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

 

[Hitman928]

Yes they can, but most people will angrily say they used fake frames to achieve it, rather than using DLSS 4 that can do no wrong...



I can easily see it - she just say: we have to do MI300 instead and make 12-15x the money you can possibly hope to make with big consumer chip, she isn't not wrong sadly...

AFAIK, AMD wouldn’t have an issue getting the wafers, so I don’t think it’s an either or situation here.

 

[adroc_thurston]

I'll believe it when I see it ...

The roadmap is out there, big dawg.

Why

Because.

The news from there is that they're building steadily and will start sampling this year and build up for HVM by Q4 next year, and see no major hurdles for high yields going into 2027.

That doesn't mean they have the wafer capacity to do anything useful.

 

[gaav87]

Well:
+10% driver
+10% (daily oc) (wonder what the daily oc will be 3.4g ? 3.5g)
Almost 4090 for 1/3 the price damn.

 

[gdansk]

Well:
+10% driver
+10% (daily oc) (wonder what the daily oc will be 3.4g ? 3.5g)
Almost 4090 for 1/3 the price damn.

He's hyper hype-training. Look out everyone get off the tracks.

 

[ToTTenTranz]

Went there right now.

It's running on an engineering sample GPU so I'm assuming it's the 9070/XT.

Frame rate seems to be around 30FPS (there's no frame counter) but for some reason it's set with VSync off nor freesync, so lots of tearing when you pan around.

As for how the demo looks, I really can't say it looks transformative. I guess it's more interesting to know what's happening behind it especially when you turn denoising off and then you definitely notice it.

View attachment 114523

View attachment 114524

Went back there and asked for a rundown. They're actually being pretty transparent about it and I really enjoyed the explanation. Here's what the rep told me:

- It's an internally developed demo using Unreal Engine.

- It's running on hardware being released in the near future, but couldn't disclose if it was on the top end model or not.

- It's doing pathtracing and the demo only has 3 toggles: denoising (which also is doing ML upscaling if I understood correctly), sampling and caching. Each step significantly increases IQ. All of these are using ML.

- Besides the 3 toggles, we can either select free movement or a scripted scene sequence. This is just to show that it's running in real-time.

- Native resolution was 1440p, presentation on 4k.

- Performance was around 35 FPS, but without any ML and doing the 3 stages through shaders they say it would be running at 4 FPS.

- They say there's still room for improvement as they have only recently started serious work on ML optimizations for raytracing on the new GPUs.

 

[PJVol]

wonder what the daily oc will be 3.4g ? 3.5g

If gpu-z screenshot is authentic, gpu gfx fmax is already there. The question is the power draw at these frequences and high gfx activity. Eg my 6800xt eats up to 500W in timespy at 2750mhz.

 

[Z O X]

If gpu-z screenshot is authentic, gpu gfx fmax is already there. The question is the power draw at these frequences and high gfx activity. Eg my 6800xt eats up to 500W in timespy at 2750mhz.

6800XT@2450MHz (1036mV) uses a hair under 330W when allowed to go full power on TS. So this could mean +1GHz on max frequency (at same power), the rest of perf on architecture (hopefully positive)...

 

[PJVol]

So this could mean +1GHz on max frequency (at same power), the rest of perf on architecture (hopefully positive)...

I'm not sure I get your point but anyway, I think you can't extrapolate the power consumption between different archs on a different process nodes

 

[gaav87]

If gpu-z screenshot is authentic, gpu gfx fmax is already there. The question is the power draw at these frequences and high gfx activity. Eg my 6800xt eats up to 500W in timespy at 2750mhz.

Im talking about OC 3060mhz was on the gpu-z but we don't kow what model that was.
nitro+ 3.1ghz boost ? My eats 380 at 2600mhz

6800XT@2450MHz (1036mV) uses a hair under 330W when allowed to go full power on TS. So this could mean +1GHz on max frequency (at same power), the rest of perf on architecture (hopefully positive)...

Well lets resonable oc say +10% power slider so ~360w 3.4Ghz doable i think if base is 3.1G.
Prolly 3.25G at some good UV...

 

[Z O X]

I'm not sure I get your point but anyway, I think you can't extrapolate the power consumption between different archs on a different process nodes

Sorry, the point was that at same power, assuming no arch regressions, it should be cca 40% faster than 6800XT due to frequency alone...

 

[gaav87]

He's hyper hype-training. Look out everyone get off the tracks.

10% off after oc vs uv power limited 4090 would be, a win
hooo hooo motherf.... xD

 

[PJVol]

Im talking about OC 3060mhz was on the gpu-z but we don't kow what model that was.
nitro+ 3.1ghz boost ? My eats 380 at 2600mhz

It's not OC, but rather part of so called "driver reported limits" structure, whose purpose is still unknown to me (the same goes for "gaming clock"). What I'm saying is that Gpu-z reporting 3050 could easily indicate Fboost-max to be some 200-300mhz higher. For example at stock settings gpu-z shows 2250mhz boost clock for the ref. 6800XT, while in fact it's stock Fboost-max is 2577

 

[Bryo4321]

That “kilo buck” price point with a hearty amount of vram is my sweet spot. I would’ve bought RDNA 4 at that price point too but alas.

 

[Z O X]

That would be 40% faster than reference 6800XT...

 

[gaav87]

It's not OC, but rather part of so called "driver reported limits" structure, whose purpose is still unknown to me (the same goes for "gaming clock"). What I'm saying is that Gpu-z reporting 3050 could easily indicate Fboost-max to be some 200-300mhz higher. For example at stock settings gpu-z shows 2250mhz boost clock for the ref. 6800XT, while in fact it's stock Fboost-max is 2577

What u talking about fmax for 6800xt is not 2577mhz but 2800mhz

 

[gaav87]

View attachment 114600

That would be 40% faster than reference 6800XT...

Well its 45% faster than my oced 6800xt if this BO6 benchmark is accurate. And im like top30 in timespy for 6800xt+9800x3d
Its also 50% faster then my stock 6800xt+9800x3d in timespy extreem.
If all thous leeks are accurate its here after OC/UV and that is crazy.
I take this with, a grain of salt i like to meme, a little i put it at 4070tiS from the info i had.

 

[PJVol]

What u talking about fmax for 6800xt is not 2577mhz but 2800mhz

I said "stock"

 

[Heartbreaker]

Digital Foundry presents a sneak peak at FSR4:

Yes, the DF reference is a joke.

Actual DF takes a look:

 

[gaav87]

Hm wonder if RDNA4 also have decoupled clocks (front-end and shaders) this could explain higher TBP and a little more perf.

 

[GodisanAtheist]

 

[gaav87]

Well i think the TSE and SPEEDWAY chinese leaks are fake...
7900xt right there... Scores fall in line with top30-40 7900xt oc

 

[gaav87]

Back to 4070tiS boys hoo hoo crashed

 

[gdansk]

7900xt right there

In many of the pictures it says 7800 XT. Either way it's easy to fake such things so it's good to derail the train.

 

[SolidQ]

Either way it's easy to fake such things so it's good to derail the train.

chinese guy, who leak to us 4080 -5%, was surprised by speedway results were too low, despite in games like ~4080 RT

 

[gdansk]

chinese guy, who leak to us 4080 -5%, was surprised by speedway results were too low, despite in games like ~4080 RT

I won't believe anything until I see it.