Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

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

branch_suggestion · Jan 22, 2025

Timorous said:
A theoretical N31 as a monolithic die would be about 400mm^2.

Nah, 450mm^2+ easy, analog and cache don't change much.

Heartbreaker · Jan 22, 2025

Heartbreaker said:
IMO, software isn't ready yet.

adroc_thurston said:
It is.

HW has to be locked down much longer before release, so what else could it be, but software?

Software is never really done and can always be made better. Especially with new ML base FSR, they likely had beta level for a while, but could use all the training they can get. Not to mention they may have scrambled on Multi-Frame Generation.

Plus now they claiming that's what they are delaying for...

adroc_thurston · Jan 22, 2025

Heartbreaker said:
HW has to be locked down much longer before release, so what else could it be, but software?

Software is never really done and can always be made better. Especially with new ML base FSR, they likely had beta level for a while, but could use all the training they can get. Not to mention they may have scrambled on Multi-Frame Generation.

They're literally waiting for Nvidia to drop the turds.

Heartbreaker · Jan 22, 2025

adroc_thurston said:
They're literally waiting for Nvidia to drop the turds.

That doesn't even make sense. They already know NVidia's pricing and NVidia's own performance claims.

If they had a ready product, it makes no sense to wait.

But if the software lacks polish it makes no sense to rush.

Bryo4321 · Jan 22, 2025

With the extra time I would hope the software stack is ready for release day.

branch_suggestion · Jan 22, 2025

Heartbreaker said:
That doesn't even make sense. They already know NVidia's pricing and NVidia's own performance claims.

They want to hijack the 70's launch to maximise exposure, makes a far bigger impact to have direct comp being on bar graphs.

gaav87 · Jan 22, 2025

adroc_thurston said:
They're literally waiting for Nvidia to drop the turds.

Well at least linux drivers still have some kinks they are working on. Spent like 3h reading notes and code.

Heartbreaker · Jan 22, 2025

branch_suggestion said:
They want to hijack the 70's launch to maximise exposure, makes a far bigger impact to have direct comp being on bar graphs.

If they could drop now, reviews would compare to 40 series and look better, and have product in the market first to sell before NVidia competition.

Launching first means they also get top billing again in NVidia reviews as well...

So again, waiting make no sense if they have a fully ready product.

adroc_thurston · Jan 22, 2025

Heartbreaker said:
If they could drop now, reviews would compare to 40 series and look better, and have product in the market first to sell before NVidia competition.

Wrong, people will just wait™ for Blackwell.

marees · Jan 22, 2025

799 ????

Timorous · Jan 22, 2025

branch_suggestion said:
Nah, 450mm^2+ easy, analog and cache don't change much.

Even without them changing at all the maths is the maths. 96MB L3 is 60mm on the MCDs as is 6x 64 bit GDDR6 PHYs. The GCD is 300mm of which 20mm is taken up by the GCD to MCD PHYs so that is not needed and you end up at 280 + 120 which is 400mm. Even if you use 90mm from the below diagram for the L3 size then absolute top end we are talking 430mm for monolithic N32 and it has 50% more stuff than N48. I just do not see N48 being 390mm with the spec it has on N4 which does offer a small shrink Vs N5.

I simply do not see N48, even with all of the extra transistors thrown at RT and the CUs I don't see it bloating up by that much.

gdansk · Jan 22, 2025

marees said:
799 ????

View attachment 115450

That's how you can tell it's made up. At no point was it considered to sell the 7800 XT successor for $800.

branch_suggestion · Jan 22, 2025

Timorous said:
Even without them changing at all the maths is the maths. 96MB L3 is 60mm on the MCDs as is 6x 64 bit GDDR6 PHYs. The GCD is 300mm of which 20mm is taken up by the GCD to MCD PHYs so that is not needed and you end up at 280 + 120 which is 400mm. Even if you use 90mm from the below diagram for the L3 size then absolute top end we are talking 430mm for monolithic N32 and it has 50% more stuff than N48. I just do not see N48 being 390mm with the spec it has on N4 which does offer a small shrink Vs N5.

The L3 including fabric was 90mm^2, 430mm^2 is about as low as N31 mono could go with Zen5 densified L3.
And N48 is ~350mm^2 based on more thorough measurements, smaller than GB203 and uses far cheaper and plentiful GDDR6.

Timorous said:
I simply do not see N48, even with all of the extra transistors thrown at RT and the CUs I don't see it bloating up by that much.

Cache.

adroc_thurston · Jan 22, 2025

marees said:
799 ????

View attachment 115450

Nope, never was that.

poke01 · Jan 22, 2025

branch_suggestion said:
And N48 is ~350mm^2 based on more thorough measurements, smaller than GB203 and uses far cheaper and plentiful GDDR6.

but NV sells that 378mm^2 die for $999 so the use of GDDR7 isn’t a cost issue for NV.

insertcarehere · Jan 22, 2025

gdansk said:
That's how you can tell it's made up. At no point was it considered to sell the 7800 XT successor for $800.

