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

 

[Meteor Late]

I've fought for AMD for a long time now, including 580 being a decent mid range card, and most every CPU since Ryzen (I'm still running a x370 with a 1700X > 5950 in it)... but they are taking the wee wee with RDNA4. Something has gonna majorly wrong in their GPU division. Ms CEO of the Year needs to get her house in order.

I think Coreteks nailed it, they tried the chiplets strategy with RDNA4 and failed miserably and had to backtrack, it's the only explanation. Even with NAVI, AMD talked about the 5700 XT along with the much expected Ryzen 3000 processors.

 

[Heartbreaker]

Some info was released to media though:

 

[adroc_thurston]

I think Coreteks nailed it, they tried the chiplets strategy with RDNA4 and failed miserably and had to backtrack, it's the only explanation. Even with NAVI, AMD talked about the 5700 XT along with the much expected Ryzen 3000 processors.

LMAO no, N4c never reached the TO stage to fail anywhere.

Ms CEO of the Year needs to get her house in order.

She did. She killed client discrete graphics more or less.

 

[blckgrffn]

They don't have a winner and they can't put a price sticker on it

Fully in wait and see mode on nvidia. Awesome. /s

 

[del42sa]

I think Coreteks nailed it, they tried the chiplets strategy with RDNA4 and failed miserably and had to backtrack, it's the only explanation. Even with NAVI, AMD talked about the 5700 XT along with the much expected Ryzen 3000 processors.

Well, given AMD's stance on dedicated graphics cards and how RDNA4 turned out, I think we can safely forget about chiplets for the RDNA5 generation as well ...

 

[adroc_thurston]

Well, given AMD's stance on dedicated graphics cards and how RDNA4 turned out, I think we can safely forget about chiplets for the RDNA5 generation as well ...

They're never happening in client indeed.

 

[Panino Manino]

I've fought for AMD for a long time now, including 580 being a decent mid range card, and most every CPU since Ryzen (I'm still running a x370 with a 1700X > 5950 in it)... but they are taking the wee wee with RDNA4. Something has gonna majorly wrong in their GPU division. Ms CEO of the Year needs to get her house in order.

Who would had thought that Radeon would get worse after Raja?
Raja must be laughing while smoking two cigars at the same time.

 

[adroc_thurston]

Who would had thought that Radeon would get worse after Raja?
Raja must be laughing while smoking two cigars at the same time.

It got much much better since it no longer bleeds cash.

 

[Bryo4321]

They're never happening in client indeed.

Guess I need a 5090 to cope.

Seriously though, that’s really disappointing but it’s impossible to ignore the huge contrast between the rdna 3 announcement and whatever we just saw was….
I understand the business case for it but yeesh this ain’t the Radeon I remember.

 

[adroc_thurston]

Seriously though, that’s really disappointing but it’s impossible to ignore the huge contrast between the rdna 3 announcement and whatever we just saw was….

RDNA3 still had ambition. That's fully and forever gone now.

but yeesh this ain’t the Radeon I remember.

Indeed but they've been kinda dead after the RV770 seppuku.

 

[Heartbreaker]

Who would had thought that Radeon would get worse after Raja?
Raja must be laughing while smoking two cigars at the same time.

RDNA 4 sounds pretty good, closing the feature gap.

After that it's mostly price.

 

[PJVol]

They're never happening in client indeed.

This sounds promising, if true )

 

[Bryo4321]

TPU article on RDNA 4

Interesting tidbit:
“We also got our first peek at what the "Navi 48" GPU powering the Radeon RX 9070 series looks like—it features an unusual rectangular die with a 2:1 aspect ratio, which seems to lend plausibility to the popular theory that the "Navi 48" is two "Navi 44" dies joined at the hip with full cache-coherency. The GPU is rumored to feature a 256-bit GDDR6 memory interface, and 64 compute units (4,096 stream processors). The "Navi 44," on the other hand, is exactly half of this (128-bit GDDR6, 32 CU). AMD is building the "Navi 48" and "Navi 44" on the TSMC N4P (4 nm EUV) foundry node, on which it is building pretty much its entire current-generation, from mobile processors, to CPU chiplets.“

 

[jpiniero]

On the assumption almost no one DIY will buy it they cut everything to get OEM deals in laptops. Which they won't get. Hard to understand the market for that part.

Well, what I was thinking was that they would price the N44 16 GB the same as the 5060 8 GB... Going slashy slashy would make it easier to do that as long as NV isn't generous on pricing. But it'd look bad against the 7700 XT at it's fire sale price given that N44 is probably slower than it.

