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

 

[branch_suggestion]

https://twitter.com/x/status/1790552041577644461

I wonder which RT patent this aligns with.
Just a shame that DXR remains a trainwreck.

 

[Mahboi]

Yeah, it's called mindshare. Nvidia wins by a landslide in that regard. Also, this AI boom has catapulted Nvidia to the top of everyone's mind.

The problem is that the mindshare is highly controllable IF your marketing/PR is up to snuff. I remember that case where some FPS tourney got ruined because someone's 4080 crashed due to an NV driver midgame.
If I had been at AMD marketing then, I'd have shaken heaven and earth internally to push my product. Instead AMD slept through a perfect PR showcase.

4090s burn by the thousands? AMD sleeps.
XTXs had insufficient vapor chamber coolant? Nv shills were all fed the narrative and I actually had this unbelievably eerie Reddit encounter years ago where some Plebbitor asked "how could people even consider buying a Radeon 7000 due to all ((we)) know about it?!"

At first I thought it was just some other NV shill trying to spread misinfo, but no, the guy actually readily accepted the facts when I told him XTXs had insufficient coolant and were going to be fixed, and he was completely honest.
DerBauer makes a video about serious cooling problems on XTXs? NV is all over that.
THE NEXT DAY, DerBauer has opened up the XTX and shown that the vapor chamber is perfectly fine and that it's likely to just be lacking coolant? AMD is gone to pick mushrooms and can't get anyone up to speed. And even just basic communication takes weeks.

It's always like this, like NV has WW1 artillery and AMD just has infantry. Even though between burning HPWHRs, awful pricing and low VRAM, they had ample opportunities.

 

[Mahboi]

Like this posts
(BeyondFPS)

I have literally blocked this guy on Twitter for being GraphicallyChallenged/Prakhar tier.

 

[H433x0n]

Then, directly or indirectly, reviews matter.

Can you show me the positive reviews for Nvidia’s Ada lineup? The only SKU that you could argue had a positive reception was the 4090 and even then there was always an obligatory 5 minute virtue signal about how $1600 is a lot of money for a graphics card.

Arguably, there hasn’t been a glowing review of a single graphics card in like 5+ years, this includes the Radeon cards.

 

[adroc_thurston]

Can you show me the positive reviews for Nvidia’s Ada lineup

Anything shill foundry did?
How is this even a question lmao.

 

[Mahboi]

Rather than positive reviews, it's more about how reviews don't really change public opinion.
They just help evaluate a product to people that actually care to do so. Most buyers set a budget, then either do not care to evaluate the products within that budget and go for "the obvious" (Nvidia), either evaluate briefly and typically ask a nerd friend.

If you want to impress the general public, you need to destroy that position where Nvidia is the default choice and AMD the weird choice. Which again, AMD simply has a big F in terms of effort there.

 

[Hans Gruber]

I hope everybody here realizes that youtubers talking tech are simply reading web stories and blogs that anybody can find. That is not breaking news or finding real leaks about upcoming tech products. They call it content creation.

 

[NTMBK]

From last-minute change to Nvidia GPU in OG Xbox announced on stage with AMD folks sitting on the front row, to last-minute cancellation of whatever mobile/handheld Van Gogh was built for, to Microsoft deciding not to include AMD's hardware tessellation in DX10 (after supporting and older version of it on X360 for years).. perhaps AMD is better off not depending on Microsoft for absolutely anything.

I think you're misremembering- it was an Intel CPU that got the last minute switch (ditching an Athlon), not the Nvidia chipset. And ATi was an independent company back then anyway.

 

[Mahboi]

I hope everybody here realizes that youtubers talking tech are simply reading web stories and blogs that anybody can find. That is not breaking news or finding real leaks about upcoming tech products. They call it content creation.

And yet people look at HWUB and think "wow, high quality reporting".
Steve is a decent man doing a very decent job trying to put a balanced opinion up. Rarely insightful and sometimes overly insisting on some points but broadly a good addition to the tech sphere.
Next to him you have Tim the Giant NV Shill, who buys 4090s for himself and basically laughs at anything that's not a 4090 tier of performance for not being a 4090. An utter shill who buys the most expensive thing and then handles everything else based on his personal choice.

You can have very good and very bad. Even on the same channel.
And despite how much flak they get, LTT is far from being the worst. I even think that LTT gets this much popularity because they stay broadly positive about tech and engage people who don't care about numbers and just want to know what's good to buy.
Hopefully LTT does a load of shilling for RDNA 4 when it comes out.