Judging by all those chonky cards we saw at CES, I am not sure '7800 XT successor' would have been the first thing that springs to mind here. AIBs don't manufacture these high-end (read: expensive) designs out there if their design brief (and margins) were for $400-$500 cards ala Navi 32.

gdansk · Jan 22, 2025

insertcarehere said:
AIBs don't manufacture these high-end (read: expensive) designs out there if their design brief (and margins) were for $400-$500 cards ala Navi 32.

They did for the 7800 XT.

AMD Radeon RX 7800 XT Specs

AMD Navi 32, 2430 MHz, 3840 Cores, 240 TMUs, 96 ROPs, 16384 MB GDDR6, 2438 MHz, 256 bit

www.techpowerup.com

Ctrl+F "Triple-slot" 1/13

branch_suggestion · Jan 22, 2025

poke01 said:
but NV sells that 378mm^2 die for $999 so the use of GDDR7 isn’t a cost issue for NV.

Because AMD didn't build anything that smashes it.
If they did, NV would have to charge less or would've built a larger GB203.

lucasworais · Jan 22, 2025

marees said:
799 ????

View attachment 115450

Maybe if it performs better than the 5070 Ti

Heartbreaker · Jan 22, 2025

marees said:
799 ????

View attachment 115450

Much more likely, that the people making these claims are just not reliable.

Rigg · Jan 22, 2025

gdansk said:
They did for the 7800 XT.

AMD Radeon RX 7800 XT Specs

AMD Navi 32, 2430 MHz, 3840 Cores, 240 TMUs, 96 ROPs, 16384 MB GDDR6, 2438 MHz, 256 bit

www.techpowerup.com

Ctrl+F "Triple-slot" 1/13

Yeah, @insertcarehere 's comment is completely out of touch with reality even on lower tier cards than 7800XT. There are multiple triple fan 7600 cards. There are even a few triple slot 7600 cards made by chinese AIBs FFS. I highly doubt that adding an oversized brick cooler and some unicorn vomit adds much to the BOM for them. The likes of Power Color and XFX don't even charge much of a premium for these types of cards.

Kaluan · Jan 22, 2025

adroc_thurston said:
They're literally waiting for Nvidia to drop the turds.

Yup. I mean, I'm not sure how more are not catching on to the obvious.
They want to see the ACTUAL review/day1 reception of 5070/5070Ti, do a final price adjustment if needed, then swoop in a try to seem they're "rescuing" us gamers from nVidia's Intel-ization of gaming GPUs.

If the reception is mildly positive/neutral, they'll price very aggressively.
If the reception is bad/very bad (which is the likelier of the 2), they'll price less aggressive, but still be very competitively marketed/priced... and likely plausibly be painted in a good light by "the press" and consumers (certainly one better than that nVidia can/will be in).
(particularly if the software is solid on review day/shortly post-launch)

Now, that being said, it's still possible AMD will bork it, even with the 12th hour "stand back and strategize" leeway.

Edit: IF this is "confirmed" to be the case, than fail or win... and as I much as I hate marketing and "PR", I'll hand it to AMD, it is very ballsy. Pissing off people even more short term but knocking it out of the park later on, knowing full well people nowadays have very short memory. Especially when you start telling them what they want to hear. Certainly one of the more... colorful... of product launches.

Dezii90 · Jan 23, 2025

4060 was reception was terrible, 4060ti reception was terrible, 4070 so so and amd still didn't capitalize and we expect them to do it this time? Why allow people to buy lovelace, blackwell without strong marketing to counter it then just wait for 5070/ti reviews. People will be hooked by Nvidia's new features and gimmicks before amd could even show their value.

igor_kavinski · Jan 23, 2025

Nvidia has already done their marketing "magic". 4090 performance from 5070, regardless of the caveats involved in the reality of that statement. Most sheep will happily buy that, convincing themselves that 4090 level performance at $549 is simply unbeatable value. This is why AMD needs to launch now. Software can be refined along the way. Holding up the launch to add FSR4 to more titles? Do they even understand that most people wanting to upgrade will grab it immediately based on just strong native performance alone?

insertcarehere · Jan 23, 2025

igor_kavinski said:
Nvidia has already done their marketing "magic". 4090 performance from 5070, regardless of the caveats involved in the reality of that statement. Most sheep will happily buy that, convincing themselves that 4090 level performance at $549 is simply unbeatable value. This is why AMD needs to launch now. Software can be refined along the way. Holding up the launch to add FSR4 to more titles? Do they even understand that most people wanting to upgrade will grab it immediately based on just strong native performance alone?

Coming from the same people who marketed FSR3 for like, a year before that got really implemented into games, to somehow believe that this time they can meaningfully improve FSR4 implementation in 60 days? Please don't take your potential market for fools AMD...

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Senior member

Diamond Member

Diamond Member

Diamond Member

Member

Senior member

Senior member

Diamond Member

Diamond Member

Senior member

Golden Member

Diamond Member

Senior member

Diamond Member

Diamond Member

Senior member

Diamond Member

Senior member

Member

Diamond Member

Senior member

Senior member

Junior Member

Lifer

Senior member