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

 

[Win2012R2]

not wanting the Nvidia comparisons, because 8800 was a great GPU.

Likely this

 

[DaaQ]

Likely this

Yea it's weird, you have to go all the way back to 2001 to find a 8xxx labeled GPU.

Here's a link. ( I know VCs, but has decent info) also going to imbed.
LINK

 

[linkgoron]

Yeah, made 0 profit and caused an untold amount of brand damage.

Relying on x2 cards, crossfire and small dies was just an invitation to get destroyed. I'm sure, however, that they made money off of the cards. I still remember that "small-die" strategy piece on Anandtech. In hindsight, this was the most correct thing from that article, and it's clear that Nvidia's approach has been the winner:

Much like the R300 days, the success of the RV770 was partially ensured by NVIDIA’s failure. Unlike NV30 however, GT200 wasn’t delayed nor was it terribly underperforming - it was simply overpriced. ATI got very lucky with RV770, NVIDIA was tied up making a huge chip and avoided two major risks: 55nm and GDDR5, both of which ATI capitalized on.

The next round won’t be as easy, NVIDIA will be at 55nm and they’ll eventually transition to GDDR5 as well. ATI can’t pull off another Radeon HD 4800 launch every year, so chances are 2010 will be closer. Even today NVIDIA has managed to close the gap quite a bit by aggressively pricing the GeForce GTX 260 Core 216, but there’s still the problem of there not being any mainstream GT200 derivative nor will there be until sometime in 2010. Not to mention the impact of selling a 576mm^2 die at the same price as ATI selling a 260mm^2 die will have on NVIDIA’s financials.

If $600 (and probably closer to 700€ in Euroland) for RTX 4070 Ti performance in 2025 is AMD's idea of price-point leadership then there's little hope for AMD in the consumer graphics department.

I haven't used a Nvidia GPU in my main gaming desktop for almost decades now, but I will again if AMD is set on doing the silly "Nvidia minus 5%" dance again.
I'm not going to sidegrade my 4 year-old 6900XT for 700€, that's just stupid. Might as well just pay 1500€ for an actual upgrade from Nvidia.

What were you expecting? Even if they hit 7900XT performance (and rumors are now putting N48 at 5-10% slower), this card was never going to be a meaningful upgrade for a 6900XT (read: at least 50%). Would you sidegrade for $499? What would be the point. Realistically, Nvidia was always your only option.

 

[gaav87]

What were you expecting? Even if they hit 7900XT performance (and rumors are now putting N48 at 5-10% slower), this card was never going to be a meaningful upgrade for a 6900XT (read: at least 50%). Would you sidegrade for $499? What would be the point. Realistically, Nvidia was always your only option.

True... Even if it is 7900xt raster only option 5070ti... or wait for rtx5000 super in 12-16months or just buy 5090... Sad gpu market

 

[techjunkie123]

MI300X vs H100 vs H200 Benchmark Part 1: Training – CUDA Moat Still Alive

Intro SemiAnalysis has been on a five-month long quest to settle the reality of MI300X. In theory, the MI300X should be at a huge advantage over Nvidia’s H100 and H200 in terms of specifications an…

semianalysis.com

Mi300X vs H100 vs H200 for AI. MI300X doesn't come close to achieving paper performance using public code.

 

[ToTTenTranz]

Would you sidegrade for $499?

Up to 550€ I would.

What would be the point.

Better upscaling and raytracing performance at a small cost, assuming I could still sell my 6900XT for ~350-400€ rather fast in the 2nd hand market. The upgrade would cost me 150-200€ and to me that's fine.
What makes no sense whatsoever is to pay 300-350€ for that sidegrade.

Realistically, Nvidia was always your only option.

N48 for me was going to be a stopgap until UDNA and I was fine with dipping again on a true high-end offer a year later.

Of course if AMD is coming up with these ridiculous prices and I'll be spending ~1500€ on a new GPU from Nvidia, I'm not upgrading in another 3 or 4 years.

 

[linkgoron]

N48 for me was going to be a stopgap until UDNA and I was fine with dipping again on a true high-end offer a year later.

