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

 

[Krteq]

Some more info from Locuza

https://twitter.com/x/status/1410602684604846081

https://twitter.com/x/status/1410609918705475595

 

[uzzi38]

Greymon55 on Twitter: "navi33 monolithic tsmc 6nm" / Twitter

Yeah, so remember before I suggested the idea of a split product lineup between 5nm and 6nm for RDNA3?

 

[Gideon]

Greymon55 on Twitter: "navi33 monolithic tsmc 6nm" / Twitter

Yeah, so remember before I suggested the idea of a split product lineup between 5nm and 6nm for RDNA3?

Anything else makes little sense IMO. AMD can also maximize wafer throughput that way

 

[jrdls]

Greymon55 on Twitter: "navi33 monolithic tsmc 6nm" / Twitter

Yeah, so remember before I suggested the idea of a split product lineup between 5nm and 6nm for RDNA3?

Out of curiosity, why do you trust greymon55? His account has been created recently and doesn't have a track record with leaks that can be evaluated. I hope you don't mind me asking.

 

[DisEnchantment]

Out of curiosity, why do you trust greymon55? His account has been created recently and doesn't have a track record with leaks that can be evaluated. I hope you don't mind me asking.

He is probably one of those hydra leaker accounts, which the #silicongang is tracking
You take down one (which they did) , and another one will take its place.

 

[Leeea]

He is probably one of those hydra leaker accounts, which the #silicongang is tracking
You take down one (which they did) , and another one will take its place.

That seems rather unlikely.

When has twitter ever banned someone for leaking legal information?


the fortnight/bf6 guys were running giveaways, against the TOS.
Kottmann was banned for hacking into peoples systems and posting the results on twitter.

I have not seen twitter go after people for just sharing legal information.

 

[leoneazzurro]

By the way:

Pawsey Prepares for the Data Deluge from the Square Kilometre Array

As the largest-ever radio telescope, the Square Kilometre Array (SKA) will be a behemoth. As the name implies, the instruments of the massive radio telescope will span well over one square kilometer, using hundreds of dishes and hundreds of thousands of low-frequency aperture array telescopes...

www.hpcwire.com

This should use the MI200 GPUs.

 

[uzzi38]

That seems rather unlikely.

When has twitter ever banned someone for leaking legal information?


the fortnight/bf6 guys were running giveaways, against the TOS.
Kottmann was banned for hacking into peoples systems and posting the results on twitter.

I have not seen twitter go after people for just sharing legal information.

Twitter didn't. Other people from Chinese forums did because the person in question made some comments on Twitter that he shouldn't have, especially in regards to this guy's "source" of information.

Anyway, it's an old rumour that goes back a few months.

 

[uzzi38]

Btw @DisEnchantment it's the same guy we discussed about in DMs just a couple of days ago. He made a new Twitter account that he wants to keep a bit more secret now

 

[GodisanAtheist]

Two

That seems rather unlikely.

When has twitter ever banned someone for leaking legal information?


the fortnight/bf6 guys were running giveaways, against the TOS.
Kottmann was banned for hacking into peoples systems and posting the results on twitter.

I have not seen twitter go after people for just sharing legal information.

- Largest chips (250mm2 up) on 5nm, smaller chips stay on an enhanced, cheaper, high volume enhanced 7nm process.

Good way to get GPU volume out there without competing with the contracted parts and eating into their CPU production.

 

[uzzi38]

Out of curiosity, why do you trust greymon55? His account has been created recently and doesn't have a track record with leaks that can be evaluated. I hope you don't mind me asking.

Sorry, I forgot to reply to this one directly. It's someone that was on the Twitter space until recently before other Chinese forum dwellers got angry at him spreading info on Twitter. And yes he had legitimate information, courtesy of one guy on said Chinese forums who is without a doubt a legitimate leaker.

 

[Kepler_L2]

Greymon55 on Twitter: "navi33 monolithic tsmc 6nm" / Twitter

Yeah, so remember before I suggested the idea of a split product lineup between 5nm and 6nm for RDNA3?

I find this a bit hard to believe. The Navi3x GCD should be an SoC (unlike Ryzen CCDs) so it makes sense that Navi33 would be a single GCD by its own (with or without 3D IC if more B/W is needed). Also an 80 CU die on 6nm would be at a minimum 400mm², that's way too expensive for anything mid-range.

 

[uzzi38]

I find this a bit hard to believe. The Navi3x GCD should be an SoC (unlike Ryzen CCDs) so it makes sense that Navi33 would be a single GCD by its own (with or without 3D IC if more B/W is needed). Also an 80 CU die on 6nm would be at a minimum 400mm², that's way too expensive for anything mid-range.

