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

 

[xilli_fiberbit]

I don't think it's realy a launch of MI200, but a validation/test phase for some customers ?

 

[moinmoin]

One doesn't validate against a moving target.

 

[gdansk]

It's shipping to select customers. I wouldn't normally call that a launch. But in CDNA's case who else but select customers buy those accelerators.

 

[uzzi38]

I don't think it's realy a launch of MI200, but a validation/test phase for some customers ?

Initial shipments, the validation cycle is long since over. Official launch is already confirmed for this year:

Advanced Micro Devices, Inc. (AMD) CEO Lisa Su Presents at JPMorgan 49th Annual Global Technology, Media and Communications Conference (Transcript)

Advanced Micro Devices, Inc. (NASDAQ:NASDAQ:AMD) JPMorgan 49th Annual Global Technology, Media and Communications Conference May 24, 2021 8:45 AM ETCompany...

seekingalpha.com

 

[uzzi38]

It's shipping to select customers. I wouldn't normally call that a launch. But in CDNA's case who else but select customers buy those accelerators.

It'll be available in systems you can buy from specific OEMs as well, just like MI100 was. Pretty sure it had a couple of Dell systems or something anyway.

 

[leoneazzurro]

Unless I missed something, the patents suggested there would not be an IO die? Did something change?

I wouldn’t complain if the entire product line was on 6nm TBH, my gut tells me that there is still room for growth.

EDIT: Also, any parts that are on 6nm should have great availability as Zen 4 enters the market.

For I/O I meant mainly the video input/outputs - I don't think AMD will replicate video circuitry for a GCD when these are thought to be used in a MCM configuration only. Of course there will be also cache in this interconnecting die. About the video memory interconnections, it's difficult to say if they will be located on the GCDs or the cache die, I personally think the latter is more probable, but who knows at the moment.

 

[Kepler_L2]

It's really unfortunate AMD isn't sharing info like they did at the beginning of Zen. CDNA2's MCM setup should relate to the one in RDNA3.

Not really, CDNA doesn't have infinity cache and a lot of other stuff.

 

[moinmoin]

Not really, CDNA doesn't have infinity cache and a lot of other stuff.

So you expect the kind of MCM interconnects to differ between CDNA2 and RDNA3?

 

[Kepler_L2]

So you expect the kind of MCM interconnects to differ between CDNA2 and RDNA3?

Yes, MI200 is just basic LSI or something similar. RDNA3 is a lot fancier.

 

[Gideon]

So you expect the kind of MCM interconnects to differ between CDNA2 and RDNA3?

Oh they will differ for sure. People have claimed in interviews for a while that multi-chip gpu in datacenter isn't that hard (they run tons of them in parallel anyway) but executing client workloads is a much more complex problem.

This is also highlighted in all the AMD patents regarding the subject, some of which I've shared in this very thread (and I'm far from being the only one to do so):

Page 7 - Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Page 7 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

https://www.reddit.com/r/Amd/comments/nrr9gl

 

[TESKATLIPOKA]

Some info about N33
Twitter

Additional info

Spoiler

perf:n33>n21

the very bad thing is

Spoiler

only 8GB Vram

 

[Mopetar]

Some info about N33
Twitter

Additional info

Spoiler

perf:n33>n21

the very bad thing is

Spoiler

only 8GB Vram

There's an issue with the Twitter link for me. For whatever reason it's adding extra garbage to the URL. I'm not sure if it's the forum software or my phone. Here's what it's supposed to be.

https://twitter.com/x/status/1422542778223976449

Edit: Is it even possible to have more than 8 GB with a 128-bit bus? I'm not aware of any company making larger memory chips that you'd need to get beyond 8, at least that I'm aware of.

 

[Leeea]

Edit: Is it even possible to have more than 8 GB with a 128-bit bus? I'm not aware of any company making larger memory chips that you'd need to get beyond 8, at least that I'm aware of.

Yes

The DDR4 ram you most likely have your computer has a 64-bit bus, and odds are you have more then 8 GB.

 

[blckgrffn]

Yes

The DDR4 ram you most likely have your computer has a 64-bit bus, and odds are you have more then 8 GB.

Pretty sure GDDR and DDR are not the same in crucial ways.

Now, don't expect me to know them off hand

 

[Gideon]

AMD announces Radeon Pro W6800X Duo with two Navi 21 GPUs for Apple Mac Pro - VideoCardz.com

New AMD Radeon PRO W6000X Series GPUs Bring Groundbreaking High-Performance AMD RDNA 2 Architecture to Mac Pro – New AMD GPUs harness the high-performance, energy efficient AMD RDNA™ 2 architecture, AMD Infinity Cache and other advanced technologies to power demanding professional design and...

videocardz.com

So Mac Pro has a W6800X with two N21 chips with 120 CUs enabled, 256mb of L3 and 2x 256bit memory interface (64GB total), all in a 400W TGP. Judging by the TFLOPS rating it's also clocked only a bit lower than the 60CU version.


