hi end SKUs.

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

 

[adroc_thurston]

Which is factually wrong.

It's so factually wrong RDNA5 generation has zero purpose-built dGFX parts.
They're at zerodies level of competing with NV.
RDNA2 had 4 dies. RDNA3 3 dies. RDNA4 2 dies.
RDNA5 brings us a novel solution of having literally zero dies built around competing in dGFX. yay.

 

[inquiss]

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.



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



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 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?




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

AT0 is a cloud gaming part. Those rejects have all those gfx bits because they'll be used for cloud gaming.

Doesn't mean they designed a part to specifically take on Nvidia in dGPU

 

[adroc_thurston]

Doesn't mean they designed a part to specifically take on Nvidia in dGPU

None of them are.
What if NV 103 rubin is 96SM with the chop being like 88SMs.
How will le epick Navi48 replacement aka AT2 compete?

 

[soresu]

AT0 is a cloud gaming part. Those rejects have all those gfx bits because they'll be used for cloud gaming.

Doesn't mean they designed a part to specifically take on Nvidia in dGPU

As long as it's actually vailable to buy for financially irresponsible shut ins like myself I don't care what the business strategy was behind its design 🤪

Anyways I thought cloud gaming was not so profitable?

Or was that just Stadia falling flat on its face?

 

[adroc_thurston]

Anyways I thought cloud gaming was not so profitable?

It's an alright market.

 

[soresu]

Navi31 was horribly undersized relative to ad102.

That's what I meant about 7900 series.

6900 was barely competitive and the 16 extra CUs + a bit of extra clock wasn't nearly enough to compete against nVidia's next gen.

They played up the limited use dual issue FLOPS of RDNA3 like it was a true wunderkind moment to the extent that they might as well have shot themselves in the foot multiple times.

 

[marees]

As long as it's actually vailable to buy for financially irresponsible shut ins like myself I don't care what the business strategy was behind its design 🤪

Anyways I thought cloud gaming was not so profitable?

Or was that just Stadia falling flat on its face?

.microsoft has big plans for cloud gaming

1. free tier — 5 hours free per month
2. more regions
3. more resolutions
4. console hardware getting costlier & less subsidized
5. bundled cloud gaming with all gamepass plans
6. bring cloud gaming to steam, android & apple devices
7. tie up with LG & Samsung to bring cloud to all smart displays (including TVs, cars etc.)

 

[adroc_thurston]

6900 was barely competitive

Oh no, that one was nightmarishly scary.
All that on a 256b 16/18G GDDR6 was spoopy.

They played up the limited use dual issue FLOPS of RDNA3 like it was a true wunderkind moment to the extent that they might as well have shot themselves in the foot multiple times.

Oh no no one at AMD gafs about math rates.
RDNA3 just clocked 800MHz below target.

 

[soresu]

.microsoft has big plans for cloud gaming

1. free tier — 5 hours free per month
2. more regions
3. more resolutions
4. console hardware getting costlier & less subsidized
5. bundled cloud gaming with all gamepass plans
6. bring cloud gaming to steam, android & apple devices
7. tie up with LG & Samsung to bring cloud to all smart displays (including TVs, cars etc.)

Oh lol I always forget about MS, but then so does everyone else 😂🤣

Oh no no one at AMD gafs about math rates.

I'll assume that one was sarcasm given it was on all the PR and retail box info.

 

[adroc_thurston]

I'll assume that one was sarcasm given it was on all the PR and retail box info.

They didn't quote doubled SPs or anything.

 

[Win2012R2]

he said the exact same thing previously.

And I did not call him on that one.

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.

Prediction bookmarked.

They're not gaining AMD any mss.

I am nVidia from Riva128, but 9070XT (or 9060 XT) is the part I recommend to people I know now, RDNA4 is excellent and it's clearly selling well.

It's so factually wrong RDNA5 generation has zero purpose-built dGFX parts.

Oh wow, what a shocker, and one can say Zen generation has "zero purpose-built parts" for client too, did not stop AMD from succeeding there.

