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

 

[lucasworais]

The cards are reasonably priced (considering the market) and widely available here in Brazil.
The cheapest 8gb model I could find is at ~$450 and the 16gb at around $530.
For reference, the cheapest 5060 is at $450, the 5060ti 8gb at $540 and the 16gb one at $685.

 

[PJVol]

Just watched the Daniel's review of both 8 and 16 gb 9060 xt's and I wonder why I'm not surprised to see the HWUB salty comment there )

 

[Timorous]

Just watched the Daniel's review of both 8 and 16 gb 9060 xt's and I wonder why I'm not surprised to see the HWUB salty comment there )

Salty about what? It was clearly dripping with sarcasm.

 

[PJVol]

Salty about what? It was clearly dripping with sarcasm.

Yeah, it'd be better to say that his trolling seemed a little nervous to me.

 

[eek2121]

Honestly, I am not as angry as some are regarding 8gb VRAM for some SKUs. I get that those cards don’t run the most demanding games at the highest settings.

Consoles tend to hold back progress a lot, so until we see 24-32gb of memory, we likely won’t see any real issues beyond PC specific titles.

I get the reason reviewers are hounding AMD/NVIDIA, however. My only real beef there is that at least AMD included a PCIE X16 interface.

 

[igor_kavinski]

Honestly, I am not as angry as some are regarding 8gb VRAM for some SKUs.

If you thought about how it's a waste of a perfectly good GPU with VRAM that cannot be upgraded or replaced, you would be. There's absolutely no good reason to hobble a 9060 XT with lower VRAM. They should've reserved that amount of VRAM only for RX 9050 XT and maybe 6GB for RX 9040.

 

[Thunder 57]

If you thought about how it's a waste of a perfectly good GPU with VRAM that cannot be upgraded or replaced, you would be. There's absolutely no good reason to hobble a 9060 XT with lower VRAM. They should've reserved that amount of VRAM only for RX 9050 XT and maybe 6GB for RX 9040.

What 9050 and 9040?

 

[igor_kavinski]

What 9050 and 9040?

I'm expecting them to release those because why not?

 

[Rigg]

W1zard's relative performance charts are completely borked in comparison to the average FPS chart on all of the 9060 XT reviews. Someone pointed this out in the Asus Prime OC review when they noticed the positioning relative to the 3070. W1z responded and said that TLOU crashing is throwing this off. That response didn't really make any sense to me since the average FPS and relative performance charts should directly correlate. At least according to the description in the relative performance section of the reviews.

It looks like the TLOU results aren't counted in the relative performance chart but are factored into the average FPS chart. The 1080P RP chart shows 4-5% advantage for the 5060 Ti while the FPS charts shows a 10-11% advantage. Apparently this wasn't worthy of mentioning in the review anywhere though. I'm not sure how he couldn't get a valid result in TLOU either. HUB was able to test this game without issue. His sycophants on the TPU forum seem to hold his data in pretty high regard. I really don't get it.

To W!zard's credit, he added a note about this in the 9060 XT reviews after I suggested it in the PowerColor Radeon RX 9060 XT Reaper 8 GB review thread. It appears this calculation is not actually using geomean either. He explained as follows for those who don't want to read the TPU thread:


I'm still not 100% clear on how this is calculated from this explanation. I suggested he improve the description in the relative performance section and add a clear explanation of how this is calculated. I also asked him to address concerns about textures in the Hogwarts and Monster Hunter ray tracing data. Since other reviewers have shown that there are issues with textures on 8 GB cards while using the same settings, I sincerely hope he takes a closer look at his testing. Maybe he'll watch the Daniel Owens video i linked and realize some of his testing methodology could be much better for testing VRAM limitations. I'm sure the usual suspects over there will have snarky replies to my posts, but hopefully W1z will address the texture concerns.

 

[igor_kavinski]

At least he's receptive to your suggestions. I've read around here that he has a pretty twitchy ban finger.

 

[blckgrffn]

Honestly, I am not as angry as some are regarding 8gb VRAM for some SKUs. I get that those cards don’t run the most demanding games at the highest settings.

Consoles tend to hold back progress a lot, so until we see 24-32gb of memory, we likely won’t see any real issues beyond PC specific titles.

I get the reason reviewers are hounding AMD/NVIDIA, however. My only real beef there is that at least AMD included a PCIE X16 interface.

