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]

That might be their only revenue left if Tenstorrent, Cerberas and Gaudi destroy CUDA overnight

None of them have any chances against NV.

 

[maddie]

see that's a problem.
You want them to lose and stay poor.
That's bad.

You want them to win and we get poor.

 

[Win2012R2]

They aren't losing.

AMD GPU line is not good enough to sustain company

 

[igor_kavinski]

AMD GPU line is not good enough to sustain company

Creating space heaters is not the solution though and AMD management agrees. Hopefully we see real innovation with UDNA.

 

[Timorous]

Creating space heaters is not the solution though and AMD management agrees. Hopefully we see real innovation with UDNA.

It is a solution if they think people will buy them. May not be the best solution though.

Also the creation of chungus parts does not mean smaller parts can't also exist. If AMD get chiplet GPUs working with multiple compute dies it allows for a lot more scalability than they currently have so would give them the ability to make small N44 like parts all the way to chungus 800mm+ parts all with the same basic building blocks.

 

[Josh128]

heheh dumb mericans with their crappy power

FSR4 has helped AMD's perception far more then i expected. A Halo card would do well now if they had one.......
i have a 7900xtx i would have bought a halo rdna4, now im waiting /hoping to get a PSSR/RDNA3 target'd FSR4.

edit: if you want to learn how to do power properly come to Australia, i shudder every time i go to the US of A and have to touch a power point.

Your prejudice is a bit misplaced. 50Hz power and 50Hz refresh on CRTs suck, USA has you there. Also 120V is both less dangerous and less finicky maintaining <6mA GFCI safety standards without nuisance tripping.

 

[gdansk]

If you care about efficiency you can also cut N48 down to where it is most efficient. It's quite a bit less than 300W. But you can't set an RTX 4090 below 300W without changing the vBIOS power limits. Or at least I cannot with a reference Founders Edition. 300W is the lower limit. Where as that's on the high end for N48.

 

[Mopetar]

Monetary inflation has next to nothing to do with the current GPU pricing problem.

Like most issues, the causes are multifaceted and blaming any one thing for the entirety is shortsighted and even if removed would not fix the problem.

Since 2020, the US dollar has experienced a bit under 25% inflation cumulatively. In other words the same GPU that cost $200 inflation cumulatively 2020 would be $250 in today's dollars due to inflation alone.

Obviously it does not explain the higher costs alone, but ignoring it leaves out a substantial part of the cost increases.

The largest factor is that new nodes are seeing price scaling that's close to in line with the increases in transistor density. In the past, node costs tended to scale linearly whereas density gains were exponential, leading to decreasing prices even as GPUs had more and better transistors.

There's also a greater demand for wafers to be allocated to other, more profitable uses, leading to reduced supply. Naturally these also leads to price increases as there was not an equal reduction in demand.

All of these (and other) factors are contributing to higher prices. If it were simply tariffs (or some other single cause explanation) you could look at pricing differences across countries to discriminate between those causes. However, the prices are higher everywhere, which seems to contradict a simple single cause explanation like tariffs that only the USA is imposing.

Recently Sony raised prices on their PlayStation consoles in much of the world, but not the USA: https://blog.playstation.com/2025/04/13/ps5-price-to-rise-in-europe-australia-and-new-zealand/. The reasons cited were inflation and currency exchange rates. Microsoft also raised Xbox prices globally as well. Tariffs in the USA do not explain global price increases. Monetary inflation (which is of a similar rate in Europe as the USA over the last ~5 years) does, at least in part. A regression analysis could better determine what part (if any identified factors) of price increases can be explained by inflation, but it's obviously not zero.

 

[gdansk]

Obviously it does not explain the higher costs alone, but ignoring it leaves out a substantial part of the cost increases.

No, in this case it can be entirely ignored. GPUs were reasonably priced in November, 2024. GPUs were no longer reasonably priced by the end of January, 2025. Monetary inflation has negligible contribution (<0.5% over the time period) to the current pricing problem (fake MSRPs).

Perhaps for those complaining about the MSRPs being too high or the performance increases being too small then one could blame monetary inflation and transistor costs. But for the most part people who are building new PCs or upgrading from older cards would be happy to buy a $750 RTX 5070 Ti or $600 RX 9070 XT. And AMD and Nvidia said those parts should exist.

What happened is a supply crunch, new taxes and perhaps some increased demand (from panic buying and those who were waiting for one more generation before buying).

 

[Mopetar]

No, in this case it can be entirely ignored. GPUs were reasonably priced in November, 2024.

I would not agree with that assessment in the first place, but it ignores a different reality of companies trying to clear remaining inventory of previous generation products and consumers delaying purchases in anticipation of next generation products.

And of course people would be happy to pay an MSRP that many still consider to be too high for the simple fact that it's several hundred dollars below actual market prices. Others who aren't willing to do so have to sit this generation out and hope that the market situation changes or for everyone willing to pay the higher prices to buy their cards first so that the market rate falls to MSRP.

 

[gdansk]

I would not agree with that assessment in the first place, but it ignores a different reality of companies trying to clear remaining inventory of previous generation products and consumers delaying purchases in anticipation of next generation products.

