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

 

[gaav87]

No way N44 can reach that level of performance.

All the watts tweeted
N33<<N44<N32
This equals ~7700xt
I think it will be close to 7700xt but NOT between 7700xt and 7800xt.

 

[Mahboi]

Ironically Twitch actually made a major contribution to AV1 development in the form of S-Frames, which makes their slow transition all the stranger.

Certainly at this point AV1 hw decoding access is fairly widespread, and becoming ever more so since Qualcomm and Apple finally took the stick outta their posteriors and implemented it.

Even where it isn't there on older generations there is plenty of CPU power to decode it efficiently using dav1d, which at this point is quite possibly the most well optimised video decoder library in history.

I don't think that a Vulkan GPU implementation would be a significant improvement for perf/watt over CPU given how complex AV1 is as a codec.

Not really. Tons of people still run basic Turing/Pascal era GPUs without it. AV1 decode is only a 2020 onwards thing.
Twitch is forced to allow transcoding from AV1 to h264. Especially in countries where changing hardware often means changing from a dead GPU/PC to a used one.
Now the real question is in a world where MA35D exists, where the hell is it difficult for Twitch to do that? It's an expensive card I guess if you take in bulk, but we're talking about needing a few thousands of them to at least get a basic server farm and test bed. Getting ONE batch of a few hundreds in a server room should be peanuts for a company the size of Twitch, and getting that testbed going with a few big streamers to see how far it can go is nothing. It's freaking 35W of usage per card, and each can transcode 32 h264 1080p streams. We're not talking about a farm of MI300Xs here. Even 50000 viewers boils down to about 1500 cards. And that's not 50000 viewers WATCHING, it's 50000 needing h264 transcode, which should be about 30% of all viewers give or take.

I don't know from where you think dav1d is "efficient enough that there is plenty of CPU power". Last I looked AV1 SW decode ate something around 8 times more than h264. And even Zen 3 CPUs take around 10% of their power to decode h264. I don't see 60% Zen 3 usage as "enough" and for people running old 8000ks or for the one crazy still running Bulldozer, definitely not.

 

[soresu]

I don't know from where you think dav1d is "efficient enough that there is plenty of CPU power"

I was talking about mobile devices.

Anything with 3-4 cores of A76 at a reasonable clock speed can tank 1080p60 10 bpc content with ease, and possibly even 4K30 10 bpc too (4K60 might be pushing it without more cores).

Anything >X1 or comparable Axx/Mx based (NEON wise) from Apple can handle 4K60 in software with dav1d.

Like I said, it's a very efficient decoder.

I've been following dav1d development from pretty much the day it was first announced.

It's been a while since I saw a comprehensive CPU comparison with dav1d so if I'm wrong do please enlighten me.

 

[marees]

Will AMD introduce RDNA 4 (in a month's time?) when RDNA 2 sales are still 1/3rd of total discrete sales ???

https://twitter.com/x/status/1830598117445521719

 

[maddie]

Will AMD introduce RDNA 4 (in a month's time?) when RDNA 2 sales are still 1/3rd of total discrete sales ???

Market % really does not matter. Months of inventory is the key data point.

 

[ToTTenTranz]

Will AMD introduce RDNA 4 (in a month's time?) when RDNA 2 sales are still 1/3rd of total discrete sales ???

https://twitter.com/x/status/1830598117445521719

Does it matter? 1/3rd of very few is almost nothing.

Without new products bringing higher performance-per-dollar, AMD's sales of Radeon GPUs will continue to go in a downward spiral to the point of making the business unsustainable.
In Q2 2024, AMD's gaming division revenue went down 59% YoY, and the quarter before it was 48%. It doesn't matter if they're still selling 4 year-old GPUs that were produced during the 2nd crypto-craze with margins that will (hopefully) never happen again in the consumer market.

 

[jpiniero]

Does it matter? 1/3rd of very few is almost nothing.

Without new products bringing higher performance-per-dollar, AMD's sales of Radeon GPUs will continue to go in a downward spiral to the point of making the business unsustainable.
In Q2 2024, AMD's gaming division revenue went down 59% YoY, and the quarter before it was 48%. It doesn't matter if they're still selling 4 year-old GPUs that were produced during the 2nd crypto-craze with margins that will (hopefully) never happen again in the consumer market.

Like nearly all of the Gaming revenue is from the Consoles. AMD even said they made more on Radeon YoY but that doesn't tell you that much.

 

[marees]

Does it matter? 1/3rd of very few is almost nothing.

Without new products bringing higher performance-per-dollar, AMD's sales of Radeon GPUs will continue to go in a downward spiral to the point of making the business unsustainable.
In Q2 2024, AMD's gaming division revenue went down 59% YoY, and the quarter before it was 48%. It doesn't matter if they're still selling 4 year-old GPUs that were produced during the 2nd crypto-craze with margins that will (hopefully) never happen again in the consumer market.

Looking at the top selling cards, I think a navi 48 xtx will sell like hot chips for $600, if it can trade blows with 4070 ti super (or even the 4070 ti) in RT

 

[ToTTenTranz]

Looking at the top selling cards, I think a navi 48 xtx will sell like hot chips for $600, if it can trade blows with 4070 ti super (or even the 4070 ti) in RT

Yes, the consumer hasn't had any significant price-performance upgrade since 2019. We're probably paying as much per-FPS today than we were back then.