Who knows what UDNA will bring to the table... It might end up as another midrange only gen like RDNA4 or a relative flop like RDNA3.

Of course if AMD is coming up with these ridiculous prices and I'll be spending ~1500€ on a new GPU from Nvidia, I'm not upgrading in another 3 or 4 years.

Why not just get a 7900xtx? It's supposedly about equal to the N48 in RT, but still faster by a margin in raster, and would cost you much less than 1500€.

 

[adroc_thurston]

I'm sure, however, that they made money off of the cards

Hell no, early node adoption and GDDR5 were a tax.
There's a good reason the first lineup released under Rory Read (Southern Islands) was waayyy more expensive across the stack.

Who knows what UDNA will bring to the table...

No such thing.

It might end up as another midrange only gen like RDNA4 or a relative flop like RDNA3.

It has one more die at probably like $999 and that's it.

 

[tajoh111]

If Navi 48 is indeed less than 250 mm2, along with it being monolithic and made on 4nm, don't expect a big cu bump and clock speed bump. I think similar CU and 10% higher clocks along with a architecture improvements would be something more realistic. Something like this might yield a 20% increase in performance.

4nm is only 6% denser than 5nm and the performance characteristics aren't that much better either.

A 20% performance increase vs a 7800xt would actually be a good result considering the modest die size regression, and modest manufacturing node improvement.

If you want to look at something with a similar increase in specs and a modest die shrink, look at navi 23 vs navi 33.

7nm to 6nm yields a greater transistor density improvement vs 5nm to 4nm and considering the regression in die size will likely be similar, a 20% increase in performance would be great since the increase in performance of navi 33 was like 7% or 8% vs navi 23.

A 10 percent performace boost on a die 240mm2, would represent a 54% performance per mm2 increase on a process that is relatively the same. That's actually really good. Something like a bit better than a 7900xt at 240mm2 would represent a 90% increase in performance per mm2. That's a miracle. That has never happend any time recently and certainly without the aid of a much better node increase.

It would be something akin to the RTX 3090 to an imaginary RTX 4090 ti which had a massively different process node improvement.

I don't know how people are getting such wild performance expectations from a 240mm2 4nm die. Huge race tracing improvement + a 90% improvement in performance per die area? That seems simply impossible. If Navi4 was such a good architecture, AMD should have just gone for a monolithic rdna4. I.E a 450mm2 die with 80-90% more performance on 4nm which would have been rediculously good.

But that's why I have doubts about the performance...Look back at my performance predictions about RDNA3... and you will see I was the only one on this forum that predicted it's performance accurately.

Predicted a 20% improvement over at 7800xt last year.

I was the only one here that predicted RDNA3 performance remotely correct last time around.

Those timespy results are indicating I could be right again. Lets see.

 

[Jan Olšan]

Videocardz has a database with HD8xxx cards, but they have no MSRPs and the release dates coincide with Radeon 200 series.

They came out, they were an OEM-only refresh of HD 7000 series. They launched at CES 2013 if you want to look up old news coverage.
Oland GPU first appeared in that series.

Not sure how much these SKUs actually appeared in OEM PCs and if you can find them on auction servers and so on as consequence.

 

[poke01]

Shill points will move elsewhere.
You need to behead the comp, or you die.

This, Nvidia has incredible mindshare. Total I mean 200% effort is needed to break that. This is what happens when your comp is you know actually competent for the last 15 years.

 

[poke01]

Also any rumours on if RDNA4 will include Neural Texture AI feature like RTX 50. If not that’s another “feature” that Radeon will be behind on and will likely come later.

I hope for once AMD actually puts some proper effort next generation for RDNA5 or whatever because not having a halo SKU is silly and defeats the purpose of increasing mindshare.

 

[Joe NYC]

It has one more die at probably like $999 and that's it.

Monolithic or chiplet?

 

[Kepler_L2]

Also any rumours on if RDNA4 will include Neural Texture AI feature like RTX 50. If not that’s another “feature” that Radeon will be behind on and will likely come later.

