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

 

[igor_kavinski]

Maybe like this?
8600 = 4060ti+
8600XT = ~7800XT
8800 = ~7900XT
8800XT = ~7900XTX

My prediction:

8600 = >4060
8600XT = >7700XT
8800 = ~7900XT (agree)
8800XT = ~7900XTX (agree)

 

[igor_kavinski]

isn't it's cost now like 4060ti?

In UAE, 4060 Ti 8GB is cheaper than XFX 7700 XT by about $14 which makes the 7700 XT good value at least from VRAM perspective.

 

[maddie]

Yeah no way this is true with the info we already have about die size.

Or the die size is wrong.

 

[SolidQ]

8600 = >4060

In't 7600 is already = 4060? N44 should have much higher clocks

 

[marees]

FP8

Mb 8700 is 12GB N48?
for me doesn't make sense skip 8700 unless, it's like was with NV 8xxx series only 8600/8800
I think possible like this
8600 - N44 8GB
8600XT - N48 12GB

8800 - N48 16GB
8800XT - N48 16GB

The are 5 skus

• N48 xtx — 64 CU 16gb (8800xt)
• N48 xt — 56 CU 16gb? (8700xt)
• N48 xl — 48 CU? 12gb (8600xt - may launch Q3 2025)
• N44 xt — 32 CU 8gb (8500xt)
• N44 xl — 28 CU? 8gb (8500 - may not launch)

 

[igor_kavinski]

In't 7600 is already = 4060? N44 should have much higher clocks

Maybe. Depends on if AMD will let it run at highest possible clock or if it will restrict the speeds in firmware to upsell 8800. I honestly don't expect them to compete properly with Nvidia, not after what they did with RX 7600. They will pretend that they are offering great value to existing RX 6600 XT and RX 7600 owners.

 

[ToTTenTranz]

8600XT = ~7800XT

N44 reaching N32 performance sounds like a (very big) stretch.

 

[marees]

N44 reaching N32 performance sounds like a (very big) stretch.

Correct

N44 < 7700xt

 

[eek2121]

Don’t compare model numbers, just price points.

AMD really needs to rebrand. As much as I dislike “Ryzen AI” I would take “Radeon AI 670” over the current scheme any day of the week.

 

[SolidQ]

N44 reaching N32 performance sounds like a (very big) stretch.

That about 8600XT N48 12GB, not 8600 N44
Unless is N48 12GB is 8700, but there no lines in code for 8700

AMD really needs to rebrand

That can easily use for UDNA 9xxx name, like was R300, shoe new arch, new power etc

 

[marees]

8800
&
8600

https://twitter.com/x/status/1864239965598544302

 

[marees]

8800
&
8600

https://twitter.com/x/status/1864239965598544302

AMD confirms Radeon RX 8600 and RX 8800 RDNA4 GPU series​

https://videocardz.com/newz/amd-confirms-radeon-rx-8600-and-rx-8800-rdna4-gpu-series

 

[Tuna-Fish]

The are 5 skus

• N48 xtx — 64 CU 16gb (8800xt)
• N48 xt — 56 CU 16gb? (8700xt)
• N48 xl — 48 CU? 12gb (8600xt - may launch Q3 2025)
• N44 xt — 32 CU 8gb (8500xt)
• N44 xl — 28 CU? 8gb (8500 - may not launch)

I would be surprised if the 12GB sku was not called 8700XT. It would continue the trend with both 6700XT and 7700XT. This would of course make the two 16GB models a bit awkward, but they might use XTX and XT again.

N44 with 8GB is probably called 7600, with possible clamshell model (if any) getting the XT again.

 

[misuspita]

They will pretend that they are offering great value to existing RX 6600 XT and RX 7600 owners.

Noone gives a hoot what AMD thinks. If they price them competitively it will sell. If not...

 

[adroc_thurston]

or if it will restrict the speeds in firmware to upsell 8800.

You can't really upsell a die this different and 2 segments higher.

 

[jpiniero]

Noone gives a hoot what AMD thinks. If they price them competitively it will sell. If not...

Define competitively. Compared to nVidia or compared to AMD's old products at fire sale prices?

