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

 

[Mahboi]

It's not great Jim, GPUs that don't exist have a hard time competing.

 

[Tuna-Fish]

What's the over/under on the 8900xtx (RDNA4 version of 7900xtx) reaching at least 4090 performance or better?

No such thing is going to get released. There are only 2 RDNA4 GPU dies, a small one and a tiny one. The chiplet-based high end got cancelled.

 

[darkswordsman17]

What's the over/under on the 8900xtx (RDNA4 version of 7900xtx) reaching at least 4090 performance or better?

Well, 4090 x 2 = 8180
8900 - 8180 = 720

So, maybe if you have it render at 720p?

 

[linkgoron]

What's the over/under on the 8900xtx (RDNA4 version of 7900xtx) reaching at least 4090 performance or better?

There won't be a 8900xtx, and current rumors place Navi48 at around 7900XT performance at best. Chiplets worked out so well that AMD decided to scrape them and instead go without a high-end product for the next two years.

 

[blckgrffn]

What's the over/under on the 8900xtx (RDNA4 version of 7900xtx) reaching at least 4090 performance or better?

0? Doesn’t that sound like it’s happening.

….

Edit:

Haha, definitely should have read to the end of the thread first. That I had, turns out I was just at the bottom of the page…

 

[marees]

Damn, you were faster.

Most Highly Credible And Believeable and True Leak of All Time™

He is credible for sure.

Except this time he seems to be messing with us

Did you see his zen5 DOA slide ?????

 

[PJVol]

Slides... slides...

 

[marees]

View attachment 99904
I laugh indeed.

Seriously, this is the slide Lisa Su should be presenting in computex

 

[gdansk]

Seriously, this is the slide Lisa Su should be presenting in computex

No they have to go back and fix RDNA3 to be... what is it 35-40%?
Also are we really expecting more than that for RDNA4?

 

[Joe NYC]

This is more CDNA2 / CDNA3 related. It seems that first, the work that went into Frontier supercomputer, with Cray / HPe Sligshot became the basis of Ultra Ethernet.

Now, AMD is donating majority of IP that is going into Ultra Accelerator Link, that has most of the industry (except NVidia) behind it.

It seems that on top of NVidia overcharging on the GPUs, NVidia is overcharging the entire industry on all of the networking components, which NVidia is trying to impose on the industry, as a condition of being able to obtain NVidia GPUs.

Key Hyperscalers And Chip Makers Gang Up On Nvidia’s NVSwitch Interconnect

The generative AI revolution is making strange bedfellows, as revolutions and emerging monopolies that capitalize on them, often do. The Ultra Ethernet

www.nextplatform.com

https://twitter.com/x/status/1796189975110721648

 

[Mahboi]

He is credible for sure.

Except this time he seems to be messing with us

Did you see his zen5 DOA slide ?????

Olrak is a bored student that just uses his Twitter account to mess with people, I'm not sure why anyone takes him seriously, he doesn't do anything serious ever.

 

[beginner99]

Most Highly Credible And Believeable and True Leak of All Time™

In this time and age not even slides like this can be trusted due to AI.

It's believable because it fits the other rumors, much better RT, else on 7900XT level.

 

[Mahboi]

In this time and age not even slides like this can be trusted due to AI.

It's believable because it fits the other rumors, much better RT, else on 7900XT level.

What?
Are you seriously believing anything about this slide? There's no need to mention AI...
It's 500% fake, and incredibly obviously so. AMD would never ever put a "0.9x" on any game on the presentation slide.

The writing isn't perfectly aligned.

It's not going to be called 8800 XTX (for God's sake anyway I hope it's not).
And so on. It's an RDNA 3 slide that's been doctored for funnies.

 

[coercitiv]

What?
Are you seriously believing anything about this slide? There's no need to mention AI...

The slide is perfectly legit, from the same deck as the one for Zen 5%.

Spoiler

Introducing XTX moniker for the x800 SKU is also to be expected, as a direct branding counter to RTX. All AMD cards going forward will be branded XTX.

 

[Mahboi]

Introducing XTX moniker for the x800 SKU is also to be expected, as a direct branding counter to RTX. All AMD cards going forward will be branded XTX.

On a more serious note, I was adverse to this, but all things considered, it's not that bad an idea. Putting XTX on the top of the line of each gen is a fairly easy way to recognise the flagship. It's also practical to write, I keep just writing "XTX" rather than "7900 XTX".
Even if it's gonna sound silly on a "8700 XTX", it is indicative that RDNA 4 doesn't go above 7 class, which is very good naming. It's a zillion times better than the 7900 XT was, considering that it was going to originally face the "4080 12 Go".

 

[eek2121]

What's the over/under on the 8900xtx (RDNA4 version of 7900xtx) reaching at least 4090 performance or better?

There is no 8900XTX.

 

[eek2121]

The slide is perfectly legit, from the same deck as the one for Zen 5%.

Spoiler

View attachment 100093

Introducing XTX moniker for the x800 SKU is also to be expected, as a direct branding counter to RTX. All AMD cards going forward will be branded XTX.

I really need to mod that chart and make it -5%...

 

[Saylick]

I really need to mod that chart and make it -5%...

Or you can wait and UserBenchmark will eventually do it for you.

 

[Hans Gruber]

There is no 8900XTX.

I was in a hurry when I wrote the original post because I had to leave my computer for many hours. What I meant to say was will RDNA4's top SKU be equal to or better or worse than the 4090. Since AMD had not announced what their top SKU was at the time of the post, I didn't know what to call it.