 

[H433x0n]

Anything shill foundry did?
How is this even a question lmao.

I didn’t really think of them.

My main point isn’t an Nvidia vs AMD thing but that GPU reviews and their commentary as a whole are becoming useless since they’ve got to act negatively about every new GPU release from audience capture. For example RDNA3, the only GPU that didn’t get bad reviews was the 7800XT and even that was sort of lukewarm.

If you based your buying decisions based off of reviewers you wouldn’t buy anything, you’d have to get a time machine and go back to pascal era.

 

[Mahboi]

If you based your buying decisions based off of reviewers you wouldn’t buy anything, you’d have to get a time machine and go back to pascal era.

And the market is heavily down. Mostly due to pricing.

 

[adroc_thurston]

I didn’t really think of them.

They're really important!

For example RDNA3, the only GPU that didn’t get bad reviews was the 7800XT and even that was sort of lukewarm.

Yes but the more obnoxious shills always tow the NV line.
That's the issue even if the product is meh (or bad).

 

[Aapje]

Next to him you have Tim the Giant NV Shill, who buys 4090s for himself and basically laughs at anything that's not a 4090 tier of performance for not being a 4090. An utter shill who buys the most expensive thing and then handles everything else based on his personal choice.

This is just nonsense, since Tim goes against the Nvidia narrative plenty. Like criticizing DLSS 3 quality, criticizing how it is marketed, criticizing RT, etc.

The channel got blacklisted by Nvidia for not towing the Nvidia line!

If this is really what you perceive as Nvidia shills, then I think that the bias is very much on your end.

 

[Aapje]

My main point isn’t an Nvidia vs AMD thing but that GPU reviews and their commentary as a whole are becoming useless since they’ve got to act negatively about every new GPU release from audience capture. For example RDNA3, the only GPU that didn’t get bad reviews was the 7800XT and even that was sort of lukewarm.

If you based your buying decisions based off of reviewers you wouldn’t buy anything, you’d have to get a time machine and go back to pascal era.

Are the reviews useless, or are the products mostly pretty poor compared to previous generations?

 

[Mahboi]

Did I uh, miss something? ROCm under Windows? Google search results seem inconclusive.

edit: huh, seems like ROCm actually works under Win11 now. That's news to me.

(from Z5 thread)

I assume that RDNA 4 will be the occasion for a version bump and we're going to ROCm 7?

 

[DrMrLordX]

I assume that RDNA 4 will be the occasion for a version bump and we're going to ROCm 7?

Depends on whether AMD adds the RDNA 4 consumer cards to the supported list. The current consumer stack supports ROCm in Win11 all the way down to the 7600. So maybe the top-end RDNA4 cards will be supported?

 

[Mahboi]

Honestly if Radeon is dumb enough to cease supporting consumer cards at any point between now and Hell freezing over, they deserve to die as a brand.

I'm more wondering whether they'll stick to ROCm 6 or not. After all, MI300 is CDNA 3 and yet it did enter into ROCm 6.

 

[Krteq]

🤨

For ROCm 6.1 - MI300A (gfx942) is supported on Ubuntu 22.04.4, RHEL 9.3 & 8.9 and SLES 15 SP5. MI300X (gfx942) is only supported on Ubuntu 22.04.4.

For ROCm 6.0 - MI300A (gfx942) is supported on Ubuntu 22.04.3, RHEL 8.9 and SLES 15 SP5. MI300X (gfx942) is only supported on Ubuntu 22.04.3.

Compatibility matrix — ROCm Documentation

ROCm compatibility matrix

rocm.docs.amd.com

 

[Mahboi]

🤨

Compatibility matrix — ROCm Documentation

ROCm compatibility matrix

rocm.docs.amd.com

See this
Search for "Compatibility matrix CUDA"
Find:

In order to run a CUDA application, the system should have a CUDA enabled GPU and an NVIDIA display driver that is compatible with the CUDA Toolkit that was used to build the application itself. If the application relies on dynamic linking for libraries, then the system should have the right version of such libraries as well.

So basically, it jest werks™.

 

[Krteq]

Well, that's why you have HIP backend in ROCm to port CUDA code to HIP

... or what exactly are we discussing here?

 

[Tup3x]

I get it and mainly agree but there's definitely also a marketing impact. For example, the RTX 3050 is about 3.7x more popular than the RX 6600 (on Steam HW Survey) and the gap is growing wider. But one of those is cheaper, more power efficient, 30% better raster performance, and even has a frame gen.

