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

 

[Panino Manino]

What do you expect, precisely?

Before RDNA2 I believed that people here would be greatly disappointed (AMD would only add bad RT, performance and efficiency would not improve), now I believe that they be satisfied with whatever AMD delivers this time.
Vega Era is dead.

 

[Glo.]

Before RDNA2 I believed that people here would be greatly disappointed (AMD would only add bad RT, performance and efficiency would not improve), now I believe that they be satisfied with whatever AMD delivers this time.
Vega Era is dead.

What you are saying is that we have become AMD fanboys who will be satisfied with anything that god almighty Moma Su will give us.

.

 

[GodisanAtheist]

What you are saying is that we have become AMD fanboys who will be satisfied with anything that god almighty Moma Su will give us.

.

- Only the Wool Pant-suite Woman can counter the Leather Jacket Man...

 

[Timorous]

What do you expect, precisely?

To answer this more directly here is where I think a 7900XT will land if the 50% perf/watt target is adhered to.

350W TDP and ~1.8x rasterisation perf at 4k vs the 6900XT. Maybe a 2x+ uplift in RT vs RT.

When AMD start talking about their perf/watt target for RDNA3 I will revisit this. Since Vega they have met or exceeded their public perf/watt targets so I see no reason to doubt their claims when they do make them.

 

[Ajay]

To answer this more directly here is where I think a 7900XT will land if the 50% perf/watt target is adhered to.

I know that the RX 6000 series made a huge jump in perf/watt, but that was with a large node change. Having trouble seeing how AMD hits this target on 6nm. I hope they do, but it's pretty daunting.

 

[maddie]

I know that the RX 6000 series made a huge jump in perf/watt, but that was with a large node change. Having trouble seeing how AMD hits this target on 6nm. I hope they do, but it's pretty daunting.

From RX 5000 series? What node change?

 

[Mopetar]

Undisclosed, proprietary data is not 'legal information'. Leaking that data sets you up for criminal charges.

That really depends on your relationship to the owner of the data. If it's an employ who has a contract with the company that forbids them from sharing the data then yes, there's certainly legal liability on the part of that individual. Someone unaffiliated with the company simply sharing what they've heard isn't illegal and most countries have fairly broad protections for journalists that strictly prevent them from facing criminal charges for reporting on something.

 

[Panino Manino]

What you are saying is that we have become AMD fanboys who will be satisfied with anything that god almighty Moma Su will give us.

.

Not fanboys, just desperate to not see RTG die.

 

[Stuka87]

That really depends on your relationship to the owner of the data. If it's an employ who has a contract with the company that forbids them from sharing the data then yes, there's certainly legal liability on the part of that individual. Someone unaffiliated with the company simply sharing what they've heard isn't illegal and most countries have fairly broad protections for journalists that strictly prevent them from facing criminal charges for reporting on something.

Yes, but typically unreleased, unannounced products and/or technology can fall under 'trade secrets'. But I agree, its a very fuzzy line. But my point to the person I quoted was that not everything leaked is done so "legally".

 

[Ajay]

From RX 5000 series? What node change?

Dammit, completely forgot. I will wear the egg on my face with flair

 

[Mopetar]

Yes, but typically unreleased, unannounced products and/or technology can fall under 'trade secrets'. But I agree, its a very fuzzy line. But my point to the person I quoted was that not everything leaked is done so "legally".

Most people have no obligation to protect a company's trade secrets. You only really see that come up with regards to former employees that have started working for a competitor.

The people who spread leaks/rumors on Twitter aren't going to run afoul of such laws. Even most employees that get caught leaking just get fired and don't face legal action.

Some stuff that's "leaked" is even done so intentionally by companies to build hype for products and get free publicity.

 

[soresu]

What you are saying is that we have become AMD fanboys who will be satisfied with anything that god almighty Moma Su will give us.

I think you meant to say Susano'o.

I'll just show myself out.....

 

[Ajay]

Most people have no obligation to protect a company's trade secrets. You only really see that come up with regards to former employees that have started working for a competitor.

The people who spread leaks/rumors on Twitter aren't going to run afoul of such laws. Even most employees that get caught leaking just get fired and don't face legal action.

Some stuff that's "leaked" is even done so intentionally by companies to build hype for products and get free publicity.

This isn't generally true in my experience. I've worked for high-tech companies and had to sign NDAs, non-compete clauses (which are, legally, weak) plus statements that *any* work product by me is property of the company unless I get approval by the company in advance (aka, consulting work, open source and whatnot).

 

[Stuka87]

Most people have no obligation to protect a company's trade secrets.

Every single tech company I have worked for has required that I sign NDAs. Maybe there are some out there that don't, but I would say they are very much in the minority.

 

[GodisanAtheist]

I know that the RX 6000 series made a huge jump in perf/watt, but that was with a large node change. Having trouble seeing how AMD hits this target on 6nm. I hope they do, but it's pretty daunting.

- I suspect the answer is going to lie in stacked die cache. Instead of 128mb of on die cache, we'll get 2gigs of on-top-of-die cache.

6nm = smaller dies
Stacked Infinity Cache = smaller dies
Assload of IC = less trips to memory/narrow bus/lower power consumption.