And of course people would be happy to pay an MSRP that many still consider to be too high for the simple fact that it's several hundred dollars below actual market prices. Others who aren't willing to do so have to sit this generation out and hope that the market situation changes or for everyone willing to pay the higher prices to buy their cards first so that the market rate falls to MSRP.

If you think the MSRPs are too high then there is much to blame on monetary inflation and transistor costs.
But the current pricing problem is fake MSRPs. Not high MSRPs like the 4080 and 7900 XT.

Edit: clarified

 

[marees]

If you think the MSRPs are too high then there is little to blame but monetary inflation and transistor costs.
But the current pricing problem is fake MSRPs. Not high MSRPs like the 4080 and 7900 XT.

Part of the problem is something got seriously wrong in Nvidia's supply chain. They stopped 4xxx production way back but didn't replenish the stocks with equivalent 5xxx stuff

If I were AMD I would have launched 9060xt 1 year ago (for $400), knowing that Nvidia's competition in xx60 usually launches in Q2 thus giving a full 1 year headstart to Navi 44

Instead, AMD seems to prefer the marketing win of being able to compare 16gb 9060xt vs 8gb 5060 ti at 1440p

 

[Abwx]

Your prejudice is a bit misplaced. 50Hz power and 50Hz refresh on CRTs suck, USA has you there. Also 120V is both less dangerous and less finicky maintaining <6mA GFCI safety standards without nuisance tripping.

With 2x the voltage the same cables can supply 4x the power at same losses, also in PSUs the rectifying diodes on the primary are subject to 2x lower current for a same output power, and same for the primay switching mosfets, since diodes and mosfets losses increase exponentialy with the current a same PSU has quite better efficency when used in Europe compared to the US.

 

[igor_kavinski]

What happened is a supply crunch, new taxes and perhaps some increased demand (from panic buying and those who were waiting for one more generation before buying).

Also I think lots of "clever" non-humans who buy multiple GPUs and sell the extra at inflated prices to essentially get a "free" GPU.

The accommodation scene in UAE is a mess due to this very reason. Pricks rent a flat, rent out the individual rooms and then live rent-free in a different flat of their own.

 

[Mopetar]

If you think the MSRPs are too high then there is much to blame on monetary inflation and transistor costs.
But the current pricing problem is fake MSRPs. Not high MSRPs like the 4080 and 7900 XT.

One being too high and far worse does not preclude the other from also being true. Is the situation any better if AMD/NV changed it so MSRP was the same as the current market rate? Would it change the underlying causes for the higher prices by not being a "fake" MSRP?

Part of the problem is something got seriously wrong in Nvidia's supply chain.

Their supply chain is fine. Gamers aren't willing to pay as much as AI companies so guess who gets chips first.

If I were AMD I would have launched 9060xt 1 year ago (for $400), knowing that Nvidia's competition in xx60 usually launches in Q2 thus giving a full 1 year headstart to Navi 44

You should apply for a job at AMD. I'm sure they'll be impressed with your perfect hindsight. Not that you would need to as having such perfect insight into what Nvidia will do is more than enough to make a killing in the stock market.

 

[Thunder 57]

Like most issues, the causes are multifaceted and blaming any one thing for the entirety is shortsighted and even if removed would not fix the problem.

Since 2020, the US dollar has experienced a bit under 25% inflation cumulatively. In other words the same GPU that cost $200 inflation cumulatively 2020 would be $250 in today's dollars due to inflation alone.

Obviously it does not explain the higher costs alone, but ignoring it leaves out a substantial part of the cost increases.

The largest factor is that new nodes are seeing price scaling that's close to in line with the increases in transistor density. In the past, node costs tended to scale linearly whereas density gains were exponential, leading to decreasing prices even as GPUs had more and better transistors.

There's also a greater demand for wafers to be allocated to other, more profitable uses, leading to reduced supply. Naturally these also leads to price increases as there was not an equal reduction in demand.

All of these (and other) factors are contributing to higher prices. If it were simply tariffs (or some other single cause explanation) you could look at pricing differences across countries to discriminate between those causes. However, the prices are higher everywhere, which seems to contradict a simple single cause explanation like tariffs that only the USA is imposing.

Recently Sony raised prices on their PlayStation consoles in much of the world, but not the USA: https://blog.playstation.com/2025/04/13/ps5-price-to-rise-in-europe-australia-and-new-zealand/. The reasons cited were inflation and currency exchange rates. Microsoft also raised Xbox prices globally as well. Tariffs in the USA do not explain global price increases. Monetary inflation (which is of a similar rate in Europe as the USA over the last ~5 years) does, at least in part. A regression analysis could better determine what part (if any identified factors) of price increases can be explained by inflation, but it's obviously not zero.

Computer components have almost always gone down in price throughout history. You can try to blame it on inflation or nodes but other components have no problem maintaining prices. What you got right is supply. Nvidia and AMD would much rather devote finite fab allocation to higher margin profits so GPU's are essentialy at the bottom when it comes to priority.