 

[misuspita]

Define competitively. Compared to nVidia or compared to AMD's old products at fire sale prices?

Compared to the relative performance it has to near performing Nvidia boards. I hope this time AMD won't be doing another 6600-7600 "leap" of performance

 

[jpiniero]

Compared to the relative performance it has to near performing Nvidia boards. I hope this time AMD won't be doing another 6600-7600 "leap" of performance

Assuming the 5060 is using GB207, it's likely to not be any faster than the 4060. So N44 should def be faster.

The 5070 should be somewhere between the 4070 Ti Non Super and Super. N48... we should see.

 

[Kepler_L2]

Assuming the 5060 is using GB207, it's likely to not be any faster than the 4060. So N44 should def be faster.

The 5070 should be somewhere between the 4070 Ti Non Super and Super. N48... we should see.

5060 desktop is using GB206. GB207 is for laptop 5060 and desktop 5050.

 

[CastleBravo]

5060 desktop is using GB206. GB207 is for laptop 5060 and desktop 5050.

Are you sure about that? If 5060 is on GB206, I would expect it to be cut down enough to leave room for a 5060 Ti and/or 5060 Super.

 

[Kepler_L2]

Are you sure about that? If 5060 is on GB206, I would expect it to be cut down enough to leave room for a 5060 Ti and/or 5060 Super.

5060Ti is using a cutdown GB205.

 

[Joe NYC]

So how hard is it to break Nvidia's mindshare? Do they need fail like Intel? I mean AMD did with Ryzen, who knows maybe will.

Depends on Strix Halo reception - if it is good, AMD cam make more headway in mobile market, replacing NVidia dGPU that in desktop dGPU.

Where, BTW, most (unit wise) NVidia dGPUs are sold.

 

[darkswordsman17]

Don’t compare model numbers, just price points.

AMD really needs to rebrand. As much as I dislike “Ryzen AI” I would take “Radeon AI 670” over the current scheme any day of the week.

I fear RDNA4 might be dead in the water if they don't have ample supplies here prior to higher tariffs. It'd wipe out AMD's value attempt, with Intel cutting that out from them entirely with Battlemage. And the lower the price, the less of a hit the percentage adds to the overall price, so they'll still likely undercut AMD, and they have more parity on features, with Intel likely having better video processing (could be important for game streaming, especially with budget conscious people).

If I were AMD I'd factor in the tariff price increase in the MSRP. They can't completley obliterate it (i.e. make the MSRP be the price they intend minus the expected tariff), but they'd be smart to announce a lower price, expecting the final price with tariff to be closer to their original MSRP. It helps reduce the tariff some, plus makes them look better and they can blame the pricing on the tariffs entirely. I have a hunch Intel did this with Battlemage.

Anyone pushing rumors that it performs better than expected are malicious. It is clearly not mere stupidity anymore.

I don't know why they do it.

I mean, its like people forget that Nvidia spent years running a clandestine propaganda campaign coupled with a lot of other tactics (Gameworks breaking games on release, trying to force reviewers into insane agreements in order to receive cards to test, I forget what all else)? I'm not even saying they are perpetrating it still, but once that mindset gets into people it is very difficult to break it. It took years of sustained competitive products coupled with borderline disastrous issues by Intel to really break that. The situation is almost reversed in GPUs. Couple it with desperation by many and general common overhype that happens in the industry (happens pretty often with regards to Nvidia too but people rarely get upset about it). Add in the social media bias (everything has to be hyperbolic to try and get attention), plus AMD taking some big swings (compute heavy GPUs with GCN then Vega, chiplet design - which had pretty quickly radically altered the CPU market) and its easy to see why this keeps happening.

 

[carrotmania]

I fear RDNA4 might be dead in the water if they don't have ample supplies here prior to higher tariffs. It'd wipe out AMD's value attempt,

Cos Nvidia are gonna eat the tariff cost? That's nice of them!

 

[Tuna-Fish]

Cos Nvidia are gonna eat the tariff cost? That's nice of them!

Rather, any new low-end GPU has to compete with used old stuff, and a tariff on all the new stuff makes that hard.