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

 

[SolidQ]

Is that for real? Another 6500XT

 

[gaav87]

Is that for real? Another 6500XT

4x 5.0 349$ no hw encoders u need more ?
-Lisahawtuah
Meme aside would be horrible. But its amd they can shoot themselves in both feet at the same time.

 

[adroc_thurston]

Is that for real? Another 6500XT

That would be insanely based so I hope it's real.
God I love PPAmaxxing so much. More. MORE.

 

[Arctic Islands]

But RDNA4 was supposed to be monolithic, for both dies. Isn't even that settled?

No. We'll see.

 

[adroc_thurston]

No. We'll see.

Yes it is a monodie. Next.

 

[branch_suggestion]

Is that for real? Another 6500XT

Not an issue when paired with a laptop SoC.
Desktop Ryzen... yeah.

 

[Meteor Late]

It absolutely can be an issue even when you have an integrated GPU with encode support, local streaming is a thing, at the moment you cannot use the iGPU to encode while using the discrete GPU for gaming on the host with Moonlight and Sunshine.
AMD is just a disaster on GPU man

 

[marees]

That would be insanely based so I hope it's real.
God I love PPAmaxxing so much. More. MORE.

If they PPAmaxxed the N44 then what would be its launch price ?

& when do you think they'll launch it ???

 

[RNR_Forte]

I still think that AMD needed a card to fight toe-to-toe with the Nvidia halo card. At least amd maximum 4k performance. 8k is for the riches.

 

[Meteor Late]

I still think that AMD needed a card to fight toe-to-toe with the Nvidia halo card. At least amd maximum 4k performance. 8k is for the riches.

I think they tried to do that but failed, I mean the chiplet strategy, they are probably lying about focusing on the midrange as an overall planned strategy for a long time and instead they are just forced to do so after they plan about competing with chiplets failed.

 

[GTracing]

Is that for real? Another 6500XT

To be fair, the main issues with the 6500XT were the 4GB VRAM, the four PCIe lanes, and the price. The missing encode/decode engine was just the cherry on top.

 

[marees]

To be fair, the main issues with the 6500XT were the 4GB VRAM, the four PCIe lanes, and the price. The missing encode/decode engine was just the cherry on top.

AMD should have lied that msrp of 6500xt is $99

Crazy to launch at a price for which it actually sold on the market

 

[GTracing]

AMD should have lied that msrp of 6500xt is $99

Crazy to launch at a price for which it actually sold on the market

Yeah, that would have helped with the reception a lot. Though it's still a bad graphics card. It has issues at 1080p medium in some games.

 

[GodisanAtheist]

To be fair, the main issues with the 6500XT were the 4GB VRAM, the four PCIe lanes, and the price. The missing encode/decode engine was just the cherry on top.

-It was really all the 4gb RAM. There is an 8gb from Sapphire that runs fine for the specs, sometimes twice as fast as the stock version.

That card couldn't exist at launch though thanks to mining, the 8gb of Vram would have sent the price and the card would probably have ended up in the mines with 8gb of RAM and a lean power profile.

No reason we couldn't get that card today though.

I'd love to see a 24 CU 8GB clocked to the moon entry level card for $100-120. It would sell like hotcakes.

 

[Panino Manino]

That would be insanely based so I hope it's real.
God I love PPAmaxxing so much. More. MORE.

"I know GPU makers who prioritise PPA and they're all COWARDS."

Just sell it all to Intel.
Or Qualcomm, as they already did with the mobile part.

 

[RNR_Forte]

I think they tried to do that but failed, I mean the chiplet strategy, they are probably lying about focusing on the midrange as an overall planned strategy for a long time and instead they are just forced to do so after they plan about competing with chiplets failed.

The only time that they really tryed to do this was with the HD 2900 xt back in 2006.

 

[GTracing]

-It was really all the 4gb RAM. There is an 8gb from Sapphire that runs fine for the specs, sometimes twice as fast as the stock version.

That card couldn't exist at launch though thanks to mining, the 8gb of Vram would have sent the price and the card would probably have ended up in the mines with 8gb of RAM and a lean power profile.

No reason we couldn't get that card today though.

I'd love to see a 24 CU 8GB clocked to the moon entry level card for $100-120. It would sell like hotcakes.

A wider PCIe bus also would have increased performance. The combination of the two were what caused the terrible performance on the 6500 XT.

The fact is, you can't compromise on both VRAM capacity and PCIe bandwidth. To a degree you can get away with handicapping one or the other, but doing both is disastrous for performance. It also creates a situation where the 6500 XT is overly sensitive to exceeding the VRAM buffer, even by a little bit, whereas products such as the GTX 1650 Super have a lot more leeway.

AMD Radeon RX 6500 XT Review: A Bad, Really Bad Graphics Card

How to start this review... this thing is so bad, it's really spoiled the mood for me. In short, this is the worst GPU release since I...

www.techspot.com

 

[Win2012R2]

They didn't, N31 (and N36) had an X3D variant planned

And I planned to go out on a date with Queen of the Prom, same result...

 

[Win2012R2]

The only Blackwell that will (probably) get more cache than its predecessor will be GB202, at 128MB.

Works for me!

 

[Win2012R2]

Is that for real? Another 6500XT

Could be - AMD took customer feedback about how bad their hardware encoders are compared to Nvidia and even ARC and fixed the problem at the source...

 

[adroc_thurston]

about competing with chiplets failed.

It didn't fail, they just don't want to compete.

Just see it all to Intel.
Or Qualcomm, as they already did with the mobile part.

Intel is too poor and QC too incompetent. 0/2.

 

[Panino Manino]

Intel is too poor and QC too incompetent. 0/2.

But them it wouldn't be "my team" embarrassing itself.
Yes, I know it's an irrational feeling, it's like football teams, I can't help it since I "discovered" myself on the side of AMD back in the mid 1990.

If AMD really gave up them give up for real.
Stick doing "gaming" GPUs only for the IGPs, and at most launch only one or two discrete GPUs for generation. One the "sweet spot" mainstream and the other the low-end for old PCs. If they don't believe they can make real money to be truly competitive, or do something like this or cut their losses on exiting this market.

 

[adroc_thurston]

exiting this market.

Full exit is always too expensive to pull off.

Stick doing "gaming" GPUs only for the IGPs, and at most launch only one or two discrete GPUs for generation. One the "sweet spot" mainstream and the other the low-end for old PCs

well they're making three, one for the poor people, one for the middle class, and a kilobuck offering on top of these. But no halo or any fancy chiplet wondertech devices.

 

[reaperrr3]

The only time that they really tryed to do this was with the HD 2900 xt back in 2006.

That's just not true.

While R600 was relatively big for its time, it was smaller than Hawaii, Fiji, Vega, N21 and N31, and horribly underspecced for its overly ambitious 512bit mem interface, especially on TMUs and ROPs.

 

[Jan Olšan]

Is that for real? Another 6500XT

No, that's a good idea. Media engines take donkeyloads of die space and what use are they, really? Quality encoding needs to be done through software encoders, game streaming and game recording belongs on the CPU's integrated media engine because it's more efficient there.

It's inefficient to pay for the media engine footprint twice, and most AM5 CPUs have media engine (iGPU) now (the main exception is 7500F, with 8700F and 8400F in tow but those don't sell that much I think).

So this is 100% a smart thing to do. PCIe ×8 interfaces, too.

(But it probably will be used against AMD in the guerrilla marketing / volunteer shilling comments that tell you how paying more for the same performance on Nvidia cards is "the more you save", Huang math or something.)