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

 

[CouncilorIrissa]

RDNA40%

https://videocardz.com/newz/amd-radeon-rx-9070-series-gaming-performance-leaked-rx-9070xt-is-42-faster-on-average-than-7900-gre-at-4k

 

[gaav87]

Native no fsr frame gen

 

[gaav87]

549$

 

[Philste]

I get 37% over GRE in Raster and 53% over GRE in RT, in Computerbase Parcour that would mean Raster=4080 RT=4070Ti Super

 

[poke01]

AMD is the true hardware king these days.

Saving both CPU and GPUs in client, just pump up the production

 

[SolidQ]

9070nonXT seems 4070S in RT, maybe should be a bit faster than 5070 in RT?

 

[poke01]

AMDs biggest regret this decade will be not going with a halo this gen.

They could have owed NV, destroyed them in terms of leadership and engineering. RDNA4 isn’t even on N3E and it boasts these impressive figures.

Really excited for reviews, I wasn’t at all interested in Blackwell but this truly impressive more impressive than Zen5 IMO.

 

[jpiniero]

9070nonXT seems 4070S in RT, maybe should be a bit faster than 5070 in RT?

The 5070 is definately going to be slower than the 4070 Super, so yeah.

 

[Philste]

Let's not forget that is AMD first party gaming Benches which all had been off by a huge margin since Vanilla ZEN4. They advertised 50-70% over 6950XT for 7900XTX and ended at 35% at the end. Also Vanilla ZEN5 Claims were hilarious.

 

[marees]

5070 Ti should be about

6% faster in raster perf and about
23% faster in RT perf:

https://twitter.com/x/status/1893631198116606285

 

[jpiniero]

Looked closer to me in RT. I was using the Compubase reviews tho. It should be slower than the 5070 Ti but we're talking only a couple percent.

But we'll see...

 

[marees]

Looked closer to me in RT. I was using the Compubase reviews tho. It should be slower than the 5070 Ti but we're talking only a couple percent.

But we'll see...

I am guessing 15% slower in RT & worse in "PT"

 

[SolidQ]

I am guessing 15% slower in RT & worse in "PT"

but price also different and that very good perf without tensor cores.

 

[jpiniero]

Just based upon that, I'd say $549 and $649 maybe... with the latter having a Fake MSRP until nVidia shows they are serious about supplying the market.

 

[SolidQ]

I'd say $549 and $649 maybe..

i think 450/550, another version 480/600

 

[jpiniero]

i think 450/550, another version 480/600

Based upon those benchmarks, the 9070 should be faster than the 5070 as it is. No need to do the -$100.

 

[uzzi38]

I think it's much more likely that poster on a Chinese forum references AnandTech forums, guesstimate of the 350 mm2 die size is just wrong. It's more likely that the people estimating ~390 mm2 where closer.

The "people" you're referring to in question:

https://twitter.com/x/status/1893178700423655903

(Yes it was only one person).

 

[SolidQ]

the 9070 should be faster than the 5070 as it is. No need to do the -$100.

5070 is basically dead card for 550$, unless you need specific things, like Cuda.
Why no need -100$? Make just no room for 5070 to reduce price.

 

[jpiniero]

5070 is basically dead card for 550$, unless you need specific things, like Cuda.
Why no need -100$? Make just no room for 5070 to reduce price.

Don't want to mess up the opportunity to actually make money this time.

Keep in mind, at 53 billion transisitors, it's only slightly less than N31 is.

 

[coercitiv]

Keep in mind, at 53 billion transisitors, it's only slightly less than N31 is.

Now they're paying TSMC by transistor count?

 

[SolidQ]

I calculate only raster. got 18-19% for nonXT, 37-38% for XT
Compare with this

 

[jpiniero]

Now they're paying TSMC by transistor count?

A large portion of N31/N32 is on a far more expensive node, so yeah.

 

[maddie]

Now they're paying TSMC by transistor count?

Think about when we get FinFlex, it'll be by transistor type.

Everything to advance the agenda.

 

[gaav87]

 

[gaav87]