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

 

[jpiniero]

Oh but it is a new WGP.

The leakers are implying that it has basically the same raw specs as N32 yet is 30% smaller even excluding the small density gain of N4P.

 

[adroc_thurston]

The leakers are implying that it has basically the same raw specs as N32

it's a fair bit faster than that lmao.

yet is 30% smaller even excluding the small density gain of N4P.

Part of it is the thing entirely being on N4p and not paying for USRs but yes, much more efficient IP on core PPA metrics.

 

[SolidQ]

Would be cool if RDNA4 can beat at least 5070.

Based on this. Maybe 5070 = 4080, so RDNA4 maybe going compete with RTX 5060ti

 

[adroc_thurston]

Based on this. Maybe 5070 = 4080, so RDNA4 maybe going compete with RTX 5060ti

the guy doesn't know anything, bad idea to quote him

 

[SolidQ]

bad idea to quote him

Still true about RDNA4/5 fight?

 

[adroc_thurston]

Still true about RDNA4/5 fight?

no, again, the guy doesn't know anything.

 

[SolidQ]

no, again, the guy doesn't know anything.

ok we will see in future

 

[adroc_thurston]

ok we will see in future

not much to see here.

 

[PJVol]

It's not a driver issue, AMD has a problem with resets

If you mean the same issue on which Alex Deucher spent most of his time, then yeah... it's kinda nightmare

 

[Abwx]

That s supposed to be RDNA4 specs for two GPUs fabbed using N4P.

https://twitter.com/x/status/1776244864856051785

 

[TESKATLIPOKA]

Assuming this leak is legit, I guess I need to kneel to Teska.

Please, don't. I don't even remember what was that about.

 

[adroc_thurston]

Please, don't. I don't even remember what was that about.

I assume something something leadership PPA aka my favourite rhetoric.

 

[TESKATLIPOKA]

The leakers are implying that it has basically the same raw specs as N32 yet is 30% smaller even excluding the small density gain of N4P.

It just shows interconnects used up a lot of space.

 

[adroc_thurston]

It just shows interconnects used up a lot of space.

Naa, USRs are tiny.
It's N5 tiny SRAM shrink + a whole lot better IP in Navi4.

 

[TESKATLIPOKA]

So the specs look like this for now:

RDNA4

Size

SE

Clocks

WGP

CU

Dual-issue Shaders

TMU

ROPs

Infinity Cache

bus width

GDDR6

Mem BW

Effective BW IC BW

Vram

N48

~240mm2

4

?

32

64

4096 ?

256 ?

96 ?

64MB ?

256-bit

21.65gbps

693 GB/s

2770 GB/s

16GB

N44

~130mm2

2 ?

?

16

32

2048 ?

128 ?

64 ?

32MB ?

128-bit

18gbps

288 GB/s

515 GB/s

8-16GB ?

WGP is supposedly new so I don't know what changes that means for the specs.

N44 doesn't look that good for higher resolutions thanks to the underwhelming memory subsystem.

edit: Ok, It's probably just IC BW and not effective BW.

edit2: also included memory BW

 

[Mahboi]

So the specs look like this fro now:

Effective BW

2770 GB/s

515 GB/s

A 4090 with a 384 bit bus is around 21Gbps. Are we sure that's Effective B/W on N48?

 

[TESKATLIPOKA]

A 4090 with a 384 bit bus is around 21Gbps. Are we sure that's Effective B/W on N48?

Ok, probably It's just the BW from Infinity cache.
Infinity cache BW:
N32: 2708 GB/s
N33: 476.9GB/s

 

[Mahboi]

Are you sure about what you wrote?
RTX 4090: 384-bit paired with 21gbps GDDR6 = 1008 GB/s

Effective BW is Infinity cache hitrate + BW from memory controller.

Nope, not sure. That's why it's a question.
The number just seems very high compared to N44. It's five times more.
I'd be more willing to bet that it's the estimated clocks at 2770Mhz and 2515 Mhz, but the 2 was missed or something.

 

[Timorous]

So the specs look like this for now:

RDNA4

Size

SE

Clocks

WGP

CU

Dual-issue Shaders

TMU

ROPs

Infinity Cache

bus width

GDDR6

Effective BW

Vram

N48

~240mm2

4

?