You haven't accounted for the fact that cost per transistor is rising with newer nodes, not falling. Die size doesn't matter all that much if it'll end up being more costly on N5 than N6 anyway.

 

[Kepler_L2]

You haven't accounted for the fact that cost per transistor is rising with newer nodes, not falling. Die size doesn't matter all that much if it'll end up being more costly on N5 than N6 anyway.

That is offset by the fact you are essentially backporting the Navi3x 5nm GCD to 6nm. Unless the volume of Navi33 is absurdly high (an order of magnitude above any previous AMD GPU launch) there is no measurable cost savings.

 

[uzzi38]

That is offset by the fact you are essentially backporting the Navi3x 5nm GCD to 6nm. Unless the volume of Navi33 is absurdly high (an order of magnitude above any previous AMD GPU launch) there is no measurable cost savings.

Backporting? You're making the assumption that AMD's IP isn't already highly node agnostic.

Even dead broke AMD made sure that the original Zen was node agnostic, forget current year AMD that can't grow fast enough to fully invest the oodles of cash they're making.

 

[Stuka87]

When has twitter ever banned someone for leaking legal information?

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

 

[Glo.]

I find this a bit hard to believe. The Navi3x GCD should be an SoC (unlike Ryzen CCDs) so it makes sense that Navi33 would be a single GCD by its own (with or without 3D IC if more B/W is needed). Also an 80 CU die on 6nm would be at a minimum 400mm², that's way too expensive for anything mid-range.

N33 is designed to be as cheap as possible and as fast and powerful as possible.

Thats why AMD can ride the 6 nm node performance for this GPU.

Which kinda could tell you for which market segment this GPU is slated.

 

[Ajay]

N33 is designed to be as cheap as possible and as fast and powerful as possible.

Thats why AMD can ride the 6 nm node performance for this GPU.

Which kinda could tell you for which market segment this GPU is slated.

I thought 6N was more about an area reduction, rather than any significant performance or efficiency improvement? Of course, there are other areas of improvement available from a design perspective.

 

[Glo.]

I thought 6N was more about an area reduction, rather than any significant performance or efficiency improvement? Of course, there are other areas of improvement available from a design perspective.

The general idea behind what the rumors say about N33 is this.

"Its witchcraft."

And that should tell us everything we need to know. Nobody who heard the rumors about N33 is able to believe in them. Its not that they don't make sense.

Its just they are so good, that are truly hard to believe.

 

[Panino Manino]

Its just they are so good, that are truly hard to believe.

For what I remember RDNA2 far surpassed my expectations.
I was seeing everyone believing and optimistic here and I was all doom and gloom.
"no way it would clock that high, no way it would be so efficient, no way it would reach parity with Nvidia on rasterization, no way it would even match Turing on RT".
I was wrong about everything.
I don't expect the same degree of miracle but now I'm more willing to believe that it'll be in fact "good" and AMD will make more progress than most expect. Will I be wrong this time?

 

[Glo.]

I don't expect the same degree of miracle but now I'm more willing to believe that it'll be in fact "good" and AMD will make more progress than most expect. Will I be wrong this time?

What do you expect, precisely?

 

[GodisanAtheist]

What do you expect, precisely?

-Witchcraft.

 

[DisEnchantment]

I thought 6N was more about an area reduction, rather than any significant performance or efficiency improvement? Of course, there are other areas of improvement available from a design perspective.

TSMC has hinted that the N6 performance is similar to N7+ EUV node. N6 has one more EUV step compared to N7+
18% logic density gain which isn't much, ~10% efficiency gain and low single digit performance over N7. At worse similar perf with N7P with better power efficiency.
Also in 2021, TSMC is pushing most N7 customers to N6 because they can achieve higher WPM with N6.

 

[moinmoin]

Also in 2021, TSMC is pushing most N7 customers to N6 because they can achieve higher WPM with N6.

Which is interesting considering the former is DUV only while the latter replaces layers with EUV. That step apparently significantly simplifies the production for TSMC to push its customers to there.

 

[Ajay]

TSMC has hinted that the N6 performance is similar to N7+ EUV node. N6 has one more EUV step compared to N7+
18% logic density gain which isn't much, ~10% efficiency gain and low single digit performance over N7. At worse similar perf with N7P with better power efficiency.
Also in 2021, TSMC is pushing most N7 customers to N6 because they can achieve higher WPM with N6.

Well, I can see why companies are choosing N6 then (in addition to the PDK/RTL being very similar to that of N7).
I find this amazing, the NXE:3400C series EUV lithographs max out at 170 wph and the 'B's are around 125 wph. The original target for 'commercial success' was originally 200 wfh. I wonder what has changed.