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

 

[SolidQ]

Why is it hard to get a card that is a color other than black, gray, or white?

I heard, physics saying black color is better reduce heat, than other colors.
Who's friendly with physic, maybe can give answer for us.
If i'm right maybe this - https://en.wikipedia.org/wiki/Kirchhoff's_law_of_thermal_radiation

 

[igor_kavinski]

If i'm right maybe this - https://en.wikipedia.org/wiki/Kirchhoff's_law_of_thermal_radiation

Don't think so. The Sapphire Pure White RX 7900 GRE doesn't have worse thermal characteristics.

 

[Tuna-Fish]

I heard, physics saying black color is better reduce heat, than other colors.
Who's friendly with physic, maybe can give answer for us.
If i'm right maybe this - https://en.wikipedia.org/wiki/Kirchhoff's_law_of_thermal_radiation

This is true for radiative cooling. It's completely irrelevant when you have fans that are pushing meaningful airflow.

 

[igor_kavinski]

This is true for radiative cooling.

Any practical examples? Does this happen in DIY PCs?

 

[MrTeal]

Any practical examples? Does this happen in DIY PCs?

A perfect blackbody at 320°K (47°C) would have a total emittance of ~600W/m². A GPU isn't a blackbody obviously and even colored black it's emittance will be a lot lower, but let's run with 600W/m². Say a card like the 4090 FE was a perfect black body. At .304m x .137m x .061m, it has a surface area of 0.137m², and if the surface was a uniform 47°C it would radiate about 80W in heat. That's the max.
Googling says black anodized aluminum is 0.86 so you'd get 70W, clear anodized is 0.76 for 58.7W so a difference of 11.3W or so on a 4090 sized card with a surface temp of 47°C. nearly 50° is really hot though, the numbers would drop a bit at a more reasonable temp. Realistically as a non-expert I don't think it makes a huge difference, though color could play more of a role in how well they perform in a fan stop mode.

 

[igor_kavinski]

So is that why the FE editions have pitch black fins?

By the way, does Noctua not know this? How come their passive CPU heatsink NH-P1 isn't black?

 

[SolidQ]

Noctua not know this?

isn't they going to Black color?

 

[igor_kavinski]

isn't they going to Black color?

They have Chromax Black models but their lone passive heatsink isn't black which I find odd if black color would help it radiate heat better in the absence of a fan.

 

[Timorous]

So if I had to guess a stack based on the rumours floating around this is my stab at it.

Die	CUs	Vram	Approx Perf (scaled to 7600XT at 1440p)	MSRP (Aggressive MSRP)	die area
N48	64	16GB	220% (7900XT)	$600 ($500)	220mm
N48	56	16GB	170% (7800XT)	$450 ($400)	220mm
N44	32	16GB	130% (7700XT -10%)	$350 ($300)	110mm
N44	28	12GB	100%	$250 ($230)	110mm

Not much movement on perf/$ imo because why would they. I guess if AMD wanted to be aggressive they could go $500 for top N48 and scale everything off of that to offer similar perf/$ all down the stack. With such a small N44 die there may not be a cut part but if there is I think the bus will get cut to 96bit and AMD will slap 12GB of vram on it, given the 8GB issue I think that config for a cut part would be far better than a 128bit bus + 8GB of vram.

Given NV hit double the density of N33 I don't see why AMD can't fit a 32CU/128bit/8pcie4 part into 110mm and double that in 220mm.

As with RDNA 3 let's see how wrong I am come launch day.

 

[Tuna-Fish]

Any practical examples? Does this happen in DIY PCs?

Not really. Radiative cooling gets really important as things get much hotter, because the rate it happens is proportional to T⁴. Near NTP it usually does not have much of an impact, but numbers go up quickly in quartic equations.

 

[Mopetar]

The color of your GPU isn't going to matter enough. The reason they do t use other colors is that the plastic can discolor over time and it winds up looking awful. With black it's considerably less noticeable.

I also don't think there's a need for standardized shroud design. There's no such thing as a standard card size and the layout of the components on the card makes it further impractical. Companies should instead just give specifications that make it easy for end users to 3D print their own designs or adapt some existing design to that specific card.

 

[Abwx]

if the surface was a uniform 47°C it would radiate about 80W in heat.

That s right but it will also absorb the radiation that come from the heated matter that is around, be it the air or the PC case and other elements, if those elements are at lower temp then they will absorb more energy from the card than what they send it in return, hence the importance to keep the case interior at significantly lower temp than the card to benefit from some radiative dissipation.

 

[beginner99]

If that is the rough performance then I think $600 for the top N48 part

That would not move the bar much at all in terms of performance/dollar.

 

[Timorous]

That would not move the bar much at all in terms of performance/dollar.

I don't think there will be much movement on that front. A small step and a beefing up of VRAM on the low end but nothing groundbreaking. I don't see another 4870 moment because as great as that was for the consumer it just did not work and AMD would much rather eat margin which if they have the performance to charge $600 for an N23 ish sized GPU they absolutely will.

 

[Aapje]

As long as AMD is just reacting to what Nvidia does and what the market will bear, talking about performance/dollar seems pretty pointless unless you include Nvidia in the conversation or what consumers will accept.

