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

 

[moinmoin]

Lisa Su says AMD is on track to a 100x power efficiency improvement by 2027 — CEO outlines AMD’s advances during keynote at imec’s ITF World 2024

Less energy, more performance.

www.tomshardware.com

Absolute banger, will probably elaborate beyond this at Computex.

It's an update of last year's talk at ISSCC. It's a good look into how AMD intends to proceed with technology and a promising focused application of "AI" tech.

Question - By 2030, we will be buying massive NPUs with a CPU and GPU attached to it.

Today, NPUs such as Apple's neural engine takes less space than the CPU or GPU in a SoC. By 2030, I predict that we won't be buying "CPUs". We will all be buying NPUs with a CPU and a GPU attached to it. NPUs will become the new CPUs. More applications will start to make massive use of AI...

forums.anandtech.com

 

[branch_suggestion]

It's an update of last year's talk at ISSCC. It's a good look into how AMD intends to proceed with technology and a promising focused application of "AI" tech.

Question - By 2030, we will be buying massive NPUs with a CPU and GPU attached to it.

Today, NPUs such as Apple's neural engine takes less space than the CPU or GPU in a SoC. By 2030, I predict that we won't be buying "CPUs". We will all be buying NPUs with a CPU and a GPU attached to it. NPUs will become the new CPUs. More applications will start to make massive use of AI...

forums.anandtech.com

Over a year ago, 2027 aligns with MI500 and well, all the hints are there.

 

[SolidQ]

3.1 was announced at GDC already, 2 months ago.

I mean release date.
Interesting on Computex is zero chance talking about RDNA4?

 

[dhruvdh]

I think it's almost certain they'll talk about it. Because AI. It supports new data types for AI after all.

AI.

 

[SolidQ]

I think it's almost certain they'll talk about it.

I think so. Zero reason add alot lines, if not talking about it.

AMD Linux devs jamming nearly 24,000 lines of RDNA 4 supporting code into its mainstream driver suggests next-gen launch may be close at hand

Mesa mega merge means a magnificent moment may manifest…err…soon.

www.pcgamer.com

 

[Mahboi]

I mean release date.
Interesting on Computex is zero chance talking about RDNA4?

Zero.
They already have to introduce Zen 5, RDNA 3.5 (trust us my friends it werks now), Strix Point, the big APU plans, more AI AIA AAIAIIIAIAIAA IA AI, and probably 2 or 3 extra things.
The more messages, the less they go through.
RDNA 4 would just be a dead weight.

 

[soresu]

They already have to introduce Zen 5, RDNA 3.5 (trust us my friends it werks now)

RDNA 3.5 won't be a focus of that but rather the XDNA NPU or whatever it is AMD calls it.

RDNA 4 would just be a dead weight

If they are introducing any RDNA4 cards this year it's not dead weight to at least mention it.

Also I think the RDNA3.5 changes are more likely to be Sony collaborations for PS5 Pro than anything specifically intended for PC improvement if what Kepler says is correct about RDNA4 RT looking very different from RDNA3.

 

[Tuna-Fish]

Zero.
They already have to introduce Zen 5, RDNA 3.5 (trust us my friends it werks now), Strix Point, the big APU plans, more AI AIA AAIAIIIAIAIAA IA AI, and probably 2 or 3 extra things.
The more messages, the less they go through.
RDNA 4 would just be a dead weight.

AMD has a habit of having "one last thing" in all their presentations, where at the very end they tease a different product. I think it's very likely that RDNA4 will be that for the Zen5 presentation.

I doubt they fit it into the actual presentation, for the simple reason that whether it's a hit or a flop, Zen5 will drown it in news and discussion. Even if they are releasing RDNA4 soon, a separate event will get it a lot more attention, even if it's a lot lower profile one than a computex keynote.

 

[blckgrffn]

RDNA 3.5 won't be a focus of that but rather the XDNA NPU or whatever it is AMD calls it.

If they are introducing any RDNA4 cards this year it's not dead weight to at least mention it.

Also I think the RDNA3.5 changes are more likely to be Sony collaborations for PS5 Pro than anything specifically intended for PC improvement if what Kepler says is correct about RDNA4 RT looking very different from RDNA3.

Meh, PS5 Pro has the RT hardware from RDNA4 so I don't think its really that influenced by RDNA3.5.

Honestly that feels like something they would just want to skate right by at this point. Technically it would be interesting to know more and to know what RDNA3 maybe should have looked like but ¯\_(ツ)_/¯ Basically water under the bridge now.

 

[GodisanAtheist]

AMD has a habit of having "one last thing" in all their presentations, where at the very end they tease a different product. I think it's very likely that RDNA4 will be that for the Zen5 presentation.

I doubt they fit it into the actual presentation, for the simple reason that whether it's a hit or a flop, Zen5 will drown it in news and discussion. Even if they are releasing RDNA4 soon, a separate event will get it a lot more attention, even if it's a lot lower profile one than a computex keynote.

- One Last Things (TM Apple Corporation) are always what the whole presentation was about, not the afterthought, which RDNA4 would be to Zen 5.

It would make more sense to add Zen 5 as the One Last Thing to an RDNA4 presentation, not the other way around.

Granted AMD marketing sucks [ALL THE REDACTEDS ARE BELONG TO US] and might screw this up in the opposite way.

