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

 

[Keller_TT]

I wouldn't rule out 549 for the XT and 479 for the non-XT.

AMD is aware of their brand and feature disadvantage, and with slightly cheaper memory, lower TDP (can go with somewhat cheaper cooling and board) and no extra cost from InFO packaging, the 9070 should be at least a little cheaper to make than the 7800XT.

It is already tooo high for the non-XT and will be a non-starter. If it is indeed a 7900 XT across the board (raster & RT) and keeps a 240W wattage, the max AMD can ask for is $449, but I would expect that price to drop post launch.
XT - max $529 for MBA card.

The non-XT card's primary competitor is the 4070S, and will likely be around $500 when the 5070 hits the market. The Super tier will co-exist with the 50 series by how it's shaping up.

 

[Keller_TT]

can go with somewhat cheaper cooling and board) and no extra cost from InFO packaging

They will be reusing the 7900GRE reference design with a new design aesthetic for the MBA XT this time. It was a 7900XT reference design with a thinned heatsink.

 

[Josh128]

Just for curiosity's sake, how would you rate your confidence in that compared to, say, your confidence that Zen5 is a 40% ipc uplift?

Not sure about thurston, but I can see no other reason why they need to let Nvidia go first, unless RDNA 4 performance/price will make a huge splash and AMD want to compare directly to them.

 

[Heartbreaker]

the 5070 has less SMs than a 4070 Super and might not even match it. if 9070 non-xt is 7900xt in raster, it will be clearly faster than a 5070 with more vram.

I stand by my previous statement:

9070 XT for $599
9070 for $499

With 9070 having less CU and lower clocks.

That is the most optimistic pricing. these won't be any cheaper than that and much more likely the price is higher.

Good to see that not everyone wants to lowball AMD into oblivion. I've been seeing absurd things like the 9070 XT must be $400 or it will fail (most likely from someone that really just wants a price war, to get a cheaper NVidia card).

AMD at minimum has to at least price against itself sanely. I think ~$100 between XT and Non-XT is just about right, and doesn't repeat the initial mistake of making the 7900XT and 7700XT so close in price to the cards above, that you would question, why anyone would buy them, which was quickly followed by price cuts to correct, but that basically meant initial reviews were negative, and cards seldom get re-reviewed just for price cuts, so the initial negative reviews stand.

Pricing against NVidia is tricky. Especially with the price cut on the 5070. For the following I'm assuming that 9070 has better performance than 5070, but 9070XT underperforms 5070 Ti.

If 5070 had come in at $600, then 9070 at $500 would have been great. Since 9070, would outperform, have greater VRAM, and a 20% discount. We can't reasonably expect AMD to do more than that. But with 5070 at $550, that's more like a 10% discount, and then it will probably be a lot less compelling for a lot more people. I kind of think that with 5070 at $550 it has to be more like $459, with 9070 XT coming in about $100 more.

 

[Thunder 57]

Good to see that not everyone wants to lowball AMD into oblivion. I've been seeing absurd things like the 9070 XT must be $400 or it will fail (most likely from someone that really just wants a price war, to get a cheaper NVidia card).

AMD at minimum has to at least price against itself sanely. I think ~$100 between XT and Non-XT is just about right, and doesn't repeat the initial mistake of making the 7900XT and 7700XT so close in price to the cards above, that you would question, why anyone would buy them, which was quickly followed by price cuts to correct, but that basically meant initial reviews were negative, and cards seldom get re-reviewed just for price cuts, so the initial negative reviews stand.

Pricing against NVidia is tricky. Especially with the price cut on the 5070. For the following I'm assuming that 9070 has better performance than 5070, but 9070XT underperforms 5070 Ti.

If 5070 had come in at $600, then 9070 at $500 would have been great. Since 9070, would outperform, have greater VRAM, and a 20% discount. We can't reasonably expect AMD to do more than that. But with 5070 at $550, that's more like a 10% discount, and then it will probably be a lot less compelling for a lot more people. I kind of think that with 5070 at $550 it has to be more like $459, with 9070 XT coming in about $100 more.

If people are smart they will avoid the 5070 like the plague. The 5070 and 5080 have crap VRAM for the price.

 

[Win2012R2]

So it's about double "perf" by tripling the number of fake frames?

Seems like 50 series won't last 2 years

 

[Heartbreaker]

Someone just posted more NVidia 50 series, performance numbers without Fake Frames, and they are brutal:

Page 78 - Discussion - Nvidia Blackwell in Q1-2025

Page 78 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[SolidQ]

9070 and 9070XT

https://twitter.com/x/status/1879541037389250651

 

[blckgrffn]

Good to see that not everyone wants to lowball AMD into oblivion. I've been seeing absurd things like the 9070 XT must be $400 or it will fail (most likely from someone that really just wants a price war, to get a cheaper NVidia card).