I hope for once AMD actually puts some proper effort next generation for RDNA5 or whatever because not having a halo SKU is silly and defeats the purpose of increasing mindshare.

https://videocardz.com/newz/amd-to-present-neural-texture-block-compression-technology

 

[adroc_thurston]

This is what happens when your comp is you know actually competent for the last 15 years.

No it's a case of ATi doing a market suicide in 2008.
They should've never, EVER gone with the small die strategy. They handed NV a free win.
Gave them a 9-year gap between R600 and Fiji to do whatever they wanted, and by the time Fiji came, it was already too late.
Franky even after that they didn't try.

Also any rumours on if RDNA4 will include Neural Texture AI feature like RTX 50. If not that’s another “feature” that Radeon will be behind on and will likely come later

Who cares, PPAmaxxing is all that matters.

Monolithic or chiplet?

Of course it's a monodie. People who can make tiled GPUs work are busy with other things.

 

[H T C]

I hope for once AMD actually puts some proper effort next generation for RDNA5 or whatever because not having a halo SKU is silly and defeats the purpose of increasing mindshare.

The only way i see for AMD to increase their mindshare SIGNIFICANTLY is by offering GPUs @ "mainstream" prices WITH "enthusiast" level of performance: it would cost AMD quite a lot, ESPECIALLY @ current wafer prices, but IT SHOULD pay off in the long run, because it would COMPLETELY disrupt current pricing in their favor, which is THE WHOLE POINT.

That said, there's NO WAY that AMD will EVEN CONSIDER such an approach because:

A - it would be TERRIBLY expensive, @ 1st: it would take A LONG TIME for it to start paying off
B - nVidia could actually afford to fight back, though it would cost them EVEN MORE
C - it would CERTAINLY be a gamble: one that, in case it backfires, would put AMD in a VERY DIFFICULT position, as a GPU manufacturer

 

[adroc_thurston]

would put AMD in a VERY DIFFICULT position, as a GPU manufacturer

no becase they make AMD Instinct. Client dgfx is just a tiny side job, they need gfx IP for other, bigger and better things.

 

[H T C]

no becase they make AMD Instinct. Client dgfx is just a tiny side job, they need gfx IP for other, bigger and better things.

I didn't mean server grade GPUs: only desktop grade.

 

[poke01]

no becase they make AMD Instinct.

they make great hardware there but their software stack still sucks.

Hopefully, mangagement is serious about AMD becoming software company. Its honestly culture, Nvidia has it, AMD needs to cultivate it.

 

[adroc_thurston]

but their software stack still sucks.

No it works pretty well already but there are still many things to be done wrt training.

Hopefully, mangagement is serious about AMD becoming software company. Its honestly culture, Nvidia has it, AMD needs to cultivate it.

that's a meme. Neither is an s/w company.

 

[DaaQ]

They came out, they were an OEM-only refresh of HD 7000 series. They launched at CES 2013 if you want to look up old news coverage.
Oland GPU first appeared in that series.

Not sure how much these SKUs actually appeared in OEM PCs and if you can find them on auction servers and so on as consequence.

Yea back then I didn't even look at OEM stuff.
In that list none of the HD 8xxx have a MSRP. Pretty sure was all rebrand.

Yea Oland came out as R7 250 for $89 like 18 months later.

 

[Shmee]

Hmm, if true that AMD names their RDNA4 cards this way, that is pretty stupid IMO. I would much prefer a naming scheme that made more sense. Of course will likely be waiting for UDNA anyway before upgrading, at the least.

 

[marees]

Surely they won't price a xx70 card at $600 🤔

7900 gre stocks are sold out now. This could mean one of 2 things:

1. No old stocks to worry about. Price new stocks aggressively
2. Otoh, no pricing pressure from old cards. Price new stocks conservatively

Which way AMD 🤔🤔

 

[SolidQ]

Surely they won't price a xx70 card at $600

make it 600$ very small amount will be sold , then street prices gonna be 500$, and people starting buy, that always AMD way overprice MSRP > street prices fixes sales

 

[gdansk]

Surely they won't price a xx70 card at $600 🤔

$550 would be typical Radeon fashion (i.e. not exactly compelling but cheaper than Nvidia).