 

[Tuna-Fish]

- One Last Things (TM Apple Corporation) are always what the whole presentation was about, not the afterthought, which RDNA4 would be to Zen 5.

It would make more sense to add Zen 5 as the One Last Thing to an RDNA4 presentation, not the other way around.

Granted AMD marketing sucks [ALL THE REDACTEDS ARE BELONG TO US] and might screw this up in the opposite way.

I specifically meant it like AMD has been using it, not Apple. And yes, it's quite different. This has in the past been things like a 15-second demo of an upcoming Zen CPU running at a high clock speed after a Radeon presentation, and a demo of upcoming Radeon cards running a new game at some indeterminate frame rate after a Zen presentation.

It's something that's not related to the main presentation, and is just added to the end as a teaser on future products.

 

[gdansk]

AMD adds afterthoughts, Apple used to save the best for last. It's a different approach.

 

[kekframeX]

Future CDNA roadmap to be announced. Soon™

 

[marees]

Technically name and price can change right before announcement, but AFAIK it will look something like this:

N48:
8700 XTX 16GB
8700 XT 16GB
8600 XT 12GB
N44:
8500 XT 8GB

Thinking about this, I am skeptical of a 12 GB RDNA 4 card.

Reason is

1. Yields are likely to be good thus there will be fewer of these cut down chips
2. Easier to mark-up a 16gb card closer to $400. AMD Customers unlikely to pay $400 for 12gb card
3. 6800 16gb is selling for $355 now. It will be a huge letdown if AMD sells a 12gb less powerful than this for $350

 

[leoneazzurro]

It is done to cover more market ranges, and thus increase total sales. This helps anyway with the amortisation of fixed costs, because BOM is a small part of the total production costs.

 

[igor_kavinski]

Radeon needs a serious asskicking and even though moar CUs will do half the job, they don't know how to do the other half.

There are a lot of Geforce users using Ryzen 7000 vanilla and X3D CPUs and a lot might make the switch to Zen 5.

All AMD needed to do was include a decent iGPU (8700G level) on the V-cache CPUs. The V-cache would've offset the latency disadvantage of chiplets and these Geforce users would've been introduced to a cheap Radeon that they could experiment with and appreciate in light games. This sort of introduction is absolutely needed to get people onboard when they think your product is worse than getting the plague. I frankly DISLIKE Geforce drivers control panel and the Geforce Experience UI crap after seeing how much better Adrenaline control panel looks aesthetically. It's so easy on the eyes and easier to navigate.

 

[Mahboi]

Thinking about this, I am skeptical of a 12 GB RDNA 4 card.

Reason is

1. Yields are likely to be good thus there will be fewer of these cut down chips
2. Easier to mark-up a 16gb card closer to $400. AMD Customers unlikely to pay $400 for 12gb card
3. 6800 16gb is selling for $355 now. It will be a huge letdown if AMD sells a 12gb less powerful than this for $350

Agreed, but that's exactly what they did with Navi 32, and it worked out great.
Tons of 7800 xts sold, few 7700 xts. Just price them badly and sell 90% of the stock as 7800 XTs.
Although MLID did have that statement about "they aimed for 50/50 sales" but I don't know how much of that statement is from his imagination.

 

[igor_kavinski]

Naah. Radeon is not the best performing product and competitor has the superior product. That's the problem for AMD. Some just want the best and the best is RTX 4090. It's just as simple as that.

Yeah except when their "best" product malfunctions (like some repair shops getting 200 4090s to repair per month). But people who want bragging rights appear to put themselves through any suffering.

 

[jpiniero]

3. 6800 16gb is selling for $355 now. It will be a huge letdown if AMD sells a 12gb less powerful than this for $350

See, this is what I am talking about when I talk about the Fire Sales hurting future products.

 

[branch_suggestion]

See, this is what I am talking about when I talk about the Fire Sales hurting future products.

Fire sales are quickly forgotten once inventory is gone.

 

[Mahboi]

The problem is that apparently, fire sales do not empty inventory anymore...
How long has AMD been trying to get rid of RDNA 2 cards now??? A whole year?

 

[soresu]

How long has AMD been trying to get rid of RDNA 2 cards now??? A whole year?

That has a lot to do with the slow rollout of the wider family of RDNA3 SKUs.

Because of that many of the lower end RDNA2 GPUs were in production for longer vs the higher end.

 

[igor_kavinski]

How long has AMD been trying to get rid of RDNA 2 cards now???

Would be interesting to see data from shopping sites showing how many customers searched directly for 4060, those who clicked RDNA2 but still got 4060 and those who went with RDNA2. The revelations from that data would be mind boggling to say the least.

 

[SolidQ]

How long has AMD been trying to get rid of RDNA 2 cards now??? A whole year?

Mindfactory. Seems a lot 6800 still..

If 8600XT with 12GB for 300-330 with same perf as 6800. Would be nice.

P.S Gonna be interesting how people react, if AMD gonna sell 8700 non XT for 400$ with much better RT than 7800XT

 

[Thunder 57]

Would be interesting to see data from shopping sites showing how many customers searched directly for 3050 Tye, those who clicked RDNA2 but still got 3050/3050 Tie and those who went with RDNA2. The revelations from that data would be mind boggling to say the least.

FTFY.