Part of the problem is something got seriously wrong in Nvidia's supply chain. They stopped 4xxx production way back but didn't replenish the stocks with equivalent 5xxx stuff

If I were AMD I would have launched 9060xt 1 year ago (for $400), knowing that Nvidia's competition in xx60 usually launches in Q2 thus giving a full 1 year headstart to Navi 44

Instead, AMD seems to prefer the marketing win of being able to compare 16gb 9060xt vs 8gb 5060 ti at 1440p

You can't launch something that isn't ready. Or if you do not have enough stock.

 

[gdansk]

One being too high and far worse does not preclude the other from also being true.

It doesn't. But there we are already trending toward MSRP. Monetary inflation and transistor costs didn't suddenly start decreasing. Supply did start increasing. It also didn't impact other logic silicon suddenly in late 2024/early 2025. Phones that were $1000 are still $1000 (or less). CPUs are in stock or intermittenly in stock but near MSRP or less. If monetary inflation and transistor costs are the issue, then why not these too?

A recent example of price increases are flash memory products over the last 2 years. How did they achieve this large price increase? Industry-wide supply cuts.

And that's most similar to what happened to graphics cards in late 2024/early 2025. Industry-wide supply evaporated. But, unlike Micron-Hynix-Samsung, Nvidia-AMD weren't bleeding money when they did this.

 

[inquiss]

Part of the problem is something got seriously wrong in Nvidia's supply chain. They stopped 4xxx production way back but didn't replenish the stocks with equivalent 5xxx stuff

If I were AMD I would have launched 9060xt 1 year ago (for $400), knowing that Nvidia's competition in xx60 usually launches in Q2 thus giving a full 1 year headstart to Navi 44

Instead, AMD seems to prefer the marketing win of being able to compare 16gb 9060xt vs 8gb 5060 ti at 1440p

I would have launched it 5 years ago. Would have been easier to beat Nvidia then.

 

[inquiss]

Computer components have almost always gone down in price throughout history. You can try to blame it on inflation or nodes but other components have no problem maintaining prices. What you got right is supply. Nvidia and AMD would much rather devote finite fab allocation to higher margin profits so GPU's are essentialy at the bottom when it comes to priority.



You can't launch something that isn't ready. Or if you do not have enough stock.

There other factors here. I agree supply is a big, big part of it. Nvidia didn't show up, so the market doesn't have enough supply and AMD couldn't have predicted that. It would have been insane for them to think they should gamble unsold inventory on the hope Nvidia does something unprecedented and make enough to supply the market from the position they were in, but...

..cost per transistor is now not going down in like with moores law, instead costs per transistor are going up. We can't rely on old scaling laws anymore to reliably get future hardware cheaper in the way you describe it in that relevant paragraph.

Everyone is right.

 

[jpiniero]

It doesn't. But there we are already trending toward MSRP. Monetary inflation and transistor costs didn't suddenly start decreasing. Supply did start increasing. It also didn't impact other logic silicon suddenly in late 2024/early 2025. Phones that were $1000 are still $1000 (or less). CPUs are in stock or intermittenly in stock but near MSRP or less. If monetary inflation and transistor costs are the issue, then why not these too?

You have to be looking at nVidia's prices over time too, not AMD Fire Sale ones.

 

[Thunder 57]

There other factors here. I agree supply is a big, big part of it. Nvidia didn't show up, so the market doesn't have enough supply and AMD couldn't have predicted that. It would have been insane for them to think they should gamble unsold inventory on the hope Nvidia does something unprecedented and make enough to supply the market from the position they were in, but...

..cost per transistor is now not going down in like with moores law, instead costs per transistor are going up. We can't rely on old scaling laws anymore to reliably get future hardware cheaper in the way you describe it in that relevant paragraph.

Everyone is right.

It really is a supply thing though. Look at DRAM and NAND prices when they overproduce, they go down. When they collude (as we have seen particularly with DRAM), prices go up.

 

[adroc_thurston]

Look at DRAM and NAND prices when they overproduce, they go down. When they collude (as we have seen particularly with DRAM), prices go up.

DRAM and NAND are commodities, they're not bespoke designs targeted at a specific market.
Super different market forces at play, basically.

 

[Thunder 57]

DRAM and NAND are commodities, they're not bespoke designs targeted at a specific market.
Super different market forces at play, basically.

CPU's then. Even models in high demand are in stock at MSRP from time to time. GPU's are clearly the exception. Nvidia and AMD can get higher margins on pretty much anything else so they hardly seem to try.

 

[adroc_thurston]

CPU's then.

Did you forget the Great Laptop CPU Crunch of 2021?

GPU's are clearly the exception.

No, it's normal, they're pretty complex devices with long-tail supply chains.

 

[inquiss]

It really is a supply thing though. Look at DRAM and NAND prices when they overproduce, they go down. When they collude (as we have seen particularly with DRAM), prices go up.

Right, supply and demand. I agreed. However costs also have been going up too, so if you were to establish supply, msrps would be going up because costs are going up too. That's part of the supply bit.