A card that brought significant gains on that front would be welcomed at least by all the major youtube channels that have been complaining about it incessantly.

 

[Mahboi]

Like I said, it's a very efficient decoder.

I've been following dav1d development from pretty much the day it was first announced.

It's been a while since I saw a comprehensive CPU comparison with dav1d so if I'm wrong do please enlighten me.

I can't.
I can't find a single good review that gives proper usage. Maybe I'm just the one who assumed wrong and it's less costly than it once was.
But I am certain that the transcoding issue is one of the problems at Twitch.
However I also am personally convinced that their code is a hellhole of years of spaghettification, and since their entire stack is on h264 since forever, they are trying to roll out a new codec through over a decade of spaghettifying optimisations.

 

[soresu]

I can't find a single good review that gives proper usage. Maybe I'm just the one who assumed wrong and it's less costly than it once was.

Probably Phoronix would be the only consistent site for dav1d perf in hardware reviews.

I don't ever save them tho, so once it comes and goes it's hit and miss if it sticks in my head.

 

[Mahboi]

Probably Phoronix would be the only consistent site for dav1d perf in hardware reviews.

I don't ever save them tho, so once it comes and goes it's hit and miss if it sticks in my head.

They're not valid, they don't give any core usage. I don't really care about the max framerate, I just am curious about CPU usage.
I still doubt than say Zen 3 can have AV1 decode without significant weight. Like my 5600x would probably take something like 30-50% usage just to watch a 1440p video. That's a CPU that hardly ever consumes more than 20% for any web tasks normally.

 

[soresu]

They're not valid, they don't give any core usage. I don't really care about the max framerate, I just am curious about CPU usage.
I still doubt than say Zen 3 can have AV1 decode without significant weight. Like my 5600x would probably take something like 30-50% usage just to watch a 1440p video. That's a CPU that hardly ever consumes more than 20% for any web tasks normally.

Hol' up, just downloading a YT video to test my 3950X then you can infer from that.

 

[soresu]

This may not be a particularly high bitrate example which would affect the utilisation, but it is 1440p25:


MPC-HC at the bottom left is the player.

Another screenshot with full core utilisation graph after I shutdown Firefox:

 

[soresu]

According to amd.com specs 5600X has 300 mhz lower boost clock than 3950X, but also 300 mhz higher base clock, though Zen3 also has higher IPC over Zen2.

 

[soresu]

Sorry for non CPU talk mods 😅

 

[soresu]

Despite AV1 being a more complex codec it is actually possible to be significantly more efficient when decoding it vs the older VP9 because more possible parallelisation routes were designed into AV1 before bitstream freeze.

 

[Mopetar]

Now they just need to release it and all will be well. Planning to get a top tier RDNA4 at launch if raster is really at 7900XT levels.

I'm waiting to see prices, but if it's reasonable I might pick one up. The only issue is that AMD knows what reasonable is and how to tack an extra $100 on top of that because NVidia is even worse in that regard and will charge $150-$200 over reasonable.

I wouldn't mind seeing a pro refresh for the consoles either. My PS3 died recently and I'm without a BluRay player.

 

[jpiniero]

I wouldn't mind seeing a pro refresh for the consoles either. My PS3 died recently and I'm without a BluRay player.

I got bad news about that... the PS5 Pro is most likely $599 (or more even) and that's without the drive (which is $80)

You can pretty much take it to the bank that people are going to complain about RDNA 4's prices.

 

[marees]

Wccftech made an article out of the tweet below

https://twitter.com/x/status/1829466208145494016

Million dollar question is:
How much faster in RT

 

[marees]

Wccftech made an article out of the tweet below

No wccftech article yet on Mahboiiiii's tweet

https://twitter.com/x/status/1829880954812580196

 

[SolidQ]

How much faster in RT

That RDNA4 RT still mystery for us, we don't know how impact those changes

 

[Ghostsonplanets]

I wonder if AMD will really do RX 8600 and RX 8500. If they price it correctly, they'll have an advantage over Nvidia in the low-end market.

Navi 44 is a 130mm² die. AMD needs to use this advantage of an extremely small die to their fullest and be aggressive to win marketshare.

either less or even no Infinity Cache

AMD x86 Consoles don't have IC. They rely on hUMA with a single pool of memory.

No way N44 can reach that level of performance.

I agree. Folks forget that Navi 44 is still a 16 WGPs design and Navi 33 already clocked quite high. I believe N44 Radeon 8600 XT performance will be around Radeon 6750 XT or RTX 3070/Ti.

All the watts tweeted
N33<<N44<N32
This equals ~7700xt
I think it will be close to 7700xt but NOT between 7700xt and 7800xt.

Something like RTX 3070/Ti performance, yes.

 

[marees]

Looking at the top selling cards, I think a navi 48 xtx will sell like hot chips for $600, if it can trade blows with 4070 ti super (or even the 4070 ti) in RT

View attachment 106696

Looking at NewEgg there is $150 gap between the 7900gre & 7900xt

That is a significant hole in the segment of the market which is selling big now

If only AMD had a $600 card with RT of 4070 ti or 4070 ti super that they could launch 🤔

 

[Mahboi]

This may not be a particularly high bitrate example which would affect the utilisation, but it is 1440p25:

MPC-HC at the bottom left is the player.

wow, impressive, and I was wrong.