AMD at minimum has to at least price against itself sanely. I think ~$100 between XT and Non-XT is just about right, and doesn't repeat the initial mistake of making the 7900XT and 7700XT so close in price to the cards above, that you would question, why anyone would buy them, which was quickly followed by price cuts to correct, but that basically meant initial reviews were negative, and cards seldom get re-reviewed just for price cuts, so the initial negative reviews stand.

Pricing against NVidia is tricky. Especially with the price cut on the 5070. For the following I'm assuming that 9070 has better performance than 5070, but 9070XT underperforms 5070 Ti.

If 5070 had come in at $600, then 9070 at $500 would have been great. Since 9070, would outperform, have greater VRAM, and a 20% discount. We can't reasonably expect AMD to do more than that. But with 5070 at $550, that's more like a 10% discount, and then it will probably be a lot less compelling for a lot more people. I kind of think that with 5070 at $550 it has to be more like $459, with 9070 XT coming in about $100 more.

IF the AMD parts come in performing above "expectations" like we've been seeing leaked, I honestly think AMD might do something "crazy" like 9070 = $529 and 9070 XT = $629 and the "best" AIB boards at $699.

It it has 7900 XTX perf +/- 10% AND reduced power consumption (even ~350W TBP on the highest models would be significant IMO) then... it just makes sense to me.

I just paid $980 after sales tax for a premium 7900 XTX, I'd rather pay $770 after tax and get a more forward looking part to boot.

And if Nvidia moves the line I expect that AMD to keep prices just under the 5070/5070Ti.

nvidia brought almost nothing with the 5xxx series in terms of actual performance so AMD might actually have an opening.

And yeah, I expect them to maximize profits when they can.

 

[Saylick]

So it's about double "perf" by tripling the number of fake frames?

Seems like 50 series won't last 2 years

Jensen probably doesn’t need it to when they probably prefer that the N4 wafers get used for AI servers where the margins are better, but this would imply they are basically stagnating on the consumer GPU side.

 

[Heartbreaker]

IF the AMD parts come in performing above "expectations" like we've been seeing leaked, I honestly think AMD might do something "crazy" like 9070 = $529 and 9070 XT = $629 and the "best" AIB boards at $699.

It it has 7900 XTX perf +/- 10% AND reduced power consumption (even ~350W TBP on the highest models would be significant IMO) then... it just makes sense to me.

I just paid $980 after sales tax for a premium 7900 XTX, I'd rather pay $770 after tax and get a more forward looking part to boot.

And if Nvidia moves the line I expect that AMD to keep prices just under the 5070/5070Ti.

nvidia brought almost nothing with the 5xxx series in terms of actual performance so AMD might actually have an opening.

And yeah, I expect them to maximize profits when they can.

We now have more information to confirm that NVidia is definitely coming in below expectations. Without the fake frames 50 series are single digit gains over 40 Supers. 5070 in particular seems identical performance to 4070 Super. I think AMD will have no trouble pummeling that one with 9070, though it remains to be seen how close XT will come to 5070 Ti...

 

[Keller_TT]

TPU posted a Blackwell 50 series deep dive, and the neural rendering using "Cooperative Vectors API for DirectX" got me intrigued.
MS and Nvidia have been thick as thieves for ever, and it's likely that there's a lot more Nvidia way of doing ML, RTRT being built into DX12 than AMD's, while AMD have been upstream contributors to Vulkan (the whole project proceeded from Mantle). Vulkan also feeds Sony's proprietary GNM/X and helps Valve/Steam do great stuff.
NV too contribute to Vulkan but DX is their raison d'être.

I'm just curious how all this is going to shape the next few years because I see RDR2, and Indiana Jones do great using Vulkan and other games that run better using Vulkan than DX12. This "Project Amethyst" stuff will be more going into Vulkan, and I would think the API would be best repository of GPUOpen algorithms under research. But the API needs to sort out tooling for more Indie hobbyists to be more productive and feel some joy for the industry to thrive.

The point is, I'm more of a Vulkan guy, dabble a bit with it, and I would like it to become more of the standard. Can AMD provide documentation on how RDNA4 features can be exploited using Vulkan, and present it well with a dedicated website or something? Because that would be really being GPU Opened.

 

[SolidQ]

I see RDR2

My Pascal in RDR2 on Vulkan was work flawless and very smooth

Looking at RTX 5xxx results, and want asking: Lisa WHY?

 

[Win2012R2]

want asking: Lisa WHY?

Use limited chiplet assembly capacity for MI300 instead? That's 10x money

 

[eek2121]

It is already tooo high for the non-XT and will be a non-starter. If it is indeed a 7900 XT across the board (raster & RT) and keeps a 240W wattage, the max AMD can ask for is $449, but I would expect that price to drop post launch.
XT - max $529 for MBA card.

The non-XT card's primary competitor is the 4070S, and will likely be around $500 when the 5070 hits the market. The Super tier will co-exist with the 50 series by how it's shaping up.

IF the AMD parts come in performing above "expectations" like we've been seeing leaked, I honestly think AMD might do something "crazy" like 9070 = $529 and 9070 XT = $629 and the "best" AIB boards at $699.