32

64

4096 ?

256 ?

96 ?

64MB ?

256-bit

21.65gbps

2770 GB/s

16GB

N44

~130mm2

2 ?

?

16

32

2048 ?

128 ?

64 ?

32MB ?

128-bit

18gbps

515 GB/s

8-16GB ?

WGP is supposedly new so I don't know what changes that means for the specs.

N44 doesn't look that good for higher resolutions thanks to the underwhelming memory subsystem.

It will be fine at 1080p / 1440p with compromises or upscaling and N48 will be fine for 1440p and 4k with compromises or upscaling.

 

[TESKATLIPOKA]

Nope, not sure. That's why it's a question.
The number just seems very high compared to N44. It's five times more.
I'd be more willing to bet that it's the estimated clocks at 2770Mhz and 2515 Mhz, but the 2 was missed or something.

You were too fast. I changed that post of mine, check It out.
I think It's very unlikely he was talking about clockspeed.

 

[Mahboi]

It will be fine at 1080p / 1440p with compromises or upscaling and N48 will be fine for 1440p and 4k with compromises or upscaling.

32 CUs of non-broken RDNA 3/RDNA 3.5 would already probably handle 1440p quite well. The current broken 7600 is already very close. Assuming 20% extra perf, it would honestly surprise me that it can't reach 6700xt or so performance. And with the expected RDNA 4 improvements, maybe it'll be closer to 7700 xt performance...except it has that tiny bus.
Same with 64 CUs of non broken RDNA 3, that would already gain a solid 15% extra, it would put the 7800 xt around 4070 Ti performance? And then RDNA 4 improvements on top of it. I don't think there's going to be a lot of compromises compared to the current top range. Unless you think a 7900 xt is too weak for 4K.

 

[Abwx]

So the specs look like this for now:

RDNA4

Size

SE

Clocks

WGP

CU

Dual-issue Shaders

TMU

ROPs

Infinity Cache

bus width

GDDR6

Effective BW IC BW

Vram

N48

~240mm2

4

?

32

64

4096 ?

256 ?

96 ?

64MB ?

256-bit

21.65gbps

2770 GB/s

16GB

N44

~130mm2

2 ?

?

16

32

2048 ?

128 ?

64 ?

32MB ?

128-bit

18gbps

515 GB/s

8-16GB ?

WGP is supposedly new so I don't know what changes that means for the specs.

N44 doesn't look that good for higher resolutions thanks to the underwhelming memory subsystem.

edit: Ok, It's probably just IC BW and not effective BW.

256b bus 693 GB/s and 2770MHz with 64CUs
128b bus 288 GB/s and perhaps 3515MHz along with 32CUs, would be odd that frequency is only 2515MHz for the smaller GPU.

 

[TESKATLIPOKA]

32 CUs of non-broken RDNA 3/RDNA 3.5 would already probably handle 1440p quite well. The current broken 7600 is already very close. Assuming 20% extra perf, it would honestly surprise me that it can't reach 6700xt or so performance. And with the expected RDNA 4 improvements, maybe it'll be closer to 7700 xt performance...except it has that tiny bus.
Same with 64 CUs of non broken RDNA 3, that would already gain a solid 15% extra, it would put the 7800 xt around 4070 Ti performance? And then RDNA 4 improvements on top of it. I don't think there's going to be a lot of compromises compared to the current top range. Unless you think a 7900 xt is too weak for 4K.

That depends on what you mean by handle quite well. >60FPS?

In raster most likely yes, unless It's a game like City Skylines II.
RT is another story.

P.S. I am talking about N44

 

[TESKATLIPOKA]

256b bus 693 GB/s and 2770MHz with 64CUs
128b bus 288 GB/s and perhaps 3515MHz along with 32CUs, would be odd that frequency is only 2515MHz for the smaller GPU.

What? I have question marks for clock.
I don't think this was meant for me.

 

[Abwx]

What? I don't think this was meant for me.

You confused the GPU frequency with the BW, 2770 is the GPU frequency not the BW.
Same for the other GPU.