Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

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 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[Stuka87]

I never considered that the same thing as an 'APU' since it was 2 entirely separate chips from intel and amd talking together on an EMIB or whatever it was called with dedicated memory. It was more like a new design idea for laptop discrete GPU's than an iGPU.


Yes, this has always been the problem of APU's that AMD understood despite a vocal minority in the communities. Every few months on reddit there would be a "Hey, why doesn't AMD make an APU with HBM attached or dedicated GDDR? man what a home run that would be!!" .... it's more of a novalty fascination with people than a compelling alternative to buying a PC with separate CPU & GPU. An APU locks your cpu/gpu choice together and requires the fastest system RAM possible to get the best performance which will still fall short of mid range dGPUs?

Its more like an SoC than an APU. If you look at say, an Apple M1, it has the CPU, GPU, and unified memory. No external memory at all for CPU or GPU. This makes it super fast, but also unupgradable. The performance would be there, and it would be great for laptops. But a non-starter for desktops. And well, Windows would need a lot of work to properly handle unified memory.

 

[Heartbreaker]

Its more like an SoC than an APU. If you look at say, an Apple M1, it has the CPU, GPU, and unified memory. No external memory at all for CPU or GPU. This makes it super fast, but also unupgradable. The performance would be there, and it would be great for laptops. But a non-starter for desktops. And well, Windows would need a lot of work to properly handle unified memory.

Effectively SoC = APU. These are essentially equivalent, monolithic complete solutions.

Kaby Lake G. Is really just a CPU and discrete GPU (with it's own independent VRAM). There is very little different between this and other laptops with discrete GPUs. The only real technology difference is the memory connection.

 

[Stuka87]

Effectively SoC = APU. These are essentially equivalent, monolithic complete solutions.

Kaby Lake G. Is really just a CPU and discrete GPU (with it's own independent VRAM). There is very little different between this and other laptops with discrete GPUs. The only real technology difference is the memory connection.

Kind of. APU's don't have onboard memory. They go outside the package to access RAM. An SoC has any required RAM onboard the package.

 

[Heartbreaker]

Kind of. APU's don't have onboard memory. They go outside the package to access RAM. An SoC has any required RAM onboard the package.

Completely Disagree.

Putting devices on their own small PCB and connecting that to a larger PCB is no real difference than just mounting them to larger PCB in the first place. It's just a VERY minor packaging difference.

Thus Kaby-G is no difference than just having a separated CPU with it's own memory, and PCIe connected dGPU with it's own memory. Because that is exactly what both are.

Likewise. An APU/SoC on MB, with one shared memory pool with LPPDDR soldered next to on the motherboard is really no different than an APU/SoC on it's own tiny PCB, with one shared memory pool with LPPDDR soldered next to it, is no different.

Kaby-G is Nothing like an APU/SoC, and completely like normal laptop CPU and PCIe connected dGPU combo.

 

[Glo.]

Completely Disagree.

Putting devices on their own small PCB and connecting that to a larger PCB is no real difference than just mounting them to larger PCB in the first place. It's just a VERY minor packaging difference.

Thus Kaby-G is no difference than just having a separated CPU with it's own memory, and PCIe connected dGPU with it's own memory. Because that is exactly what both are.

Likewise. An APU/SoC on MB, with one shared memory pool with LPPDDR soldered next to on the motherboard is really no different than an APU/SoC on it's own tiny PCB, with one shared memory pool with LPPDDR soldered next to it, is no different.

Kaby-G is Nothing like an APU/SoC, and completely like normal laptop CPU and PCIe connected dGPU combo.

It was 100% separate CPU and dGPU that was on single package. It never was an APU. Its still MCM connection. Its NOT an APU, its seperate CPU and GPU.

 

[Glo.]

Yes, so AMD should overbuild all their APUs with overkill iGPUs that the majority of its customers do not need or want...in case Intel makes an overbuilt tiled product with an overkill iGPU that the majority of its customers do not need or want. Great business strategy.

The last time Intel and AMD made an APU with a ton of hardware dedicated for iGPU was Kaby Lake-G. Notice that neither AMD nor Intel have been anxious to produce spiritual successors to that product, for good reason.

Kaby Lake-G was always an MCM product. CPU + dGPU on single package. NEVER an APU.