It it has 7900 XTX perf +/- 10% AND reduced power consumption (even ~350W TBP on the highest models would be significant IMO) then... it just makes sense to me.

I just paid $980 after sales tax for a premium 7900 XTX, I'd rather pay $770 after tax and get a more forward looking part to boot.

And if Nvidia moves the line I expect that AMD to keep prices just under the 5070/5070Ti.

nvidia brought almost nothing with the 5xxx series in terms of actual performance so AMD might actually have an opening.

And yeah, I expect them to maximize profits when they can.

I am telling you guys, The leak for the price of $479 for the MBA 9070XT was legit. The only way it won’t come in at that price is if it absolutely kills the 5070 series, even then, AMD was this launch to go off without a hitch. IIRC the launch even t is Jan 22nd. We will find out soon.

 

[adroc_thurston]

Use limited chiplet assembly capacity for MI300 instead? That's 10x money

It's not limited. They just saw GB202 and kinda gave up.

 

[Win2012R2]

It's not limited. They just saw GB202 and kinda gave up.

Hope somebody gets fired for that losing attitude.

 

[reaperrr3]

Hope somebody gets fired for that losing attitude.

If what adroc said is true, I doubt Lisa will fire herself over this...

 

[SolidQ]

I doubt Lisa will fire herself over this...

I think someone from Radeon department whispered in her ear that we wouldn't be able to do anything to GB202, and she made her decision.

 

[Novacius]

It's not limited. They just saw GB202 and kinda gave up.

I don't really get that. While GB202 might be out of reach, they could have made a GPU that is 1.5x N48, even after killing off chiplet RDNA4. The rest of the Blackwell looks mighty disappointing. The 5080 is only 15% faster in rasterization than the 4080, which means it's only 10-15% faster than the 9070 XT potentially. 1.5x N48 would've killed that thing and could've delivered ~4090 performance for $999. No need to compete with GB202. On the other hand I get that AMD's time, money and talent is limited, so it probably makes sense on some level. Sigh.

 

[blckgrffn]

It makes you wonder what the sales pyramid looks like in terms of volume/profit. If C level folks are doing their job and listening to their MBA hires ( ) we can expect that there was more than gut feel to the decision.

How many GPUs are sold via OEM and Pre-Built? Most?

How many of them are going to be 4090/5090 class? By volume? Not many?

I think the most prolific prebuilt "premium" card, in rigs north of $1k, has got to be 4070/4070S.

So if you are AMD and thinking how can we sell 80% of the volume of GPUs but cut off a huge amount of time and work on what is likely the riskiest segment (the doom guy chainsaw) then I guess it makes sense.

The GPU AMD needs to kick in the balls to be relevant is the 3060/4060 tier card. I'll be interested to see how that plays out. Upping the game there is a tide that will lift all boats in terms of PCMR.

I GET THAT HALO INFLUENCES MANY PURCHASERS ALL THE WAY DOWN THE STACK so you can save that if you really feel the need to type it out. It doesn't influence everyone or everything, though.

 

[gdansk]

Most probably AMD didn't expect Nvidia would leave a $1000 hole in their lineup.

 

[gaav87]

I don't really get that. While GB202 might be out of reach, they could have made a GPU that is 1.5x N48, even after killing off chiplet RDNA4. The rest of the Blackwell looks mighty disappointing. The 5080 is only 15% faster in rasterization than the 4080, which means it's only 10-15% faster than the 9070 XT potentially. 1.5x N48 would've killed that thing and could've delivered ~4090 performance for $999. No need to compete with GB202. On the other hand I get that AMD's time, money and talent is limited, so it probably makes sense on some level. Sigh.

This aged great xD

 

[Heartbreaker]

I think someone from Radeon department whispered in her ear that we wouldn't be able to do anything to GB202, and she made her decision.

IMO, 5090 barely barely makes sense at NVidia price and volume. It would make no sense at all for AMD to chase it with the volume and price they would have received.

 

[adroc_thurston]

I think someone from Radeon department whispered in her ear that we wouldn't be able to do anything to GB202, and she made her decision.

it's just normal program review stuff.

The 5080 is only 15% faster in rasterization than the 4080, which means it's only 10-15% faster than the 9070 XT potentially. 1.5x N48 would've killed that thing and could've delivered ~4090 performance for $999

dats da plan for RDNA5! one more die for the kilobuck range.

While GB202 might be out of reach, they could have made a GPU that is 1.5x N48, even after killing off chiplet RDNA4.

when you build omega chungus, you always go to the top.

No need to compete with GB202

there always is a need.

The GPU AMD needs to kick in the balls to be relevant is the 3060/4060 tier card.

oh nyo that segment is dying off. ASPs are creeping up.

IMO, 5090 barely barely makes sense at NVidia price and volume. It would make no sense at all for AMD to chase it with the volume and price they would have received.

not the product itself that drives volume, it amplifies it down the stack. halo effect.