I guess the 8800xtx is what it's going to be called. But with RDNA3 losing 25-30% performance due to a design error with inoperable silicon on the GPU for higher tier GPU's. Does this mean even a gimp but fully functional RDNA die produce performance numbers equal to a 4090? Nvidia being Nvidia will have a 5090 that smokes the 4090 with the Blackwell release.

Anybody have thoughts on this? Let's assume RDNA gives a 15% boost in performance simply because of the silicon N4P vs N5 (RDNA3). If you add generational optimizations and improvement add another 30-40%. Then factor in the missing performance because of the design issue with RDNA3. Should the 8800xtx not be close to a 4090 in performance?

If forum members say no and the 5090 is a 50% or more improvement over the 4090. AMD would be irrelevant in the GPU sector. Add to it that ARC was broken as well but they used software workarounds to make the GPU pretty functional. If Battlemage fixes all the design issues with ARC and solidifies their drivers. What does this say about AMD?

If if the top SKU of Battlemage is equal to a 4070 and close to a 4070ti. Where would the 8800xtx fit in? Are we talking 4080 performance or 4090 performance for AMD?

 

[blckgrffn]

I was in a hurry when I wrote the original post because I had to leave my computer for many hours. What I meant to say was will RDNA4's top SKU be equal to or better or worse than the 4090. Since AMD had not announced what their top SKU was at the time of the post, I didn't know what to call it.

I guess the 8800xtx is what it's going to be called. But with RDNA3 losing 25-30% performance due to a design error with inoperable silicon on the GPU for higher tier GPU's. Does this mean even a gimp but fully functional RDNA die produce performance numbers equal to a 4090? Nvidia being Nvidia will have a 5090 that smokes the 4090 with the Blackwell release.

Anybody have thoughts on this? Let's assume RDNA gives a 15% boost in performance simply because of the silicon N4P vs N5 (RDNA3). If you add generational optimizations and improvement add another 30-40%. Then factor in the missing performance because of the design issue with RDNA3. Should the 8800xtx not be close to a 4090 in performance?

If forum members say no and the 5090 is a 50% or more improvement over the 4090. AMD would be irrelevant in the GPU sector. Add to it that ARC was broken as well but they used software workarounds to make the GPU pretty functional. If Battlemage fixes all the design issues with ARC and solidifies their drivers. What does this say about AMD?

If if the top SKU of Battlemage is equal to a 4070 and close to a 4070ti. Where would the 8800xtx fit in? Are we talking 4080 performance or 4090 performance for AMD?

What?

No.

There is no high end competitor chip from AMD.

Remember the RDNA1 launch? The 5700XT being the top of the range and competing against the ~2070 but definitely not the 2080ti?

That's what we are looking at again.

It's looking like we might get a ~$550 card that is uses ~7800xt levels of power and delivers performance somewhat north of a 7900XT but likely less in raster maybe better in RT than the 79000XTX. Maybe even lower on the power usage front. And maybe lower cost. All extracted from a very efficiently laid out die, as Adroc has told us that PPA was the defining characteristic of RDNA4.

That's the pretty definite steam, anyway.

 

[linkgoron]

I was in a hurry when I wrote the original post because I had to leave my computer for many hours. What I meant to say was will RDNA4's top SKU be equal to or better or worse than the 4090. Since AMD had not announced what their top SKU was at the time of the post, I didn't know what to call it.

I guess the 8800xtx is what it's going to be called. But with RDNA3 losing 25-30% performance due to a design error with inoperable silicon on the GPU for higher tier GPU's. Does this mean even a gimp but fully functional RDNA die produce performance numbers equal to a 4090? Nvidia being Nvidia will have a 5090 that smokes the 4090 with the Blackwell release.

Anybody have thoughts on this? Let's assume RDNA gives a 15% boost in performance simply because of the silicon N4P vs N5 (RDNA3). If you add generational optimizations and improvement add another 30-40%. Then factor in the missing performance because of the design issue with RDNA3. Should the 8800xtx not be close to a 4090 in performance?

If forum members say no and the 5090 is a 50% or more improvement over the 4090. AMD would be irrelevant in the GPU sector. Add to it that ARC was broken as well but they used software workarounds to make the GPU pretty functional. If Battlemage fixes all the design issues with ARC and solidifies their drivers. What does this say about AMD?

If if the top SKU of Battlemage is equal to a 4070 and close to a 4070ti. Where would the 8800xtx fit in? Are we talking 4080 performance or 4090 performance for AMD?

You're not reading what people wrote you:

There won't be a 8900xtx, and current rumors place Navi48 at around 7900XT performance at best. Chiplets worked out so well that AMD decided to scrape them and instead go without a high-end product for the next two years.

The top RDNA4 card is going to be slower than the 7900xtx (I would bet - slower than 7900xt). Rumors state that it's a 240mm^2 GPU.

 

[adroc_thurston]

that is uses ~7800xt levels of power

That's not mainstream.

as Adroc has told us that PPA was the defining characteristic of RDNA4.

Actually ray tracing.
PPA is still a milestone you gotta hit or people gotta get lobbed off (again).

 

[blckgrffn]

That's not mainstream.

Sorry, I am not so well informed.

I am reading that as you saying that's "too high for the mainstream the card is targeting."

Is that what you are saying?

Accepting Yes/No/Higher/Lower as answers if you want to keep it signature brief. I enjoy your posts since they aren't walls of text, FWIW.

 

[adroc_thurston]

I am reading that as you saying that's "too high for the mainstream the card is targeting."

Yeah, ~>250W is high-end territory for either vendor.

 

[blckgrffn]

Yeah, ~>250W is high-end territory for either vendor.

I like/support this mentality and design goal. I plan to support it with dollars.