I get what you are saying, but the consoles are clear on how much vram to include: 12GB. The PS5 is not going away. The series s continues to be the boat anchor for the current gen, but that’s sort of laughable when it gets the Xbox One graphics presets anyway.

12GB was the sweet spot for a minimum of this GPU generation with aspirations of solid performance and avoiding stuttering with out of the box settings for games.

Yet here we are.

Hopefully 3GB GDDR7 modules really do clean this up in 2026 onwards.

 

[DAPUNISHER]

To W!zard's credit, he added a note about this in the 9060 XT reviews after I suggested it in the PowerColor Radeon RX 9060 XT Reaper 8 GB review thread. It appears this calculation is not actually using geomean either. He explained as follows for those who don't want to read the TPU thread:

View attachment 125165
I'm still not 100% clear on how this is calculated from this explanation. I suggested he improve the description in the relative performance section and add a clear explanation of how this is calculated. I also asked him to address concerns about textures in the Hogwarts and Monster Hunter ray tracing data. Since other reviewers have shown that there are issues with textures on 8 GB cards while using the same settings, I sincerely hope he takes a closer look at his testing. Maybe he'll watch the Daniel Owens video i linked and realize some of his testing methodology could be much better for testing VRAM limitations. I'm sure the usual suspects over there will have snarky replies to my posts, but hopefully W1z will address the texture concerns.

Well done dude!

Perhaps he isn't shilling, just an old dog having trouble learning new tricks? I despise geomean, and he may still be pulling the lies, damned lies, and statistics schtick, but I doubt it is more deceptive than geomean can be. If he starts taking the time and doing the work to expose issues with vram (it has been a hot button topic for years now) he will dispel most of the suspicions arising from his current testing methodology. Even if he does it as follow up content.

The 8GB 9060xt would be compelling with a $70 price drop. FSR 4 not being in nearly as many games is a dumb talking point. Most of those games this card can run at the advertised/targeted resolution of 1080 without the need for upscaling, and without framebuffer overflow. The list of games where you are stuck with older FSR and need it is a fraction of the size, and every new title not sponsored and engineered by Nvidia, for Nvidia hardware, should have the newest iteration.

The other marketing/used car salesman tactic is calling the Nvidia showcase suite of what? 6 games? a transformative experience. It gets a chuckle from me, because my gut says it'll be like DLSS since 2. Every version gets praised until the next hits, then it's suddenly garbage. 🤡 Maybe that's just me though. I do get overly aggro about inane/dumb 💩

 

[marees]

Just watched the Daniel's review of both 8 and 16 gb 9060 xt's and I wonder why I'm not surprised to see the HWUB salty comment there )

So besides PS5 games (Spiderman-2, Ratchet & clank etc.) all open world games with nanite like tech seem to have an issue with 8gb cards:

1. Oblivion Remastered — UE5
2. Monster Hunter Wilds — Red Engine
3. Kingdom Come Deliverance II — Cry Engine
4. Assassin's Creed shadows — Anvil engine

 

[Thunder 57]

Well done dude!

Perhaps he isn't shilling, just an old dog having trouble learning new tricks? I despise geomean, and he may still be pulling the lies, damned lies, and statistics schtick, but I doubt it is more deceptive than geomean can be. If he starts taking the time and doing the work to expose issues with vram (it has been a hot button topic for years now) he will dispel most of the suspicions arising from his current testing methodology. Even if he does it as follow up content.

The 8GB 9060xt would be compelling with a $70 price drop. FSR 4 not being in nearly as many games is a dumb talking point. Most of those games this card can run at the advertised/targeted resolution of 1080 without the need for upscaling, and without framebuffer overflow. The list of games where you are stuck with older FSR and need it is a fraction of the size, and every new title not sponsored and engineered by Nvidia, for Nvidia hardware, should have the newest iteration.

The other marketing/used car salesman tactic is calling the Nvidia showcase suite of what? 6 games? a transformative experience. It gets a chuckle from me, because my gut says it'll be like DLSS since 2. Every version gets praised until the next hits, then it's suddenly garbage. 🤡 Maybe that's just me though. I do get overly aggro about inane/dumb 💩

Geomean is a person? I thought you meant you despise geomean. If there is anyone to despise there it is justbenching. But I don't spend much time there so I could be wrong.

 

[DAPUNISHER]

Geomean is a person? I thought you meant you despise geomean. If there is anyone to despise there it is justbenching. But I don't spend much time there so I could be wrong.

https://en.wikipedia.org/wiki/Geometric_mean

 

[Rigg]

Well done dude!

