Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

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...

forums.anandtech.com

 

[marees]

The "Ultimate Navi 48" is said to have a higher GPU core frequency, expected to be between 3.2GHz and 3.4GHz, and will be equipped with a higher frequency GDDR6 video memory, and the power consumption may exceed 340W.

In addition, the graphics card may be equipped with 24GB or 32GB of video memory to meet the needs of high-end games and professional applications.

AMD还有杀招！Navi 48“终极版”GPU曝光：介于RTX 5070 Ti和5080

快科技3月7日消息，据最新爆料，AMD正在研发一款名为“Ultimate Navi 48”的RDNA 4架构显卡，这款显卡预计将在2025年第二季度发布。这款显卡性能定位介于NVIDIA的RTX 5070 Ti和RTX

m-mydrivers-com.translate.goog

 

[lucasworais]

The "Ultimate Navi 48" is said to have a higher GPU core frequency, expected to be between 3.2GHz and 3.4GHz, and will be equipped with a higher frequency GDDR6 video memory, and the power consumption may exceed 340W.

In addition, the graphics card may be equipped with 24GB or 32GB of video memory to meet the needs of high-end games and professional applications.

AMD还有杀招！Navi 48“终极版”GPU曝光：介于RTX 5070 Ti和5080

快科技3月7日消息，据最新爆料，AMD正在研发一款名为“Ultimate Navi 48”的RDNA 4架构显卡，这款显卡预计将在2025年第二季度发布。这款显卡性能定位介于NVIDIA的RTX 5070 Ti和RTX

m-mydrivers-com.translate.goog

No gddr7. Too bad

 

[Win2012R2]

Not when the the company making a 600 card is being told next time to make a halo card because they can’t compete

I think you've missed the point of that video completely - it shows that 9070 XT in 4K (!) provides excellent 100-ish fps frame rate, obviously 5090 is faster - but not even 2 times for 5 times the price (plus possible house fire).

And yes AMD simply must make a halo GPU next time - bring back Big Navi v2!

No gddr7. Too bad

 

[Bryo4321]

The "Ultimate Navi 48" is said to have a higher GPU core frequency, expected to be between 3.2GHz and 3.4GHz, and will be equipped with a higher frequency GDDR6 video memory, and the power consumption may exceed 340W.

In addition, the graphics card may be equipped with 24GB or 32GB of video memory to meet the needs of high-end games and professional applications.

AMD还有杀招！Navi 48“终极版”GPU曝光：介于RTX 5070 Ti和5080

快科技3月7日消息，据最新爆料，AMD正在研发一款名为“Ultimate Navi 48”的RDNA 4架构显卡，这款显卡预计将在2025年第二季度发布。这款显卡性能定位介于NVIDIA的RTX 5070 Ti和RTX

m-mydrivers-com.translate.goog

How would you end up with 24gb? That doesn’t seem possible. And what faster gddr6 is there? Factory OC?

 

[marees]

what faster gddr6 is there? Factory OC?

20gbps is the lowest common denominator across diff manufacturers (same as PS5)

If you go to a specific memory maker then faster memory should be possible

 

[Hail The Brain Slug]

The fastest GDDR6 I remember hearing about was 24gbps from Samsung. That would give them 20% more bandwidth. There also was supposed to be 3GB packages but they never seem to have materialized.

I'd be surprised if they managed to find those MIA 3GB packages. But using 24gbps would be great.

 

[adroc_thurston]

The fastest GDDR6 I remember hearing about was 24gbps from Samsung.

Doesn't exist.

There also was supposed to be 3GB packages but they never seem to have materialized.

Also doesn't exist.
Not even in JESD250D.

20gbps is the lowest common denominator across diff manufacturers (same as PS5)

PS5 is like 14Gbps tho.

 

[Hail The Brain Slug]

Doesn't exist.

Also doesn't exist.
Not even in JESD250D.

PS5 is like 14Gbps tho.

24gbps is real, it was sampled by samsung but never entered mass production.

3GB packages were announced, but never materialized.