The whole point of an APU is to combine CPU and GPU to be unified memory architecture. Kaby Lake - G wasn't one.

 

[jpiniero]

The whole point of an APU is to combine CPU and GPU to be unified memory architecture. Kaby Lake - G wasn't one.

The problem is where you:

A.) Get the memory bandwidth
B.) Have enough of a cost reduction to make it worthwhile over just including some skimpy IGP and having an external dGPU

Fully knowing that the OEMs will mostly want the skimpy IGP.

 

[eek2121]

The Steam Deck 2 would be a prime candidate for an 8 core CPU/24 CU GPU (responding to comments I read about it the past few pages, have not had much free time).

When unplugged, the GPU can simply run at lower clocks to save power. When plugged in, clocks can boost to allow higher resolutions and refresh rates. One of the few complaints about the Steam Deck was that it doesn’t scale performance when plugged in.

Just a thought.

I used the best estimate I could find. Industry sources in the know have disputed the claims of 7nm wafers costing $10K as being way too high. Maybe in the early days but its been three years and 6nm is even cheaper to make.

You are overpricing by several thousand dollars. N6 was significantly cheaper than N5 as of 6 months ago. A customer like AMD would pay somewhere around $8,000-$9,000 for N7 (note this was $7,000-$8,000 in 2019) based on the numbers I have seen. TSMC made N6 cheaper because of less machine time involved which leads to higher volume. The early numbers I heard for N6 were around $4,000-$5,000, but that was before supply chains blew up. The real (post supply chain issues) number is likely somewhere between $5,000-$7,000. TSMC really wants everyone to transition from N7 to N6 because they can output more wafers per month, which leads to more revenue.

With the economy struggling, those prices will possibly even drop a bit.

Note that most of the numbers I have referenced above came from various leaks in 2018/2019 and a few from last year. I don’t have access to a price sheet or anything, but the sources that provided the numbers were reliable ones.

 

[Glo.]

The problem is where you:

A.) Get the memory bandwidth
B.) Have enough of a cost reduction to make it worthwhile over just including some skimpy IGP and having an external dGPU

Fully knowing that the OEMs will mostly want the skimpy IGP.

OEMs want: lowest possible manufacturing costs, simplest possible engineering costs.

That is the brutal truth. OEMs will happily take powerful iGPU if it will make them not have to include dGPU. They will take APU over CPU+dGPU combo in a hearthbeat if the manufacturing costs, design costs, engineering costs will be lower than just buying CPU+dGPU combo.

2) To get high enough memory bandwidth you have few choices.

You integrate system level caches, you increase memory bandwidth by using GDDR6 memory, or you widen the memory controller bus, and go from 128 to 192 or 256 bit bus, or you use faster memory.

 

[jpiniero]

You integrate system level caches, you increase memory bandwidth by using GDDR6 memory, or you widen the memory controller bus, and go from 128 to 192 or 256 bit bus, or you use faster memory.

But that would drive up the cost for the majority of sales that wouldn't need that. Maybe AMD could get to a point where they could split out the memory controllers into their own chiplet, like they do with Navi 31... but that feels like some ways off. Would also need a bigger package too.

 

[Glo.]

But that would drive up the cost for the majority of sales that wouldn't need that. Maybe AMD could get to a point where they could split out the memory controllers into their own chiplet, like they do with Navi 31... but that feels like some ways off. Would also need a bigger package too.

If integration of SLC into the die will not cost the die space too much, you can manage.

Which is why, if the rumors of big.Little are true, I think AMD decided to go for 8P/4E combo with 24 CUs and 32 MB L4 cache.

 

[jpiniero]

If integration of SLC into the die will not cost the die space too much, you can manage.

Which is why, if the rumors of big.Little are true, I think AMD decided to go for 8P/4E combo with 24 CUs and 32 MB L4 cache.

That's going to cost too much if it's an monolithic N3 chip.

 

[Glo.]

That's going to cost too much if it's an monolithic N3 chip.

As always: It depends.

 

[insertcarehere]

Kaby Lake-G was always an MCM product. CPU + dGPU on single package. NEVER an APU.

The whole point of an APU is to combine CPU and GPU to be unified memory architecture. Kaby Lake - G wasn't one.

