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

 

[RnR_au]

If Jack is serious about getting 40-50% market share, he should launch N48 XTX for $399

Folk would still buy Nvidia...

 

[beginner99]

If Jack is serious about getting 40-50% market share, he should launch N48 XTX for $399

I think $499 will do just fine if NV sells a 5070 12 GB for $599. Assuming it then actually sells for $499 outside of US. The difference here where I live tends to be smaller between the brands in terms of price (AMD more expensive than in US)

 

[ToTTenTranz]

I think $499 will do just fine if NV sells a 5070 12 GB for $599. Assuming it then actually sells for $499 outside of US. The difference here where I live tends to be smaller between the brands in terms of price (AMD more expensive than in US)

You should expect the 5070 to perform well above the Navi 48 XTX though, assuming it'll be between the 4070ti super and the 4080.

So if the price delta between the top Navi48 and the 5070 is 20% for 15% more performance, everyone is definitely going for nvidia again.

If AMD tries to undercut Nvidia's price/performance by $50 or less again, their abysmal marketshare is going to fade away completely.

 

[Kepler_L2]

I think $499 will do just fine if NV sells a 5070 12 GB for $599. Assuming it then actually sells for $499 outside of US. The difference here where I live tends to be smaller between the brands in terms of price (AMD more expensive than in US)

It won't. Reviewers will do the usual "AMD has better value and more VRAM, but the 5070 has DLSS and RTX and NVENC and CUDA..." and NVIDIA will continue with their 90% market share.

 

[SolidQ]

, but the 5070 has DLSS and RTX

if AMD gonna with FSR4, then reviewers gonna change their minds. About RT if N48 is real double RT engine, so should be around same perf both. Yes/no?

 

[marees]

if AMD gonna with FSR4, then reviewers gonna change their minds. About RT if N48 is real double RT engine, so should be around same perf both. Yes/no?

I expect RDNA 4 to match ada lovelace
Blackwell should be ahead

For ex:
navi 44 xt = 4060 ti 8gb
navi 48 xtx = 4070 ti super/4070 ti/4070 super?
5070 = 4080 ?

 

[Tuna-Fish]

I would expect the top model to be priced in line with NV, that is, too high to make sense, just to not leave money on the table. But the leaks describe a mildly cut variant of N48 with 16GB of 18gbps memory, I think that could possibly be very competitively priced.

 

[SolidQ]

navi 48 xtx = 4070 ti super/4070 ti/4070 super?

isn't that 7900XT perf? but there also rumors between 7900XT and XTX

 

[marees]

isn't that 7900XT perf? but there also rumors between 7900XT and XTX

I was referring to RT, AI upscaling features
Raster should be right around 7900xt

 

[marees]

But the leaks describe a mildly cut variant of N48 with 16GB of 18gbps memory, I think that could possibly be very competitively priced.

But what if the Navi 32 / 7800xt stocks are not sold out ?

 

[SolidQ]

that's an interesting info

https://twitter.com/x/status/1844355736031330379

CDNA4

 

[adroc_thurston]

that's an interesting info

That's not news. But they need to make it lower power, everyone else still leads AMD in offline playback BL by quite some margin.

 

[SolidQ]

But they need to make it lower power, everyone else still leads AMD in offline playback BL by quite some margin.

gonna be interesting to see upgrade in RDNA4 and HALO

 

[adroc_thurston]

HALO

-halo is the same(-ish) media core as Strix.
RDNA4 is discrete-only so you can't gauge just how efficient it really is in playback.

 

[Mahboi]

CDNA4

THIRTY FIVE TIMES

 

[Mahboi]

So, half NVLink 4's throughput? Thats not half bad for a first gen.

 

[SolidQ]

-halo is the same(-ish) media core as Strix.

but different instances?

https://twitter.com/x/status/1844356879771959775

 

[adroc_thurston]

but different instances?

That's a small revision.

 

[Mahboi]

So from what I gathered from it all:
- ROCm claims 2x to 3x improvement YoY
- Pensando stuff/Ultra Ethernet has started yielding its first gen and claims 400Go/s, while NVLink apparently claims 900Go/s
- Zen 5 EPYC is claiming a pretty massive improvement in general perf even core for core (unsurprisingly)
- almost everything launches H1 2025
- MI350X was announced for H2 2025 with 288Go of HBM3e, FP4 and FP6
- MI325X is here Soon:tm: with 256Go HBM3e and 6To/s

 

[soresu]

ROCm claims 2x to 3x improvement YoY

Improvement to what?

 

[Mopetar]

If Jack is serious about getting 40-50% market share, he should launch N48 XTX for $399

Doesn't matter what they charge if they don't order enough wafers to produce 50% of the chips that the market wants to consume.

The reality of an inexpensive but limited quantity N48 is that scalpers pay $399 for cards, but you still pay $499+.

 

[Tuna-Fish]

Doesn't matter what they charge if they don't order enough wafers to produce 50% of the chips that the market wants to consume.

AMD has absolutely not had this problem for a long time. Rather the opposite, they have ordered too many and they are sitting on shelves.

 

[gdansk]

AMD has absolutely not had this problem for a long time. Rather the opposite, they have ordered too many and they are sitting on shelves.

30 months ago N21 was selling for over MSRP.
I guess that's some time ago but I'm getting old so it seems like yesterday..

 

[Mahboi]

AMD has absolutely not had this problem for a long time. Rather the opposite, they have ordered too many and they are sitting on shelves.

Speaking of which, wth are they doing with RDNA 3? It's 5nm, just like Zen 4. They seem to be running out of Zen 4 chips relatively quick with Zen 5 not really selling, while having way too many RDNA 3.
Couldn't they just change their orders at TSMC and say "hey instead of making this much RDNA 3, could you change the masks/reset the machines to produce more Zen 4"?

 

[Tuna-Fish]

They have probably not made any new rdna3 for a while now.

It's not economically efficient to keep a chip in constant profuction unless there is enough demand to keep a line full, and no radeon chips qualify for this. Instead, AMD orders in lots.