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

 

[adroc_thurston]

why can’t AMD do a full lineup like that for RDNA4 and 5?

Because it costs money and the resulting sales numbers will be about 3 units in total.
Navi4C was cool but that's a lot of engineering for a hopeless market outcome.

It's clear that offering +5-10% raster at the same price point while having generally worse features, software, and no CUDA is just not worth it for most people.

Yes I know.
Need that 3.5G fmax to win.

 

[gaav87]

Because it costs money and the resulting sales numbers will be about 3 units in total.
Navi4C was cool but that's a lot of engineering for a hopeless market outcome.

Yes I know.
Need that 3.5G fmax to win.

Isn't 9070xt at 2.8G ?

 

[adroc_thurston]

Isn't 9070xt at 2.8G ?

No? No one dumped fmax or the v/f curve for N48 just yet.

 

[gaav87]

No? No one dumped fmax or the v/f curve for N48 just yet.

I was talking about the leaked AIB timespy score of ~23K
6800 (non-xt) / 20k TS score = 2.6G
7800xt / 2.9-3.0G for 23k
7900gre / 2.45G for 23k
Speculation W/O any arch changes vs 7800xt/rx6800 +4cu we are looking at ~2.8G (for that ~23k score)

If it reaches 3.0G and release drivers i could see it at 7900xt TS score.
Not talking gaming improvments.

 

[Win2012R2]

Need that 3.5G fmax to win.

Doable with N3P in 2026?

If AMD can do cheaply 3D cache for CPUs, why isn't it done for GPUs?

128 MB worked wonders in RDNA2 and after that it's gone downhill, AMD go BIG or go home - 256 MB unified 3D cache in UDNA1 is needed, their own stats shown that 128 MB is min for 4K and clearly gains will be very healthy for 256 MB.

The way things go in 2026 we'll finally have good 8k monitors and NVidia will yet again be ahead leaving AMD only scraps.

 

[Saylick]

I was talking about the leaked AIB timespy score of ~23K
6800 (non-xt) / 20k TS score = 2.6G
7800xt / 2.9-3.0G for 23k
7900gre / 2.45G for 23k
Speculation W/O any arch changes vs 7800xt/rx6800 +4cu we are looking at ~2.8G (for that ~23k score)

If it reaches 3.0G and release drivers i could see it at 7900xt TS score.
Not talking gaming improvments.

If the rumored 220W TDP stock is true, I hope AIBs give people 300W options, assuming RDNA4 is a good overclocker. But knowing history, it’s probably not lol

 

[jpiniero]

If the rumored 220W TDP stock is true, I hope AIBs give people 300W options, assuming RDNA4 is a good overclocker. But knowing history, it’s probably not lol

Probally not going to be worth it unless you get a good memory overclock.

 

[gaav87]

Probally not going to be worth it unless you get a good memory overclock.

At least they giving 20GT/s instead of 19.5 (garbage bin) this time around so 22GT/s (2750mhz) daily oc is possible.

 

[ToTTenTranz]

Doable with N3P in 2026?

If AMD can do cheaply 3D cache for CPUs, why isn't it done for GPUs?

128 MB worked wonders in RDNA2 and after that it's gone downhill, AMD go BIG or go home - 256 MB unified 3D cache in UDNA1 is needed, their own stats shown that 128 MB is min for 4K and clearly gains will be very healthy for 256 MB.

The way things go in 2026 we'll finally have good 8k monitors and NVidia will yet again be ahead leaving AMD only scraps.

Not sure they need that much LLC once they go with GDDR7. At least not for a while.

 

[gaav87]

If the rumored 220W TDP stock is true, I hope AIBs give people 300W options, assuming RDNA4 is a good overclocker. But knowing history, it’s probably not lol

So good AIB +maxed power slider ~265W. Hope the power play tables are not locked this time around...
I can see this at 3.1G and 2750mhz mem 265w surpasing 7900xt

 

[marees]

At least they giving 20GT/s instead of 19.5 (garbage bin) this time around so 22GT/s (2750mhz) daily oc is possible.