 

[Josh128]

Meh. The frequency and power is already juiced enough on XT...the price is too, now. They should just produce as many of those as possible and not waste time on this.

The "Ultimate Navi 48" is said to have a higher GPU core frequency, expected to be between 3.2GHz and 3.4GHz, and will be equipped with a higher frequency GDDR6 video memory, and the power consumption may exceed 340W.

In addition, the graphics card may be equipped with 24GB or 32GB of video memory to meet the needs of high-end games and professional applications.

AMD还有杀招！Navi 48“终极版”GPU曝光：介于RTX 5070 Ti和5080

快科技3月7日消息，据最新爆料，AMD正在研发一款名为“Ultimate Navi 48”的RDNA 4架构显卡，这款显卡预计将在2025年第二季度发布。这款显卡性能定位介于NVIDIA的RTX 5070 Ti和RTX

m-mydrivers-com.translate.goog

 

[adroc_thurston]

24gbps is real, it was sampled by samsung but never entered mass production.

That's the cartelese for "doesn't exist".

3GB packages were announced, but never materialized.

Not even in the JEDEC spec!
Only GDDR7 supports 24Gb ICs.

 

[Hail The Brain Slug]

That's the cartelese for "doesn't exist".

Not even in the JEDEC spec!
Only GDDR7 supports 24Gb ICs.

Can you stop being pedantic for like one second? Is it even physically possible for you?

 

[igor_kavinski]

Can you stop being pedantic for like one second? Is it even physically possible for you?

If he really is a military guy, the short answer is NO

 

[Mopetar]

Can you stop being pedantic for like one second? Is it even physically possible for you?

* Puts on @adroc_thurston suit *

No.

 

[adroc_thurston]

Can you stop being pedantic for like one second?

No.

Is it even physically possible for you?

I just kinda am a memory guy.

 

[igor_kavinski]

I just kinda am a memory guy.

OK, let's hear your thoughts on the memristor.

Why was it good? Why was it bad? Was there any potential?

Is it possible for Optane to make a comeback?

Why don't they arrange CPU cores AROUND the RAM slots?

Why not put the RAM slots BEHIND the CPU socket (shorter distance so better latency, no?) ?

I get that AMD doesn't trust Intel but shouldn't they have bought the rights to Optane and used it for expanding the storage capacity of Radeons instead of constantly shuffling data over the slow PCIe bus? Wouldn't that have been awesome for loading super large LLMs?

 

[adroc_thurston]

Why was it good? Why was it bad? Was there any potential?

Meme.

Is it possible for Optane to make a comeback?

Chalcogenide PCMs inherently suck at being cost effective.
3DXP made it worse since crosspoint structure means every layer has to be fabbed individually.

Why don't they arrange CPU cores AROUND the RAM slots?

What

Why not put the RAM slots BEHIND the CPU socket (shorter distance so better latency, no?) ?

Tracelen impacts not much but signal integrity.

I get that AMD doesn't trust Intel but shouldn't they have bought the rights to Optane and used it for expanding the storage capacity of Radeons instead of constantly shuffling data over the slow PCIe bus

This is insane and you should know it.

 

[blackangus]

Is it possible for Optane to make a comeback?

SCM isnt likely going anywhere, however it also will likely remain a niche product for the near/mid future.

 

[adroc_thurston]

SCM isnt likely going anywhere, however it also will likely remain a niche product for the near/mid future.

It's not a tier issue per se, it's just that 3DXP sucks.
There's a damn good reason Micron basically never shipped it.

 

[blackangus]

It's not a tier issue per se, it's just that 3DXP sucks.
There's a damn good reason Micron basically never shipped it.

Well I mean from the cost/use case standpoint. (Niche)
But I dont disagree with what you are saying from a technology standpoint.
Lets leave it at this for this is far from this threads topic.

 

[gaav87]

24gbps is real, it was sampled by samsung but never entered mass production.

3GB packages were announced, but never materialized.

Ye i said the same in this thread after i heard about the 24gb rdna4 like 20 days ago. But its in sampling since like 2years

That's the cartelese for "doesn't exist".

