Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

Man I have been dying to make this one for a while now.

First rumours for RDNA3 are here so new thread time!

Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[igor_kavinski]

Using that reasoning, I guess Intel should pay you for the privileged of using their cards?

Maybe they are ALREADY paying him

 

[adroc_thurston]

Using that reasoning, I guess Intel should pay you for the privileged of using their cards?

Well they've literally paid Argonne to do just that, so...
...why not me?

You seem to like anime and manga a lot.

I also like Farscape, Stargate SG-1 and other bits of good TV sci-fi.

 

[igor_kavinski]

Whenever you see N32 crawl out of the dungeon, which I don't really know when anymore since they've missed all the launch windows for it, both DT and mobile.

So the reason for the delay is that it's THE RDNA3 silicon done right?

 

[adroc_thurston]

So the reason for the delay is that it's THE RDNA3 silicon done right?

They've been keeping it downstairs to figure out what exactly went wrong, yea.
Usually wouldn't be that important but there's a whole batch of very very important RDNA3 parts coming.

 

[TESKATLIPOKA]

So the reason for the delay is that it's THE RDNA3 silicon done right?

Performance of N32 is supposedly Rx 6800xt level, and that is too low for a >3GHz 60CU GPU. So either the frequency will stay comparable to N31 or performance will be a lot higher.

@adroc_thurston That would be comparable to RTX 4070 both in efficiency and performance excluding RT.

 

[adroc_thurston]

Performance of N32 is supposedly Rx 6800xt level

Yea at 190W or so.
Still a wee bit lower than it should be iirc.

 

[Ajay]

They've been keeping it downstairs to figure out what exactly went wrong, yea.
Usually wouldn't be that important but there's a whole batch of very very important RDNA3 parts coming.

So, this has been an ongoing debate. Has AMD done a major respin on RDNA3, even though it has RDNA4 coming out next year? Or, are your talking about something else.

 

[MrTeal]

Meh it has issues while NV sells a working product.

Could you expand more on that? What kind of issues does it have that haven't been revealed in reviews like the performance and higher power draw? Complaints seem quite low with this generation, I haven't even really seen the usual gnashing of teeth about drivers we see with every AMD launch.

 

[adroc_thurston]

Has AMD done a major respin on RDNA3

They have no choice but to figure it all out, since whole 4 extremely important Strix parts use it.

even though it has RDNA4 coming out next year?

Different team, different gfx IP.

 

[Mopetar]

I'm assuming that RDNA4 shares some common design elements with RDNA3 and that it wouldn't make much sense for them to redesign everything from scratch.

 

[adroc_thurston]

I'm assuming that RDNA4 shares some common design elements with RDNA3 and that it wouldn't make much sense for them to redesign everything from scratch.

Of course, but every big gfx IP number change tends to be a rather big break those days.

 

[Joe NYC]

Yea at 190W or so.
Still a wee bit lower than it should be iirc.

If that's the case, then it would be possible to release Navi 30, that is 2x Navi 32 die size and be within 350-400 Watts.

 

[adroc_thurston]

If that's the case, then it would be possible to release Navi 30, that is 2x Navi 32 die size and be within 350-400 Watts.

Eh scaling doesn't quite work like that still, and they've canned the big boy that was to be on top of N31 anyway.
Better luck next time, fellas!

 

[maddie]

They have no choice but to figure it all out, since whole 4 extremely important Strix parts use it.

Different team, different gfx IP.

AMD fixing/optimizing RDNA for strix doesn't automatically mean improved discrete parts to the ones we have now, which was the question.

After the RDNA3 team is finished, I thought they go to RDNA5. Saying RDNA5 will be late?

 

[Ajay]

They have no choice but to figure it all out, since whole 4 extremely important Strix parts use it.

I was too vague - meant for AIB GPU. If AMD doesn't have silicon back from TSMC, then that's clearly a no - since that puts the marketing window into RDNA4 timeframe. If we are talking APU, etc. - different story.

Different team, different gfx IP.

I understand. Hope they don't blow it. Do you know if it's going to N4 or N3 - again, topic of debate, unless I've missed something. Seems like most are leaning N4.

 

[adroc_thurston]

AMD fixing/optimizing RDNA for strix doesn't automatically mean improved discrete parts to the ones we have now, which was the question.

It really does, since N32 is the text subject.
Like, they're not doing it in vacuum or on Strix parts actual (which should TO rather soon anyway).

Saying RDNA5 will be late?

People that triage parts on bringup aren't the same people calling major uArch shots and you know that.

I was too vague - meant for AIB GPU.

The answer is probably, depending on how soon(tm) they work this all out.

Do you know if it's going to N4 or N3 - again, topic of debate, unless I've missed something. Seems like most are leaning N4.

RDNA4 seen some shuffles so honest to god no idea what it'll be.

Hope they don't blow it

RDOA3 is a rather uncharacteristic pre-Si miss.
All parts in the world have funny errata (parts like Navi10 were rather buggy but had no PPA misses of note) but this is the first time in years they've majorly missed on perf/power targets in CPUs or GPUs both.

 

[igor_kavinski]

but this is the first time in years they've majorly missed on perf/power targets in CPUs or GPUs both.

Could nGreedia have bribed TSMC to create process hurdles etc. for AMD? I'm sure they don't want AMD enjoying the same silicon quality as they do for their GPU dies. Or more likely, AMD didn't pay as much as nGreedia so they got lesser quality resources from TSMC?

 

[adroc_thurston]

Could nGreedia have bribed TSMC to create process hurdles etc. for AMD?

no lol.

I'm sure they don't want AMD enjoying the same silicon quality as they do for their GPU dies

They don't really care.
NV does its own thing while being super-wary of the competition, that's it.

Or more likely, AMD didn't pay as much as nGreedia so they got lesser quality resources from TSMC?

no lol, AMD is a major growth driver for $TSM given the smartphone slump.

 

[Joe NYC]

Could nGreedia have bribed TSMC to create process hurdles etc. for AMD? I'm sure they don't want AMD enjoying the same silicon quality as they do for their GPU dies. Or more likely, AMD didn't pay as much as nGreedia so they got lesser quality resources from TSMC?

AMD was a bigger and more important customer for TSMC than NVidia in the recent past. That may be gradually changing in the future, but not in the time frame you are discussing.

 

[GodisanAtheist]

OMFG guys it's a joke based on the AyyyeMD blocking DLSS thread

 

[Ajay]

OMFG guys it's a joke based on the AyyyeMD blocking DLSS thread

Sorry, what?

 

[RnR_au]

AMD fixing/optimizing RDNA for strix doesn't automatically mean improved discrete parts to the ones we have now, which was the question.

It would be an easy 7950XT and 7950XTX for AIB's.

 

[maddie]

It would be an easy 7950XT and 7950XTX for AIB's.

What I meant was, that if the problem was deep in the details, it would mean almost the equivalent of a new design. The xx50 parts traditionally were slight improvements based on process/binning, not a big change. Seeing as we don't know the failure points in the present parts, and haven't been told, it's questionable to trust its a simple fix and that it will happen. Maybe, or not. One insistent poster saying so means naught without more details.

 

[GodisanAtheist]

Sorry, what?

People seem to be taking Igor's comment about Nvidia sabotaging AMD through TSMC seriously, when it's obviously stated with heroic levels of sarcasm.

 

[igor_kavinski]

People seem to be taking Igor's comment about Nvidia sabotaging AMD through TSMC seriously, when it's obviously stated with heroic levels of sarcasm.

Shhhhhhhh....