The LCM is 20 GT/s

The real speed depends on who is the supplier of the memory chip that you got

 

[gaav87]

The LCM is 20 GT/s

The real speed depends on who is the supplier of the memory chip that you got

Well at worst 2700mhz daily oc at best 2800mhz 21.5/22.5 so yeah w/e +-2%

 

[Saylick]

So good AIB +maxed power slider ~265W. Hope the power play tables are not locked this time around...
I can see this at 3.1G and 2750mhz mem 265w surpasing 7900xt

That’s all they need to do. Offer 7900XT performance, even if it takes an OC to get there, at $500 max.

 

[vanplayer]

If the rumored 220W TDP stock is true, I hope AIBs give people 300W options, assuming RDNA4 is a good overclocker. But knowing history, it’s probably not lol

IIRC reference card is 6+8pin while some non-ref cards are designed to be up to insane 8+8+8pin. Sorry I forget where I saw the leak.

 

[jpiniero]

That’s all they need to do. Offer 7900XT performance, even if it takes an OC to get there, at $500 max.

AIBs don't normally OC the memory (unless the chip is rated for higher)... right?

Also I would bet there would be a premium for those.

 

[gaav87]

AIBs don't normally OC the memory (unless the chip is rated for higher)... right?

Also I would bet there would be a premium for those.

No i was talking about manual oc.
Also its unreal that it only scores 23k in timespy when my oced 6800xt scores 23300...
There must be some shenanigans going on and release score will be closer to 25-26k.
Else its such, a turd even 400$ price wont save it.

 

[SolidQ]

We need game benchs, but less than 2 weeks is left

 

[Saylick]

No i was talking about manual oc.
Also its unreal that it only scores 23k in timespy when my oced 6800xt scores 23300...
There must be some shenanigans going on and release score will be closer to 25-26k.
Else its such, a turd even 400$ price wont save it.

What clocks and power does your OC’ed 6800XT achieve?

 

[jpiniero]

No i was talking about manual oc.
Also its unreal that it only scores 23k in timespy when my oced 6800xt scores 23300...
There must be some shenanigans going on and release score will be closer to 25-26k.
Else its such, a turd even 400$ price wont save it.

Average for TS Graphics on 3dmark is 19k... that's a decent OC.

 

[Panino Manino]

I agree RDNA2 was fantastic, why can’t AMD do a full lineup like that for RDNA4 and 5?

They can't even do the names right, it's so telling about their lack of confidence in themselves and their product.

 

[jpiniero]

They can't even do the names right, it's so telling about their lack of confidence in themselves and their product.

I think they used 7 because it's meant to compete with the 5070...

 

[Kepler_L2]

I think they used 7 because it's meant to compete with the 5070...

Correct

 

[adroc_thurston]

128 MB worked wonders in RDNA2 and after that it's gone downhill, AMD go BIG or go home - 256 MB unified 3D cache in UDNA1 is needed, their own stats shown that 128 MB is min for 4K and clearly gains will be very healthy for 256 MB

They're never building another big config ever again.

 

[linkgoron]

They can't even do the names right, it's so telling about their lack of confidence in themselves and their product.

Giving Navi 31 the 7900 moniker was a dumb move, and a seven moniker (8700 or 9070 or whatever) for Navi 48 makes much more sense.

I think that going with the 9070 specifically is a weird decision. It's taking Nvidia's naming scheme, and starting with a 9 makes this look like a one time thing. We'll see how it goes, I guess. It worked out with the r3/r5/r7 vs i3/i5/i7, but Ryzen was much more competitive, had a top to bottom stack and also started from 1000 (ancient members would also remember the Athlon "plus" naming days that were IMO also basically a way to reference to Intel) .

 

[jpiniero]

Well they did use 9700. Granted it was a long time ago.