 

[adroc_thurston]

This sounds promising, if true )

no, that means they surrendered.

TPU article on RDNA 4

Interesting tidbit:
“We also got our first peek at what the "Navi 48" GPU powering the Radeon RX 9070 series looks like—it features an unusual rectangular die with a 2:1 aspect ratio, which seems to lend plausibility to the popular theory that the "Navi 48" is two "Navi 44" dies joined at the hip with full cache-coherency. The GPU is rumored to feature a 256-bit GDDR6 memory interface, and 64 compute units (4,096 stream processors). The "Navi 44," on the other hand, is exactly half of this (128-bit GDDR6, 32 CU). AMD is building the "Navi 48" and "Navi 44" on the TSMC N4P (4 nm EUV) foundry node, on which it is building pretty much its entire current-generation, from mobile processors, to CPU chiplets.“

no, AMD just tall and skinny aspect ratios. see Navi21/31/whatever.

 

[Heartbreaker]

TPU article on RDNA 4

Interesting tidbit:
“We also got our first peek at what the "Navi 48" GPU powering the Radeon RX 9070 series looks like—it features an unusual rectangular die with a 2:1 aspect ratio, which seems to lend plausibility to the popular theory that the "Navi 48" is two "Navi 44" dies joined at the hip with full cache-coherency. The GPU is rumored to feature a 256-bit GDDR6 memory interface, and 64 compute units (4,096 stream processors). The "Navi 44," on the other hand, is exactly half of this (128-bit GDDR6, 32 CU). AMD is building the "Navi 48" and "Navi 44" on the TSMC N4P (4 nm EUV) foundry node, on which it is building pretty much its entire current-generation, from mobile processors, to CPU chiplets.“

IMO, ZERO chance of that.

They may have duplicated Navi 44 layout across, but it will be one piece of Silicon.

 

[PJVol]

no, that means they surrendered

What's stopping them from making big die?

 

[Meteor Late]

I'm surprised to see AMD using just GDDR6 while Nvidia is using GDDR7. This is a massive difference, though I don't know how much GDDR7 costs in comparison to GDDR6, but surely it is cheaper than having to use a higher bus? so a 192 bit bus GDDR7 card will have the advantage against a 256 bit bus GDDR6 with anything else equal right? and a terrible 128 bit bus GDDR7 card will almost be there against a 192 bit bus GDDR6 card.

 

[adroc_thurston]

What's stopping them from making big die?

Nvidia making a bigger die.

but surely it is cheaper than having to use a higher bus?

it's a product-level thing. RDNA4 supports GDDR7 just fine.

 

[PJVol]

Nvidia making a bigger die

So does intel? (in a budget segment)

 

[Heartbreaker]

I'm surprised to see AMD using just GDDR6 while Nvidia is using GDDR7. This is a massive difference, though I don't know how much GDDR7 costs in comparison to GDDR6, but surely it is cheaper than having to use a higher bus? so a 192 bit bus GDDR7 card will have the advantage against a 256 bit bus GDDR6 with anything else equal right? and a terrible 128 bit bus GDDR7 card will almost be there against a 192 bit bus GDDR6 card.

AMD may be prepping to bring very good value. Small die, and less expensive VRAM.

 

[adroc_thurston]

So does intel? (in a budget segment)

Well Intel isn't a good benchmark and my statement was about just not winning against NV with a bigger die since NV is always willing to ship moar.
Like N31 versus AD102 is NV shipping outright 50% more shader cores.

AMD may be prepping to bring very good value. Small die, and less expensive VRAM.

they, in fact, are not. lol.

 

[Meteor Late]

AMD may be prepping to bring very good value. Small die, and less expensive VRAM.

I don't know, it just feels like a half assed release, it just doesn't make sense overall.
GDDR7 brings a huge bandwidth increase, it is looking like a necessity, I never saw GPU vendors being so stingy with bus width, I mean, Nvidia is selling RTX 4070 Mobile cards with a 128 bit bus ffs! it probably scales terrible with new nodes, so bandwidth provided by memory seems very important. How is AMD going to compete with their 128 bit cards against Nvidia's 5060s 128 bit cards?

 

[gaav87]

Radeon 9070xt prime oc 3x-8power like iv been yapping about

 

[PJVol]

Like N31versus AD102 is NV shipping outright 50% more shader cores.

True, but still I'm not sure they couldn't keep up had they followed single-die design path

Кadeon 9070xt prime oc 3x-8power

Too bad my PSU doesn't have that many )