The question for me is whether Nvidia is going to act like they did with the 10-series and intended to do with the 30-series, to give a pretty nice uplift after a series that didn't sell so well.

It's hard for me to judge how RDNA4 will perform, especially since they cancelled 3 of the 4 chips they wanted to make and then added a new design (N48 is obviously new, if you look at the code name). Is this going to be a mediocre design to just hold them over until RDNA5, where they have little room to be aggressive with the price, or did they fix the major issues with RDNA3 and improve some other things, to be quite good? Hard to say.

 

[moinmoin]

As long as AMD is just reacting to what Nvidia does and what the market will bear

If you don't offer the top product in a given market the only thing you can do is react in relation to the existing market circumstances to carve out a niche. That's why the goal always is to bring a highly competitive top product to the market even if most sales are elsewhere. With GPUs both AMD and Intel failed at that lately.

 

[blckgrffn]

As long as AMD is just reacting to what Nvidia does and what the market will bear, talking about performance/dollar seems pretty pointless unless you include Nvidia in the conversation or what consumers will accept.

The question for me is whether Nvidia is going to act like they did with the 10-series and intended to do with the 30-series, to give a pretty nice uplift after a series that didn't sell so well.

It's hard for me to judge how RDNA4 will perform, especially since they cancelled 3 of the 4 chips they wanted to make and then added a new design (N48 is obviously new, if you look at the code name). Is this going to be a mediocre design to just hold them over until RDNA5, where they have little room to be aggressive with the price, or did they fix the major issues with RDNA3 and improve some other things, to be quite good? Hard to say.

I am wondering if this is the next Polaris vs being a flash in the pan. Optimized die sizes, better than or competitive with current gen consoles and the full DX 12U feature set, this could really be the “bottom” end of the AMD lineup for the next 3 years or so. We keep saying how it’s a stopgap, but maybe it’s just going to be competent.

Given AMD is like Nvidia and needs to chase the data center compute monster, it feels like a possibility to me.

 

[MrTeal]

That s right but it will also absorb the radiation that come from the heated matter that is around, be it the air or the PC case and other elements, if those elements are at lower temp then they will absorb more energy from the card than what they send it in return, hence the importance to keep the case interior at significantly lower temp than the card to benefit from some radiative dissipation.

Yeah, and something with high emissivity will absorb more as well. The GPU surface will net dissipate heat through radiation since it's going to be hotter than the rest of the case, but it's going to be pretty minimal in the grand scheme of things.

 

[GodisanAtheist]

I am wondering if this is the next Polaris vs being a flash in the pan. Optimized die sizes, better than or competitive with current gen consoles and the full DX 12U feature set, this could really be the “bottom” end of the AMD lineup for the next 3 years or so. We keep saying how it’s a stopgap, but maybe it’s just going to be competent.

Given AMD is like Nvidia and needs to chase the data center compute monster, it feels like a possibility to me.

-Yeah, AMD isn't putting work into N44/48 without the expectation that they'll be rebranded into the bottom tier of the RDNA5 stack.

 

[Aapje]

Neither AMD, Nvidia or Intel relegates their entire product stack to the bottom tier of the next gen, so...

The goal is to be able to make something better for the same amount of dollars, so they can get people to upgrade, while maintaining margins.

 

[Aapje]

If you don't offer the top product in a given market the only thing you can do is react in relation to the existing market circumstances to carve out a niche.

Most people don't buy the 4090, so that's not how it works.

Nvidia does have a premium brand, which makes people willing to spend more, but Nvidia also takes advantage of that to price their products really high. So AMD does have room to make a very price-competitive product if they make a very efficient design and accept more modest margins.

A big mistake that AMD makes is that they assume that Nvidia prices their products at an attractive level, so AMD only tries to make their products attractive relative to Nvidia. But if Nvidia asks too much for the market to bear, as they mostly did with Ada, then this means that AMD does the same. In fact, it is worse, because AMD even overestimates how attractive they are compared to Nvidia.

The end result is that AMD gets bad reviews and constantly runs into getting bad sales after launch, requiring them to cut prices very quickly after release. This is just a very bad strategy that costs them sales, positive attention and also exposes them as incompetents, who cannot read the market.

 

[SolidQ]

The end result is that AMD gets bad reviews and constantly runs into getting bad sales after launch, requiring them to cut prices very quickly after release

Despite this, AMD still got additional 7% in 2023

 

[jpiniero]

and also exposes them as incompetents, who cannot read the market.

Don't think it's a "cannot read the market" thing, it's that the margins suck of the product they do sell...

 

[Abwx]

Yeah, and something with high emissivity will absorb more as well. The GPU surface will net dissipate heat through radiation since it's going to be hotter than the rest of the case, but it's going to be pretty minimal in the grand scheme of things.

There s really not much to grab, although radiation increase with temp at a bi quadratic rate the card would had melted before reaching a radiative dissipation power barely comparable to the convective cooling, the surface at play being just too negligible.

 

[igor_kavinski]

Despite this, AMD still got additional 7% in 2023

SInce Jensen can do no wrong, I bet the AMD gain is a direct result of too much demand for Geforce RTX cards leading to cards going out of stock in many places and buyers settling for the inferior AMD cards instead of returning home empty handed and crying themselves to sleep.

/s