You're missing the point here, AMD in particular could have made a spiritual successor to Kaby Lake-G that WAS an APU (or more APU-like) any time within the last couple years (i.e big iGPU with special sauce to meet increased bandwidth requirements), they didn't, and one must assume they have a good reason.

OEMs want: lowest possible manufacturing costs, simplest possible engineering costs.

That is the brutal truth. OEMs will happily take powerful iGPU if it will make them not have to include dGPU. They will take APU over CPU+dGPU combo in a hearthbeat if the manufacturing costs, design costs, engineering costs will be lower than just buying CPU+dGPU combo.

2) To get high enough memory bandwidth you have few choices.

You integrate system level caches, you increase memory bandwidth by using GDDR6 memory, or you widen the memory controller bus, and go from 128 to 192 or 256 bit bus, or you use faster memory.

Also missing the point, the choice isn't between powerful iGPU and CPU + dGPU, its between powerful iGPU and skimpy iGPU. A powerful iGPU on die is functionally useless for a lot of laptops so OEMs will not pay more for a chip that essentially costs more to manufacture.

Widening memory controller buses and/or adding dedicated on-chip caches (neither of which scale well with die shrinks by the way) for performance which may not be valued by buyers is stupid. Law Firms and Consultants are not. going. to. pay. more. for the next Thinkpad Carbon just because the chip has a big iGPU + 64MB SLC to feed said GPU.

 

[Heartbreaker]

OEMs want: lowest possible manufacturing costs, simplest possible engineering costs.

That is the brutal truth. OEMs will happily take powerful iGPU if it will make them not have to include dGPU. They will take APU over CPU+dGPU combo in a hearthbeat if the manufacturing costs, design costs, engineering costs will be lower than just buying CPU+dGPU combo.

AMD want: lowest possible cost APU that is competitive with Intel.

That is the brutal truth.

If any OEM, or group of OEMs want a giant GPU APU, they can commission it.

AMD will build a custom APU for anyone that wants it. They do it for Sony/MS consoles. They do it for Steamdeck. They did it for Subor Z+ console.

Thus far, it appears no OEM is asking for this in the PC market.

 

[Glo.]

AMD want: lowest possible cost APU that is competitive with Intel.

That is the brutal truth.

If any OEM, or group of OEMs want a giant GPU APU, they can commission it.

AMD will build a custom APU for anyone that wants it. They do it for Sony/MS consoles. They do it for Steamdeck. They did it for Subor Z+ console.

Thus far, it appears no OEM is asking for this in the PC market.

Lowest possible cost means not the ultimate lowest possible cost, but a lowest possible cost that gives AMD advantage over their competitors.

10 mm2 more for die size may be enough high cost for AMD to improve something if it will translate into huge performance gains, increase the TAM, and increase volume of sales.

And yes, their customers indeed ask for not only something like Strix Point: 8P/4E/24 CU design, they are asking for even more powerful hardware that is able to compete with M2 Pro and M2 Max.

 

[Glo.]

You're missing the point here, AMD in particular could have made a spiritual successor to Kaby Lake-G that WAS an APU (or more APU-like) any time within the last couple years (i.e big iGPU with special sauce to meet increased bandwidth requirements), they didn't, and one must assume they have a good reason.



Also missing the point, the choice isn't between powerful iGPU and CPU + dGPU, its between powerful iGPU and skimpy iGPU. A powerful iGPU on die is functionally useless for a lot of laptops so OEMs will not pay more for a chip that essentially costs more to manufacture.

Widening memory controller buses and/or adding dedicated on-chip caches (neither of which scale well with die shrinks by the way) for performance which may not be valued by buyers is stupid. Law Firms and Consultants are not. going. to. pay. more. for the next Thinkpad Carbon just because the chip has a big iGPU + 64MB SLC to feed said GPU.

You are moving the goalpost to fit your narrative, how it is impossible for AMD to make powerful APU in todays world, because you do not believe in it.

And the facts are. It was not AMD who created Kaby Lake-G, it was Intel. And spiritual successor to KB-G are MTL and Arrow Lake-P. Fact.

Secondly. Powerful iGPU on die is functionally useful for every use case, especially for higher res experiences. You have to just be able to control its power draw in idle.

