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

 

[adroc_thurston]

Another Medium Die

Medium is Premium as they say.

 

[RnR_au]

First let's see RDNA4 before even thinking about the next hype cycle. lol.

You are still new here, so you haven't learned how to AMD.

You'll get there.

 

[soresu]

You are still new here, so you haven't learned how to AMD.

You'll get there.

Yeah Papermaster ignited the Zen5 hype train before even Zen3 got announced.

 

[SolidQ]

Another Medium Die

Full google translate/ Big core

PS6 X3D?

 

[gaav87]

Another Medium Die

I heard 3 or 4 are planed small small medium big.
Also XE3 is chiplet there is code in recent kernel.

 

[carrotmania]

Agricultural Enterprises

 

[del42sa]

I heard 3 or 4 are planed small small medium big.
Also XE3 is chiplet there is code in recent kernel.

extra small, medium small, regularly small, fatt small

 

[coercitiv]

I heard 3 or 4 are planed small small medium big.
Also XE3 is chiplet there is code in recent kernel.

AMD can't even launch their medium die today, they are paralyzed at the executive level. Plans are easy, the engineers know what to do, but watch as the leadership chickens out again.

I know it looks overly negative, but this is my conclusion, crystalized over several design and release cycles: AMD execs haven't found a way through Nvidia's mindshare moat, and now crippling doubt has set in. They only eat after the alpha is done.

 

[adroc_thurston]

AMD execs haven't found a way through Nvidia's mindshare moat

Well yeah, it's been 16 years since Radeon committed brand suicide. Not much they can do, really.

 

[gaav87]

Well they could do number of stupid things right now:
1. Rebrand rx9070xt/non-xt as rx9080 and rx 9070 with increased price
2. Leave the 9070xt naming and still increase the price
3. We heard you GUYS we love gamers, we decided not to rename our gpus (crowd cheers) here is 9800xt 649$ xD

 

[gaav87]

After thinking for, a while its briliant strategy!
Leave all 9070 non-xt alone
Recall all 9070xt for bios change (id change + oc)
Swap the name to 9080
Compare 9080 to 4080(non super) on graphs
Show RE4 and Horizon before 5080 reviews launch
Price that at 499$ MBA That would be a WIN

Remember, profanity is not allowed in the tech section, even if misspelled. Thanks, -Moderator Shmee

 

[RnR_au]

Agricultural Enterprises

FWIW.... here is the story about this Chinese nickname for AMD;


Source

 

[Tuna-Fish]

Who needs labels when it's green box versus red box, and the red box is winning by 80% of the graph height thanks to a special driver from the company that insists on using red hardware while dissing it at every corner.

Note that they are in fact praising red hardware at every corner. They are just dissing the software.

Both parts of that are part of their marketing strategy, they intend to sell their boxes to people on the idea that they are the only ones who can use the most cost-efficient hardware with the good drivers.

I am extremely sceptical of that latest leak, mainly because of the last line:

Sony will also stack 3D on the console side, but Microsoft is not sure.

There is literally no individual inside AMD who knows both what Sony is doing and what MS is doing. The teams are completely separate, work in separate offices and are not allowed to have any communication with each other. For this random person to know what they are both doing would require them to have multiple sources, which is much less likely than that they just made it up.

 

[jpiniero]

N3E for Zen6 and UDNA seems very suspicious and unlikely. N3P would already be 2 years in HVM by the time they arrive, and the price adjusted node for TSMC as they ramp up N2.

N2 is mainly going to be used for it's power savings, and by customers willing to pay dearly for that. The latter sure doesn't sound like AMD's dGPU business, lol.

I'd expect Rubin Gaming to use N3P or X too btw.

 

[Keller_TT]

N2 is mainly going to be used for it's power savings, and by customers willing to pay dearly for that. The latter sure doesn't sound like AMD's dGPU business, lol.

I'd expect Rubin Gaming to use N3P or X too btw.