Perhaps he isn't shilling, just an old dog having trouble learning new tricks? I despise geomean, and he may still be pulling the lies, damned lies, and statistics schtick, but I doubt it is more deceptive than geomean can be. If he starts taking the time and doing the work to expose issues with vram (it has been a hot button topic for years now) he will dispel most of the suspicions arising from his current testing methodology. Even if he does it as follow up content.

The 8GB 9060xt would be compelling with a $70 price drop. FSR 4 not being in nearly as many games is a dumb talking point. Most of those games this card can run at the advertised/targeted resolution of 1080 without the need for upscaling, and without framebuffer overflow. The list of games where you are stuck with older FSR and need it is a fraction of the size, and every new title not sponsored and engineered by Nvidia, for Nvidia hardware, should have the newest iteration.

The other marketing/used car salesman tactic is calling the Nvidia showcase suite of what? 6 games? a transformative experience. It gets a chuckle from me, because my gut says it'll be like DLSS since 2. Every version gets praised until the next hits, then it's suddenly garbage. 🤡 Maybe that's just me though. I do get overly aggro about inane/dumb 💩

Yeah I don't know. We might be a bit oversensitive to it, but he does seem to hit the Nvidia talking points in his review conclusions. Min-Max normalizing the data to a scale of 0 to 100 for the relative performance chart does seem like a pretty good way to do it honestly. As far as the testing and data goes generally, I find it more likely he's clinging to dated testing/data presentation standards rather than providing intentionally misleading data. I'm interested to see if he has a response to the texture issue. The odds seem pretty good that his Monster Hunter RT data for 8 GB cards is 💩.

I would have personally preferred that all fully functioning N44 dies destined for desktop were in 16 GB cards. A cut down N44 9060 for sub $250 is a better fit for 8 GB IMO. I suspect we won't see these 8 GB 9060 XT cards get much production after the initial wave and it will get effectively replaced by a cheaper 9060.

If OptiScaler is all it's cracked up to be FSR version shouldn't really matter that much. I honestly wouldn't really know. I don't play that many super demanding brand new AAA games with RT, and I upgrade my GPU fairly often, so I've never wanted/needed to use any AI upscaling. I usually let the dust settle on AAA games a few years and wait for a Steam sale. Assuming OptiScaler works well, I don't know why AMD can't make a first party tool that does the same thing.

 

[GaiaHunter]

Yeah, it'd be better to say that his trolling seemed a little nervous to me.

Why?
Daniel was just confirming what he said.
With AI gobbling silicon that could be gaming GPUs, why are the companies wasting it by pairing it with low amounts of VRAM?

Daniel new video is a good perspective.

 

[Thunder 57]

https://en.wikipedia.org/wiki/Geometric_mean

I know what geometric mean is. They way you worded it made it sound like a user. But the "he" you were referring to was clearly wizzard. That should've been made clear to me in your next sentence.

 

[Thunder 57]

Why?
Daniel was just confirming what he said.
With AI gobbling silicon that could be gaming GPUs, why are the companies wasting it by pairing it with low amounts of VRAM?

Daniel new video is a good perspective.

I feel obligated to say it every time, but "Stupid youtube image confirmed". Also, those teeth. Has this guy ever heard of a toothbrush? At least they're better than the DF guy.

 

[DAPUNISHER]

What a dumb upgrade

 

[Thunder 57]

What a dumb upgrade

Not even hyperthreading.

 

[SolidQ]

What a dumb upgrade

unless, it's temporary upgrade and next is platform upgrade

 

[igor_kavinski]

The guy might be satisfied if the fps doubled in his games. I bet the original Crysis runs well on this. I suspect his plan might be to see how long that system can loyally serve him before it croaks.

 

[Thunder 57]

The guy might be satisfied if the fps doubled in his games. I bet the original Crysis runs well on this. I suspect his plan might be to see how long that system can loyally serve him before it croaks.

PCIe 2 with a 15 year old CPU? I can't even imagine what kind of memory is in that thing. Were NVMe drives available back then? Don't think so. It clearly needs an upgrade.

 

[DAPUNISHER]

unless, it's temporary upgrade and next is platform upgrade

Guy mentioned money is really tight and it could be up to 2 years before he does the platform upgrade. If you want to call that temporary.