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

 

[Timorous]

N44 still only has 2/3 of 7700XT's memory bandwidth, so little chance of competing against the 7700XT (unless it's somehow massively less membw hungry)

Agreed which is why I said below the 7700XT. Probably fills the gap between 6700XT and 7700XT so around 4060Ti performance which would not be bad at $330 ish, especially with 16GB of vram.

 

[SolidQ]

Big problem we don't know how new WGP perfomance compared to RDNA3
at least hope N48 is 2x8pin, not new 12

 

[Timorous]

Big problem we don't know how new WGP perfomance compared to RDNA3

Of course so any guess is going to have a huge +/-, par for the course really but its more about getting an idea of what should be possible rather than trying to nail it 100%.

 

[PJVol]

Of course so any guess is going to have a huge +/-

Like what? Even if they are 1:1, 7900XT RT-less perf. at 215W and $400 looks already competitive with nv next gen. Add 5% to the average from the RT on steroids (BVH lookup) and you'll get somewhere inbetween 7900's xt and xtx for let's say $450-500

at least hope N48 is 2x8pin, not new 12

Wouldn't 8+6 (or even a single 8-pin) be enough?

 

[Kepler_L2]

https://twitter.com/x/status/1776553577928069464

The Navi 48 samples are capable of 50 TFLOPS on a ~215W board.

Yep, we're back.

 

[SolidQ]

even a single 8-pin)

75+150 = 225, no room for OC? Maybe 1x8pin is reference models, AIB 2x8 or 8+6

Yep, we're back.

People already like. Not impressive, RTX 4070super give 40TF at 2.8 ghz

 

[PJVol]

Yep, we're back.

Back to the April 1st ? )

 

[Timorous]

Like what?

IPC might go down or up. Clock speeds might be higher or lower than any assumptions made. Clock scaling may be totally different from prior gens etc.

Any guess like 'a 64 cu RDNA 4 part with 3Ghz clocks should perform like an OCd 7900 GRE' is full of such assumptions which means it can only ever be at best a rough guide and even that requires the assumptions to be broadly there or there abouts. If something fundamentally changes then it can throw the whole thing out.

An example would be trying to guess 5700XT perf based on specs when using GCN as a baseline, RDNA scales very differently so this kind of speculation would be a mile off.

 

[PJVol]

An example would be trying to guess 5700XT perf based on specs when using GCN as a baseline

Or like guessing N31 perf. based on tflops?
On a serious note, if this 50tf is not another 01/04 thingy, isn't it reasonable to assume the performance per CU at least hasn't degraded

 

[SolidQ]

if this 50tf is not another 01/04 thingy

7900XT is 51,5/52TF 84CU at 300W
N48 is 50TF 64CU at 215
I don't think it's 1/04

 

[Timorous]

Or like guessing N31 perf. based on tflops?
On a serious note, if this 50tf is not another 01/04 thingy, isn't it reasonable to assume the performance per CU at least hasn't degraded

We didn't know the Tflops until launch because we didn't know the clock speeds. Speculation was based on shader count and expected clocks which were both well off due to dual issue and the v/f curve being awful.

 

[SolidQ]

We didn't know the Tflops

Also we don't know how much RDNA3 1TF compared to RDNA4 TF

 

[PJVol]

We didn't know the Tflops until launch because we didn't know the clock speeds. Speculation was based on shader count and expected clocks

Yep, I mean shaders count

 

[adroc_thurston]

Also we don't know how much RDNA3 1TF compared to RDNA4 TF

Irrelevant metrics.
Final PPA is all that matters.

 

[SolidQ]

Irrelevant metrics.

We're talking not about revelent/irrelevant.
it's can be like RDNA4 50TF = RDNA3 60TF. So that why people compare

 

[Timorous]

Also we don't know how much RDNA3 1TF compared to RDNA4 TF

Technically they are equal but I know you mean we don't know the fps/Tflop ratio. Sure hence why the estimates would have plenty of +/- because it might go up or down.

All things being equal on that front then 3Ghz clocks should get you around OC 7900GRE / 7900XT perf. If IPC is lower then it will be worse and if IPC is more it will be higher.

And yea. PPA is the relevant metric for AMD because that determines margin. Would not be the 1st time AMD have a good PPA advantage over NV but don't do much with it because they don't build a big enough die to drive home the advantage.

 

[adroc_thurston]

it's can be like RDNA4 50TF = RDNA3 60TF. So that why people compare

useless metric.

If IPC is lower then it will be worse and if IPC is more it will be higher.

it's a SIMD machine it does not have 'IPC'.

All things being equal on that front then 3Ghz clocks should get you around OC 7900GRE / 7900XT perf.

It clocks higher than that.

they don't build a big enough die

Yes, now they have to build a big enough (1.5k+ mm^2 active Si) MCP.

 

[Mahboi]

Yep, we're back.

Group hug at AMD.
The dark days of gimped crap are leaving us. \_O_/

 

[SolidQ]

useless metric.

i know, but alot people love those comparision.
it's fine if N48 ~ 7900XT + new uarch give advantage in some games + driver not mature yet for RDNA4.
Only big question is left, N48 faster in RT than N31 or not.
Also price is matter, if 600 then meh, if 500 is fine.

 

[adroc_thurston]

driver not mature yet for RDNA4.

Doesn't even remotely matter those days.

Only big question is left, N48 faster in RT than N31 or not.

Doesn't matter either.

Also price is matter, if 600 then meh, if 500 is fine.

It's a tiny die.
They can sell it for $399 but, obviously, they will not.

 

[PJVol]

Unless they still have no bvh lookup added to isa it's not a question I think

 

[SolidQ]

Doesn't matter either.

For alot people already matter, almost every message, like "I chose NV because RT/DLSS etc", even if it's slower card, than from AMD

 

[adroc_thurston]

For alot people already matter, almost every message, like "I chose NV because RT/DLSS etc"

They choose NV because it has the green sticker.
That's it.
Been like that since 2006.
Won't change.

 

[Mahboi]

Like what? Even if they are 1:1, 7900XT RT-less perf. at 215W and $400 looks already competitive with nv next gen.

$400?
This isn't even optimism anymore.
They'll try at least $500 and probably start at $600 because they'll need to finish GimpeDNA 3 stock.

 

[SolidQ]

They choose NV because it has the green sticker.

That i know, but you see amd starting lose more since no competitor to GTX 9xx