Soon there will be big high end refresh cycle from millions of those who own 4090 (like myself), and I'll gladly look at RDNA5/6 if it can hack it, with PS6 win and general slow down in scaling down chips AMD to me looks like will be far better long term choice soon.

 

[adroc_thurston]

Prediction bookmarked.

Navi36 and Navi4c both died, big dawg. lmao.

is the part I recommend to people I know now, RDNA4 is excellent and it's clearly selling well.

It's aight but AMD is still at ~7% mss despite both RDNA4 parts launching.

Oh wow, what a shocker, and one can say Zen generation has "zero purpose-built parts" for client too, did not stop AMD from succeeding there.

See that's the thing.
Nvidia isn't inept and you can't coast in client gfx by shipping rando server scraps.
You need wholly bespoke solutions.

Plus, generally speaking, DIY DT AMD shipped more stuff (more cores, more freq, more cache, more anything) and they just can't do that in client dGFX.

 

[Tangopiper]

AT0
1. XCloud
2. dGPU

AT2
1. XBOX Magnus
2. dGPU

AT3
1. MDSH
2. dGPU

AT4
1. MDSP
2. dGPU

Pretty clear all of these parts are made for other products first, Consumer dGPU second.

The indicator of whether AMD cares about Consumer dGPU, dGFX, gaming etc. (whatever you all want to call it) is if AT1 actually materialises. That's the only part without a clear primary market outside of consumer dGPU. Take a guess why it's not part of any leaks so far.

 

[helloworld321]

Wouldn't AT1 just be a binned AT0 (36GB / 384 bit bus), but with a 320bit (30GB) or 256bit (24GB) bus and fewer CUs?

That's pretty much what they did with 7900 XTX and 7900 XT.

 

[adroc_thurston]

Wouldn't AT1 just be a binned AT0, but with a 320bit (30GB) or 256bit (24GB) bus (instead of a 36GB / 384 bit bus) and fewer CUs?

No, those are separate design candidates.
AMD denominates chops with traditional ATi scheme of XTX/XT/XL/yaddayadda.

That's pretty much what they did with 7900 XTX and 7900 XT.

N31XTX and XT.

 

[Tangopiper]

Wouldn't AT1 just be a binned AT0 (36GB / 384 bit bus), but with a 320bit (30GB) or 256bit (24GB) bus and fewer CUs?

That's pretty much what they did with 7900 XTX and 7900 XT.

Cutting AT0 down that much isn't likely to be needed. It's 512b full config, so 384 is already a heavily salvaged die.

But as adroc has said AT1 would be a different chip entirely, and I would assume aimed at GR203 (RTX 6080).

I speculate we will see a 6070 Ti competitor using AT2, nothing vs. 6080, and something between the 6080 and 6090 using AT0. If it's actually a success, NVIDIA might grace us with a 6080 Ti using salvaged GR202.

A boy can dream, eh?

(GR codename for Rubin client isn't confirmed AFAIK, just assuming.)

 

[adroc_thurston]

I speculate we will see a 6070 Ti competitor using AT2

Depends on how many SMs NV loads into GR203.

If it's actually a success, NVIDIA might grace us with a 6080 Ti using salvaged GR202.

if heavily chopped AT0 is in the ballpark then they're gonna ship a full die and go hunting.
They're not gonna win but it might just be an honorable samurai duel.

 

[dangerman1337]

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.

Isn't AT0 going to be a huge die like 700mm2 or so? if 264mm2 (basing on the MLID Xbox Magnus leak) for AT2 is with a 192-bit bus and assuming 72SMs then scaling the multiplier on a rough assumption of 512 / 192 = 2.666666666666667 and then 2.666666666666667 x 264 gets you to 704mm2 or so. Especially it's N3P. And they can switch apparently I/O dies so probably a cheap Pci-E Gen 5 N6 die for client and then Gen 6 for fancier stuff according to leaked RDNA 5 documents.

Now whether AMD will launch a AT0 SKU on desktop that isn't clam-shell that's the (near)-full die to be seen.