I meant they'll be using N3P, not N3E.
TSMC N2 didn't hit targets and they had issues with GAA and also process, yield issues, and so even Apple has delayed its 2nm plans by a year I guess.

AMD Server chips could use N3P and N3X combination, I suppose.

 

[Keller_TT]

FWIW.... here is the story about this Chinese nickname for AMD;
View attachment 115102

Source

Great find. And how apt! 🤣

 

[Win2012R2]

I'd expect Rubin Gaming to use N3P or X too btw.

What's the source that Rubin Gaming is even a thing? Nvidia is pushing HBM versions to be out in Q3 or Q4 2025, feels like that will be Volta without any gaming, they'll certainly use max of N3P manufacturing for it, but given how poor 50 series are one wonder if they'd have to push out gaming product also in 2026

 

[Jan Olšan]

It's maybe reality in Mindfactory but that's it. If you really think that's the case in most shops all over the world, I don't know what to say really.

The reactions to the Mindfactory data are always so weird...

People should understand the difference between the whole market that is largely prebuilts (similar to how notebook market works) from large companies a la HP, Acer.
Retail is sometimes estimated to be 5% of desktop market.

However, retail market is very different. In retail DIY-built PCs, people buy their own components, so basically retail DIY market is much more reflective of preferences of customers - the customers that self-build. Or that build PCs for others, small-scale.

You can't shout and point that mindfactory doesn't match JPR and not be silly, because that is mistaking retail for whole (mainly OEM) market. It's also kinda silly to assume conspiracies where AMD turbofans collude online to only buy at Mindfactory to fake stats... whereas Nvidia turbofans somehow don't.

The takeway you should notice is instead the fact that retail/DIY marketshare of Radeon is much bigger than overall marketshare which is being pulled down by OEM. And ask why. It all points at that it's because in OEM, the GPU choice is really pushed at you by the PC maker, you don't pick it. Apparently the choice the customer may make can easilly differ from what the maker has chosen. GOSH hoe surprising, who woulda thunk wew

It's really like with CPUs, but there people don't seem to dispute AMD's retail popularity by pointing at their overal destkop share being 20 %.

 

[ToTTenTranz]

If AMD had something above the 9070 XT you can bet Nvidia would have a different lineup with various price adjustments.

The 5090 was always going to be $2K or more.


A 32GB GPU that runs LLMs super fast to deploy 14B chatbots? All the universities and a bunch of enterprises will be snatching those as soon as they can. The alternative is a slower RTX 6000 Ada 48GB for $8K.

Full google translate/ Big core

PS6 X3D?

I mean.. if the dude's aunt said it, then it should be true.

 

[coercitiv]

The reactions to the Mindfactory data are always so weird...

People should understand the difference between the whole market that is largely prebuilts (similar to how notebook market works) from large companies a la HP, Acer.
Retail is sometimes estimated to be 5% of desktop market.

However, retail market is very different. In retail DIY-built PCs, people buy their own components, so basically retail DIY market is much more reflective of preferences of customers - the customers that self-build. Or that build PCs for others, small-scale.

You can't shout and point that mindfactory doesn't match JPR and not be silly, because that is mistaking retail for whole (mainly OEM) market. It's also kinda silly to assume conspiracies where AMD turbofans collude online to only buy at Mindfactory to fake stats... whereas Nvidia turbofans somehow don't.

The takeway you should notice is instead the fact that retail/DIY marketshare of Radeon is much bigger than overall marketshare which is being pulled down by OEM. And ask why. It all points at that it's because in OEM, the GPU choice is really pushed at you by the PC maker, you don't pick it. Apparently the choice the customer may make can easilly differ from what the maker has chosen. GOSH hoe surprising, who woulda thunk wew

It's really like with CPUs, but there people don't seem to dispute AMD's retail popularity by pointing at their overal destkop share being 20 %.