If it was entirely rational product choice I'd expect less one-sided results than 78.7% choosing the RTX 3050. Maybe it's all DLSS but I'm pretty sure there's halo marketing from having the comically expensive RTX 3090 and 4090.

Well, yeah, definitely. It's just that NVIDIA knows how to use their position to sell lower end garbage. It's part of their marketing. As long as it's GeForce it will sell. Maybe they can flood the market with those too.

Completely false, since when AMD does have the superior product in efficiency, cost effectiveness, raster, etc, it doesn't count.
Radeon can win 9/10 fights and NV will steer the discourse into that one bit they win.

That's the power of marketing and you do not even begin to understand how it works. The product is really only what people perceive it is.

It's not anything simple. You just don't notice any of the subtleties or how you're being manipulated by Nvidia.

Look... It doesn't matter and that's not how things work. It only matters what 12 year old kids think. They would google "best graphics card" and result would be RTX 4090. "Daddy I want PC with GeForce so that I can play CS like a pro." Guess what the daddy end up buying? Laptop with RTX 4050/3050/3060 (or prebuilt desktop PC with similar card). Value matters pretty much only for people who build their own systems (=massive minority).

Even if AMD had superior graphics (in every possible way) cards available it would take a lot of time to gain back market share. It doesn't help that their marketing is trash and probably volume is quite low too. They have had superior APUs and CPUs for quite some time yet their laptop market share is what it is.

 

[ToTTenTranz]

Can you show me the positive reviews for Nvidia’s Ada lineup?

I wonder if there's a single RTX40 card without a "reccommended" stamp on Guru3D, for example. Perhaps the RTX4060 for how attrociously bad it was.

I think you're misremembering- it was an Intel CPU that got the last minute switch (ditching an Athlon), not the Nvidia chipset.

Yes, you're right. They did a last-minute change to an Intel Coppermine instead of an Athlon.

Rather than positive reviews, it's more about how reviews don't really change public opinion.

Well, when someone found out that the RX480 was pulling a bit too much power from the PCIe slot, the card certainly did get a reputation for destroying motherboards (which it never did). The Vega 64 got a reputation for being "too loud to game" when it peaked 2 or 3 dBA above the 1080 Ti.


Regardless, at this point Nvidia knows they get better results with massive astroturfing, agents that pretend to have no bias, and sending engineers to dev houses to implement optimizations. Up until Turing the optimizations were massive geometry with sub-pixel triangles because AMD's cards were slower at geometry. Nowadays it's to implement raytracing with 10 ray bounces because AMD's hardware is so much slower at calculating bounces.
After RDNA4 catches up on RT performance, they'll need to shift to something that actually pushes those tensor cores because AMD's hardware have no tensor cores. If RDNA 3.1 works well, one can only guess where Nvidia is going to push game devs towards.

 

[Saylick]

Regardless, at this point Nvidia knows they get better results with massive astroturfing, agents that pretend to have no bias, and sending engineers to dev houses to implement optimizations. Up until Turing the optimizations were massive geometry with sub-pixel triangles because AMD's cards were slower at geometry. Nowadays it's to implement raytracing with 10 ray bounces because AMD's hardware is so much slower at calculating bounces.
After RDNA4 catches up on RT performance, they'll need to shift to something that actually pushes those tensor cores because AMD's hardware have no tensor cores. If RDNA 3.1 works well, one can only guess where Nvidia is going to push game devs towards.

Simple.

In-game AI chatbots. Nvidia will require game devs to implement them, so if your GPU doesn't have enough TOPS, your in-game NPCs will have room temperature IQs and then you can't enjoy the game. You ask game devs why they can't just have conventionally scripted NPC interaction but they'll keep hush or just say "because reasons", meanwhile they're collecting that Nvidia moolah behind their backs.

You can send the check to my mailbox, Nvidia. You're welcome.

 

[Mahboi]

Simple.

In-game AI chatbots. Nvidia will require game devs to implement them, so if your GPU doesn't have enough TOPS, your in-game NPCs will have room temperature IQs and then you can't enjoy the game. You ask game devs why they can't just have conventionally scripted NPC interaction but they'll keep hush or just say "because reasons", meanwhile they're collecting that Nvidia moolah behind their backs.

You can send the check to my mailbox, Nvidia. You're welcome.

Aha, but NV loooooooooooooooves to gimp VRAM to get customers coming back.
LLM usage is going to very much go against that strat.

 

[adroc_thurston]

In-game AI chatbots

total VRAM death? daring aren't we