AMD is already looking at a 15% IPC improvement in some tasks for their existing Zen 3 cores by pursuing this tech, and is looking at MCM with RDNA3 with this tech.

If AMD can pull this off, better hold on to our butts.

ALL ABOARD, NEXT HYPE TRAIN LEAVING THE STATION AT PRECISELY... NOW!

 

[Glo.]

Not fanboys, just desperate to not see RTG die.

Have you seen that AMD also designs and manufactures CPUs?

What "survival" do you mean?

 

[uzzi38]

- I suspect the answer is going to lie in stacked die cache. Instead of 128mb of on die cache, we'll get 2gigs of on-top-of-die cache.

6nm = smaller dies
Stacked Infinity Cache = smaller dies
Assload of IC = less trips to memory/narrow bus/lower power consumption.

AMD is already looking at a 15% IPC improvement in some tasks for their existing Zen 3 cores by pursuing this tech, and is looking at MCM with RDNA3 with this tech.

If AMD can pull this off, better hold on to our butts.

ALL ABOARD, NEXT HYPE TRAIN LEAVING THE STATION AT PRECISELY... NOW!

We won't see any die stacking at all for N33. Only N31 and maybe N32.

 

[Saylick]

We won't see any die stacking at all for N33. Only N31 and maybe N32.

Die stacking AND MCM for Biggest RDNA3? What's the source?

 

[GodisanAtheist]

We won't see any die stacking at all for N33. Only N31 and maybe N32.

- Went back and read through the RDNA2 thread and people were so wildly off in the beginning of the rumor cycle it was a hoot (even with some really solid theorycrafting going on).

Wonder if we're going to get a refreshed 6nm "RDNA 2+" series that is not MCM but does have stacked dies as a sort of pipe-cleaner generation.

I'm really only solely basing this off of rumors of a Ryzen 5xxx XT refresh that will have Ryzen Zen3 dies with stacked cache. Apparently all current Ryzen Zen 3 dies already have the hardware in place for stacked cache, so this has been a thing that's in the roadmaps for a looong time.

 

[uzzi38]

Die stacking AND MCM for Biggest RDNA3? What's the source?

Uh, just to be clear, I don't mean compute die stacking, I mean something similar to V-Cache in that SoIC is used. And so far it's just a rumour. Don't treat it as anything more, because I have no clue if it's accurate either. Especially given how crazy some of these rumours are.

You know what, I'm just going to shut up for now. What's rumoured for N31 is so whacky that even if I trust who told me, I still can't really believe it. If it ends up actually being true and/or RGT or someone else reports on it, I'd probably be just as surprised as the rest of you at this point.

 

[Mopetar]

Every single tech company I have worked for has required that I sign NDAs. Maybe there are some out there that don't, but I would say they are very much in the minority.

Random people posting leaks and rumors on Twitter (probably) haven't signed any kind of NDA. Some of the people who may be leaking info to them may have, but unless you know their sources (and they won't release them) there's no way to know.

Any employee that did want to leak something isn't going to start use their own Twitter account, even if no one knows who they are. Having someone else leak it for them is far safer and makes it much harder for them to be found out.

 

[Saylick]

Uh, just to be clear, I don't mean compute die stacking, I mean something similar to V-Cache in that SoIC is used. And so far it's just a rumour. Don't treat it as anything more, because I have no clue if it's accurate either. Especially given how crazy some of these rumours are.

You know what, I'm just going to shut up for now. What's rumoured for N31 is so whacky that even if I trust who told me, I still can't really believe it. If it ends up actually being true and/or RGT or someone else reports on it, I'd probably be just as surprised as the rest of you at this point.

Aww, where's the fun in that?

What's the rumor mill say about expected performance gains over big RDNA2? I've heard reasonable guesses at 2x and then you got people like MLID saying he's heard 2.7x. I, too, would be flabbergasted at 2.7x performance increase (assuming that it is in pure rasterization and not RT).

 

[uzzi38]

Aww, where's the fun in that?

What's the rumor mill say about expected performance gains over big RDNA2? I've heard reasonable guesses at 2x and then you got people like MLID saying he's heard 2.7x. I, too, would be flabbergasted at 2.7x performance increase (assuming that it is in pure rasterization and not RT).

It's the 2.7x number again. Honestly though, that's not the craziest thing. Which... kind of says a lot.

 

[Timorous]

It's the 2.7x number again. Honestly though, that's not the craziest thing. Which... kind of says a lot.

Nothing is crazier than 2.7x. That has never been done before. 9700Pro did not manage it, 8800GTX did not manage it. Heck as good as RDNA2 is that did not manage anywhere close if you factor in that AMD did not manufacture a 300W RDNA part to make it more like for like.

You can talk about interesting layout ideas, crazy low TDPs etc but if AMD managed to pull off a 2.7x perf increase in a sub 400W TDP part it will be the 1st time in the history of GPUs and 3D accelerators that such a leap has been done.

 

[beginner99]

Nothing is crazier than 2.7x.

2.7x with half the power use? Not gonna happen but that would be a lot crazier.

I mean with tiles/MCM 2.7x isn't really an issue if you can scale. the only limit then is power.

But I suspect it's not only the performance number but how they get there, eg, the tech. Like 4ghz clocks or something else completely unheard of for gpus.