To offer even more context to the post above, a screenshot from Amazon:

 

[Josh128]

Yeah Papermaster ignited the Zen5 hype train before even Zen3 got announced.

That was Mike Clark, and it was 1 year before Zen 4 got released.

 

[Keller_TT]

To offer even more context to the post above, a screenshot from Amazon:

View attachment 115104

Both thy posts pick up Strawman arguments. AMD simply missed the boat with RDNA3, but:
1. AMD is more popular in some markets (countries that have a higher MSRP than USD MSRP before adding a dose of high VAT - Germany, some Asian countries, South America, etc.)
2. Pt. 1 is exacerbated by NV's high pricing upon launch. $500 card would cost $630-650 in those places. In Germany, it's grudgingly okayish, but not in other markets.
3. Prices drop significantly within a year, and then people pick up their new NV cards.

Was a 4070 worth $700 or more? Was it that useful for RT? Was 7800XT better for 1440P native raster? So, that was the only relevant AMD card. Then 7600XT did the same vs 4060 for the entry tier.

 

[Heartbreaker]

Amazon Canada GPU Best sellers. I'm leaning more toward the Mindfactory being something of outlier, because it's looked like this every time I checked Amazon Canada. Plus this lines up better with Steam Data as well. With RX 580 and RX 6600 being AMDs most popular GPUs in the survey (behind about 30 NVidia GPUs).

Spoiler: Amazon GPU best sellers Canada

Also I believe this the Amazon Germany and UK, GPU best seller page, and it looks very much like North American ones:

Amazon.de Best Sellers: The most popular items in Graphics Cards

Amazon.de Best Sellers: The most popular items in Graphics Cards

www.amazon.de

Amazon.co.uk Best Sellers: The most popular items in Graphics Cards

 

[Tup3x]

The reactions to the Mindfactory data are always so weird...

People should understand the difference between the whole market that is largely prebuilts (similar to how notebook market works) from large companies a la HP, Acer.
Retail is sometimes estimated to be 5% of desktop market.

However, retail market is very different. In retail DIY-built PCs, people buy their own components, so basically retail DIY market is much more reflective of preferences of customers - the customers that self-build. Or that build PCs for others, small-scale.

You can't shout and point that mindfactory doesn't match JPR and not be silly, because that is mistaking retail for whole (mainly OEM) market. It's also kinda silly to assume conspiracies where AMD turbofans collude online to only buy at Mindfactory to fake stats... whereas Nvidia turbofans somehow don't.

The takeway you should notice is instead the fact that retail/DIY marketshare of Radeon is much bigger than overall marketshare which is being pulled down by OEM. And ask why. It all points at that it's because in OEM, the GPU choice is really pushed at you by the PC maker, you don't pick it. Apparently the choice the customer may make can easilly differ from what the maker has chosen. GOSH hoe surprising, who woulda thunk wew

It's really like with CPUs, but there people don't seem to dispute AMD's retail popularity by pointing at their overal destkop share being 20 %.

You can't assume that one shop in Germany has any correlation with, say, another shop in China. I can check top ten most sold GPUs in one local shop: 10 GeForces, one Radeon (same thing in another shop). It's pretty obvious that at least here in Finland it's very, very unlikely that AMD is selling better or even 50/50 split. Totally different story when it comes to CPUs - only AMD in top ten.

 

[coercitiv]

Both thy posts pick up Strawman arguments.

You didn't even catch the point I was trying to make, not to mention the Strawman part.

The Amazon screenshot was about challenging our preconceptions regarding pricing and SKU of choice among consumers. It said little if anything about AMD market share. Going back to @Jan Olšan 's post, he pointed out the discrepancies between prebuilt and DIY market, in the sense that one should not use data from one of them to extrapolate the results in the other.

Even if we had a perfect snapshot of DIY sales in all major markets around the world, that would not be enough to extrapolate into market share which is primarily driven by OEM sales. The same applies in reverse.