This bodes pretty well for N31, especially if this rumor is true and it's really on 5nm with 6nm cache:

AMD Navi 31, Navi 32 and Navi 33 GPU block diagrams by Olrak visualize RDNA3 rumors - VideoCardz.com

AMD RDNA3 architecture rumors visualized through GPU block diagrams Olrak delivers new block diagrams visualizing AMD’s upcoming RDNA3 graphics architecture. The last two weeks have been filled with leaks and rumors on next-gen AMD GPUs, which brought a large list of details and even more data...

videocardz.com

 

[TESKATLIPOKA]

There's an issue with the Twitter link for me. For whatever reason it's adding extra garbage to the URL. I'm not sure if it's the forum software or my phone. Here's what it's supposed to be.

https://twitter.com/x/status/1422542778223976449

Edit: Is it even possible to have more than 8 GB with a 128-bit bus? I'm not aware of any company making larger memory chips that you'd need to get beyond 8, at least that I'm aware of.

Link should work. I had some extra garbage in It. Fixed It within some minutes, but It looks like you were faster.

Yes, you can have even 16GB Vram with a 128bit bus. You can have 8* 2GB chips, but It's costly.

 

[Mopetar]

Link should work. I had some extra garbage in It. Fixed It within some minutes, but It looks like you were faster.

Yes, you can have even 16GB Vram with a 128bit bus. You can have 8* 2GB chips, but It's costly.

Aren't the GDDR memory interfaces 32-bit though? That would imply needing 4x 4GB and I don't think anyone makes 4GB memory modules. I don't even know if there are 3 GB modules yet.

At least the plus side to this is that it will be relatively inexpensive if the mining craze continues. Etherium is memory bandwidth limited and a 128-bit bus really limits what miners will pay for the card. It'll still go for more than the suspected MSRP, but not the ridiculous amounts we're seeing now.

 

[DeathReborn]

Aren't the GDDR memory interfaces 32-bit though? That would imply needing 4x 4GB and I don't think anyone makes 4GB memory modules. I don't even know if there are 3 GB modules yet.

At least the plus side to this is that it will be relatively inexpensive if the mining craze continues. Etherium is memory bandwidth limited and a 128-bit bus really limits what miners will pay for the card. It'll still go for more than the suspected MSRP, but not the ridiculous amounts we're seeing now.

GDDR6 is possible to use in x8 Clamshell mode, alloing you to double the number of chips used, so 8x 2GB chips can be used on 128bit. GDDR6 is a 2x16 device, so that is why Clamshell on that is x8 while GDDR5 Clamshell meant x16.

GDDR6 Deep Dive

 

[insertcarehere]

There's an issue with the Twitter link for me. For whatever reason it's adding extra garbage to the URL. I'm not sure if it's the forum software or my phone. Here's what it's supposed to be.

https://twitter.com/x/status/1422542778223976449

So let me get this straight, Navi 23 is ~23x mm^2 on TSMC N7, and now it's nominal successor is supposed to be nearly twice that size on more expensive TSMC N6? This is supposedly the "affordable" gpu for more mainstream gamers.

So either AMD struck the most lopsided contract alive with TSMC for wafers or that AMD isn't planning on selling any performance GPUs for less than $500 MSRP next generation...

 

[Stuka87]

Aren't the GDDR memory interfaces 32-bit though? That would imply needing 4x 4GB and I don't think anyone makes 4GB memory modules. I don't even know if there are 3 GB modules yet.

At least the plus side to this is that it will be relatively inexpensive if the mining craze continues. Etherium is memory bandwidth limited and a 128-bit bus really limits what miners will pay for the card. It'll still go for more than the suspected MSRP, but not the ridiculous amounts we're seeing now.

The bus can technically be divided up anyway you want, provided its broken up in 8bit increments. So you could have four chips with a 32bit interface, or you could have eight chips with a 16bit interface. You would lose some speed per chip, but its distributed across more chips. This is how the 3090 is setup, twenty four chips on a 384bit bus. You just end up with some chips sharing bandwidth with other chips.

 

[jpiniero]

or that AMD isn't planning on selling any performance GPUs for less than $500 MSRP next generation...

There's always selling Navi 2x below that.

 

[GTracing]

So let me get this straight, Navi 23 is ~23x mm^2 on TSMC N7, and now it's nominal successor is supposed to be nearly twice that size on more expensive TSMC N6? This is supposedly the "affordable" gpu for more mainstream gamers.

So either AMD struck the most lopsided contract alive with TSMC for wafers or that AMD isn't planning on selling any performance GPUs for less than $500 MSRP next generation...

Unlike Nvida, RTG doesn't assign codenames so that they correspond to a product tier. Remember, Navi 10 was the 5700XT (et al.). N33 isn't necessarily the successor to N23; there could be another RDNA3 design that is closer to N23.

 

[Lodix]

In which Q is expected for the first RNDA 3 product to be announced?

Sent from my SM-G998B using Tapatalk

 

[exquisitechar]

In which Q is expected for the first RNDA 3 product to be announced?

Sent from my SM-G998B using Tapatalk

Last we heard (rather recently), the parts hadn't even been taped out yet. So Q3 2022 at the earliest, I'd say.

 

[Ajay]

Last we heard (rather recently), the parts hadn't even been taped out yet. So Q3 2022 at the earliest, I'd say.

Wow, didn’t realize it hasn’t taped out yet! I thought it was typically 12-15 months to reach volume production after tapeout + time to fill the channel. Maybe my memory is betraying me.