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

 

[In2Photos]

Now to see what real world prices will be!

 

[SolidQ]

No reason to buy the 9070.

in half year will be difference

 

[poke01]

Interesting.

 

[Saylick]

Here’s a closer look at the RT improvements. Looks like it supports BVH compression and shader execution reordering?

 

[jpiniero]

Now to see what real world prices will be!

That's gonna be interesting. Amazon themselves is selling a 5070 Ti for $999 right now.

 

[poke01]

Here’s a closer look at the RT improvements. Looks like it supports BVH compression and shader execution reordering?
View attachment 118359
View attachment 118360
View attachment 118361
View attachment 118362
View attachment 118363

Can we say AMD is on par with NV now in regards to ray tracing?

I think amd is

 

[poke01]

I have this say AMD launch was more to point, less AI bull and sold me on the product.

More launches like this please. Much much better than RDNA3 presentation

 

[Win2012R2]

Can we say AMD is on par with NV now in regards to ray tracing?

Not until the features that require extra effort (like I believe shader exec reordering) from game makers to speed things up will be used, which in case of CyberPunk 2077 will be never...

 

[poke01]

Not until the features that require extra effort (like I believe shader exec reordering) from game makers to speed things up will be used, which in case of CyberPunk 2077 will be never...

I think the upcoming game AMD showed off looked good.

 

[Josh128]

^^
Yeah that was like 8 years ago, those days are gone, never to return my dude. Also was on the el cheapo GloFo / Samsung 14nm process.

$599/$699 still my prediction.

I happily eat a bit of crow with this incorrect prediction. $549/$599 are not terrible prices for the perf and features claimed. Seems 9070 will be low availability, low volume, the RX6800 / RX7700XT of this gen. Now lets see if literally any cards at all are available at this MSRP in the next few months.

 

[Saylick]

Can we say AMD is on par with NV now in regards to ray tracing?

I think amd is

Not sure if they are on par, but they’ve appeared to catch up in many ways, including the encoder/decoder as well.

 

[Saylick]

I happily eat a bit of crow with this incorrect prediction. $549/$599 are not terrible prices for the perf and features claimed. Seems 9070 will be low availability, low volume, the RX6800 / RX7700XT of this gen. Now lets see if literally any cards at all are available at this MSRP in the next few months.

Hey, don’t disparage the RX6800 by lumping it in with the 7700XT. It was a solid card for the price in hindsight and the price difference between it and the RX 6800XT made more sense.

$549 for the 9070 is literally designed to be unexciting, likely because yields are so good that they’d rather you buy the whole chip anyways.

 

[Josh128]

More launches like this please. Much much better than RDNA3 presentation

Herkleman didnt even believe what he was saying in that launch presentation, lol.

 

[Josh128]

Hey, don’t disparage the RX6800 by lumping it in with the 7700XT. It was a solid card for the price in hindsight and the price difference between it and the RX 6800XT made more sense.

$549 for the 9070 is literally designed to be unexciting, likely because yields are so good that they’d rather you buy the whole chip anyways.

I was speaking about volume / availability. 6800 is a very rare SKU compared to all other RDNA2's. Im thinking 9070 will be similar.

 

[RnR_au]

More launches like this please. Much much better than RDNA3 presentation

idk mate... no leather jackets and no stilettos... Zzzzz

 

[Meteor Late]

Not sure if they are on par, but they’ve appeared to catch up in many ways, including the encoder/decoder as well.

Does RDNA4 media engine support 4:4:4 encode/decode now? otherwise they are still playing catch up.

 

[GTracing]

The 9070 XT seems like a big improvement across the board.

• ~30% better perf/dollar compared to the best value GPU of last generation (7900 GRE). And it's roughly equal to the critically acclaimed B580.
• Improved ray tracing
• Massively better upscaling (though still behind DLSS4)
• Improved perf/watt. They finally caught Nvidia on the same node!

For gaming, it's a no-brainer. Nvidia has to drop the price of the 5070, otherwise reviews are going to tear it apart drop the rating from 5 stars to 4.5 stars.

 

[steen2]

View attachment 118358

Interesting.

DLSS 3/4 CNN/Trans are mixed precision down to fp16 for backwards compatibility. Ada/Blackwell support fp8 with the latter also supporting fp4 with additional quantization issues.

The real question is what model does FSR4 use? CNN, transformer, fp8 weights only...?

[*]Massively better upscaling (though still behind DLSS4)

Do tell.

 

[gdansk]

Good news the die size is indeed 356mm². So they should be competitive on an area basis and very close on perf per watt too.

 

[Meteor Late]

The price is decent, but not aggressive enough for what 357 mm2 really cost.
Still, it is much better compared to how uncompetitive RDNA3 really was, it seems RDNA4 will be much closer in terms of features.

 

[Keller_TT]

Can we say AMD is on par with NV now in regards to ray tracing?

Looks more or less equal in terms of per CU capability *if* independent tests confirm the claims.

The odd Punky ones built specifically around NV architecture and DLSS suite might have a slight NV advantage, but Lackwell is very incremental over Ada even for RT, far from the last 2 RTX gen gains.
Impressive from AMD nevertheless.

 

[coercitiv]

perfect per watt

Typos sometimes have their own wisdom.

 

[SolidQ]

Let's what they cook with RDNA5, that should be another major arch rework

 

[branch_suggestion]

L2 per SE has doubled, which was expected.
L1 is indeed gone, should be a buffer now which is why it is not being shown.

 

[poke01]

From reddit and techpowerup:

The new Radiance 2 Display Engine comes with major hardware updates that reduce GPU idle power draw in multi-monitor setups. The engine also comes with hardware flip-metering support (something NVIDIA also introduced with Blackwell, and which enables Multi-Frame Gen on the RTX 50-series). Flip-metering improves video frame pacing to the GPU and reduces CPU overhead for video playback. There is also a display engine level hardware image sharpening component that drives Radeon Image Sharpening. As for I/O, you get contemporary DisplayPort 2.1a and HDMI 2.1b, we are waiting to confirm the UHBR bit-rates.

Navi 48 comes a dual media engine, each of the two can perform concurrent encoding and decoding, and so the implementation resembles an NVIDIA GPU having two each of NVENC and NVDEC units. The new generation media engine offers a 25% increase in H.264 low-latency encode quality, and an 11% improvement in HEVC encode quality. AV1 encode and decode gets B-frames support, vastly improving bitrates. A B-frame is a frame that lacks image information, but instead motion-vector values, that lets the decoder reconstruct the image component using past and future I-Frames (frames with image data). The media engine posts a 50% generational performance uplift (measured in encoder/decoder frame rates), with reductions in memory overhead.