Not even in the JEDEC spec!
Only GDDR7 supports 24Gb ICs.

eda vendors provide the memory controlled IP i think amd uses synopsys like for their mi300 series or am i wrong?
Didnt they have had a board with it validated and tested anyway ?

 

[branch_suggestion]

https://twitter.com/x/status/1901885804185886739

 

[darkswordsman17]

Meh. The frequency and power is already juiced enough on XT...the price is too, now. They should just produce as many of those as possible and not waste time on this.

I agree. But it seems that competency (aka, they didn't completely screw up the launch while everyone is thoroughly sick of Nvidia, or rather, they can't even get Nvidia cards even at bonkers prices) of RDNA4 surprised even AMD, so they're looking for some way to ruin it. Maybe if they were putting in the effort to bin chips to enable this, but based on how well these do undervolting it seems like AMD is still not binning their stuff much if at all beyond basic levels. I feel like AMD should've maybe done something similar to R9 Fury, and done 3 tiers of their premium cards. A lower spec one (9070) that could be put on a smaller board, with low power use/decent efficiency (Nano), a standard one that they expect most to buy (9070XT) that with just some voltage tweaking would've been more efficient without losing performance, and then a higher tier 9070XTX which would be maxed out for performance and come with more VRAM.

Personally, I'll keep sitting out the entire market if I can't get reasonably priced hardware. There's tons of older games that I can play on old hardware and have a blast. All the while, the games improve and eventually I'll be able to buy newer hardware that will run it even better still. Personally none of the realistic graphics look good enough that they don't still look like games when actually playing them, so I have no interest in the massive hardware improvements needed to keep pushing incremental improvements.

They don’t have enough balls for that

Most likely they didn't expect Nvidia to screw up the 5000 series like they did. Although not sure why, AI is hitting critical stupid mass and so people are buying cards just for that purpose and they'll only go for Nvidia. Granted they probably didn't expect the DeepSeek situation, but they clearly weren't completely obtuse (Strix Halo with 128GB is clearly targeting AI). The Blackwell situation is likely because of AI (both them pushing it out the door when it has issues and then not having enough stock for gaming, which is further strained as people are buying the gaming cards for AI, with DeepSeek showing you don't need the dedicated AI chips/cards to do AI), so doubtful AMD could have predicted that it would happen this way. If Blackwell was typical Nvidia launch people wouldn't be going bananas for RDNA4 nearly as much. Don't get me wrong, I think they'd be popular but a bunch of people that bought one would've already bought a 5080 and/or 5070Ti had they been available. I've encountered several.

AMD would've had to make the decision on a higher tier GPU years back, and they'd already made that decision, but it was for chiplets, which haven't worked out yet. They probably would've had to make the RDNA4 decision to go monolith large by the RDNA3 launch to have it ready, and that was what 2 years ago at this point?

 

[Josh128]

.

Personally, I'll keep sitting out the entire market if I can't get reasonably priced hardware. There's tons of older games that I can play on old hardware and have a blast. All the while, the games improve and eventually I'll be able to buy newer hardware that will run it even better still. Personally none of the realistic graphics look good enough that they don't still look like games when actually playing them, so I have no interest in the massive hardware improvements needed to keep pushing incremental improvements.

We agree on this. Its probably not good for the hardware industry to have that mindset, but it is what it is. The point where visuals and hardware capability actually still held back aspects of gameplay was passed around 20 years ago. Everything since then has been incremental....and reaching into that last 5% in whatever hobby you are into always requires exponential cost / price increases. This is where we are in video game hardware and software, IMO.

 

[Kepler_L2]

AMD would've had to make the decision on a higher tier GPU years back, and they'd already made that decision, but it was for chiplets, which haven't worked out yet. They probably would've had to make the RDNA4 decision to go monolith large by the RDNA3 launch to have it ready, and that was what 2 years ago at this point?

Navi4C was cancelled in late January 2023.

 

[Jan Olšan]

So, what did AMD say about Cooperative Vectors support at GDC? I think the feature was presented yesterday?