AT0
1. XCloud
2. dGPU

AT2
1. XBOX Magnus
2. dGPU

AT3
1. MDSH
2. dGPU

AT4
1. MDSP
2. dGPU

Pretty clear all of these parts are made for other products first, Consumer dGPU second.

The indicator of whether AMD cares about Consumer dGPU, dGFX, gaming etc. (whatever you all want to call it) is if AT1 actually materialises. That's the only part without a clear primary market outside of consumer dGPU. Take a guess why it's not part of any leaks so far.

Pretty muich (though I think AT4 won't appear in dGPU, won't be surprised if AMD like just tries to push Medusa Halo kits and maybe down the line Grimlock Halo's successor is on AM6 w/ 2100 pins as the future of entry level stuff for Desktop in a mini ITX format) I think AT1 is financially hard to justify because if AMD needs say a 24GB that's faster than the current AT2... they'll probably do potentially an N3X (Since current RDNA 5 GMDs are N3P) + 4GB GDDR7 Module RDNA 5 refresh in 2028 that launches along the same time frame as Zen 7 since IIRC Lisa Su noted that Zen sales are boosted when new graphics come out? Can see Zen 7 + N3X RDNA5 refresh synergy in 2028.

 

[Tangopiper]

Depends on how many SMs NV loads into GR203.

Let's be real, bare minimum. It'll be 96 absolute maximum (my crystal ball tells me so) and only the 6080 and 6090 laptop is seeing close to that SM count. 6070 Ti will be the dregs.

if heavily chopped AT0 is in the ballpark then they're gonna ship a full die and go hunting.
They're not gonna win but it might just be an honorable samurai duel.

I can see them being opportunistic if they get close, yeah. It's the only scenario we see anything other than heavily chopped AT0 in the client space.

Makes the cancellation of Navi 4C even more laughable given how easily that would have slaughtered the 5090. Blackwell was 💩

 

[Tangopiper]

Pretty muich (though I think AT4 won't appear in dGPU

It probably will, but purely to get rid of the dies with poor V/f curves that can't be shipped in any MDSP SKU.

GR207/RTX 6050 competitor.

 

[adroc_thurston]

Isn't AT0 going to be a huge die like 700mm2 or so?

No lol.
You're mostly scaling compute and microscheduler rings on gfx13.

Let's be real, bare minimum. It'll be 96 absolute maximum (my crystal ball tells me so) and only the 6080 and 6090 laptop is seeing close to that SM count. 6070 Ti will be the dregs.

yea gonna bet on 96SM config too.

Makes the cancellation of Navi 4C even more laughable given how easily that would have slaughtered the 5090. Blackwell was 💩

Cancelling is one thing; having no replacement at all is outright boorish.
They have amazing (and cheap) IP and they have built zero notable configs with it.

 

[Tangopiper]

They have amazing (and cheap) IP and they have built zero notable configs with it.

It's very frustrating. And RDNA 5 seems to only amplify this.

RDNA 2 was a brief moment of hope that AMD had the IP and momentum to actually compete for the halo. Then RDNA 3 happened lol.

@Kepler_L2
Any info on RDNA 5 Render Backend? Still RB+ or something more chonk?

 

[adroc_thurston]

And RDNA 5 seems to only amplify this.

It's ok.

Then RDNA 3 happened lol.

Well RDNA3 would've been kind of underwhelming relative to full die ad102 in any case.
Can't have stuff when the comp ships 50% more shaders.

Any info on RDNA 5 Render Backend? Still RB+ or something more chonk?

nanitemaxxed computecels care not about grungy old RBEs.

 

[Tangopiper]

It's ok.

Sure, the IP is there to make up for the lack of fun stuff. But I'd prefer the fun stuff too

Well RDNA3 would've been kind of underwhelming relative to full die ad102 in any case.
Can't have stuff when the comp ships 50% more shaders.

IDK, +30% clock would have put it right in 4090 space. Nb for a substantially smaller die and a hell of a lot less N5 than AD102.

nanitemaxxed computecels care not about grungy old RBEs.

Nanite. SHUDDER