I will not go into the discussion of changing software and experience paradigms very soon, which will require powerful integrated solutions. A lot will change(in customers mind, only...) when AR/VR headset from Apple will be released.

https://twitter.com/x/status/1618003945913204736

Something like this is what you will need powerful APUs for. But hey, ITS ONLY A PC WORLD!!!111oneone

 

[Heartbreaker]

Lowest possible cost means not the ultimate lowest possible cost, but a lowest possible cost that gives AMD advantage over their competitors.

On the GPU side,they already have an advantage over Intel, so no need to spend more.

And yes, their customers indeed ask for not only something like Strix Point: 8P/4E/24 CU design, they are asking for even more powerful hardware that is able to compete with M2 Pro and M2 Max.

What you wish for, and what OEMs will pay for are two different things.

If OEMs really wanted that, they would have it buy now. Subor Z+ had 24 CU APU from AMD in 2019, simply because they wanted it enough to pay for it.

 

[jpiniero]

Now AMD could do something like Kaby-G.... put N33 and some GDDR6 on package on one unit.

 

[Heartbreaker]

Now AMD could do something like Kaby-G.... put N33 and some GDDR6 on package on one unit.

Or the industry keeps doing what it has always done. Choose flexibility to pair together CPU and dGPU of their choice.

AMD doesn't have to solder them together to offer package deals.

 

[biostud]

Isn't what your asking for basically the xbox/ps5 SoC? So I don't think there is a problem making them, I just don't think the market is there.

Microsoft could make a Xbox/PC laptop if they wanted. But I guess they don't....

 

[beginner99]

A powerful iGPU on die is functionally useless for a lot of laptops so OEMs will not pay more for a chip that essentially costs more to manufacture.

exactly. And hence economy of scale is missing. if you make multiple chips/apu one with small, medium and large iGPU it is simply not worth it. better to have a cpu with a small iGPU and add a dGPU in case of need.
especially the APU with large iGPU would be niche as it can't be reused on desktop while said dGPU is a repurposed desktop chip.

 

[leoneazzurro]

There is a market for SFF PCs, handheld consoles and thin & light laptops with high battery life. In all these cases a reasonably powerful IGPU APU is the better solution as it allows to play even recent games at low settings/low resolution, having good video playback support while keeping a smaller footprint and a way less complex PCBs/cooling (this is also a cost). With an dGPU you need to reserve additional space for the chip, cooling, memory...
One can say the market is small, but the recent success of the Steam deck and the fact that most of the high end corporate laptops rely on IGPUs say otherwise. Of course, the "reasonably" powerful is in relation with the cost and that means, the amount of area dedicated to IGPU. At the moment, PS5/XBox-like APUs would have small market outside consoles, yes. But, in a couple of years, who knows...

 

[Heartbreaker]

There is a market for SFF PCs, handheld consoles and thin & light laptops with high battery life. In all these cases a reasonably powerful IGPU APU is the better solution as it allows to play even recent games at low settings/low resolution, having good video playback support while keeping a smaller footprint and a way less complex PCBs/cooling (this is also a cost). With an dGPU you need to reserve additional space for the chip, cooling, memory...

No one is saying get rid of APU. AMDs current APUs are powerful enough for that role, and their continued evolution will keep them serving it.

The pushback is merely against the wishful thinking that seems to have an expectation of a sudden giant GPU in the APU. This is simply unrealistic.

One can say the market is small, but the recent success of the Steam deck and the fact that most of the high end corporate laptops rely on IGPUs say otherwise.

Again, no one is against APUs, just the realism of the Big GPU, APU wishful thinking.

Steamdeck is actually a great counter example against the Big GPU APU meme. Steam deck has a custom part, so Valve could have ordered any size GPU in their APU that they wanted, but the Steamdeck has only half the GPU size vs AMDs standard 6800U gpu.

That's a dedicated handheld game machine, and they chose only half the GPU of AMDs standard APU.

Given that it seems unlikely that there is much OEM demand for standard APU to have a much bigger GPU section.

 

[Heartbreaker]

Isn't what your asking for basically the xbox/ps5 SoC? So I don't think there is a problem making them, I just don't think the market is there.

Microsoft could make a Xbox/PC laptop if they wanted. But I guess they don't....

Sadly, because I'd love it if Microsoft put a 50% markup on the XBSX HW, included windows and sold it as a PC. I'd be clicking on the "add to cart" the moment they went on sale. I've never tried to compete for first day purchase of any HW. What makes the XBSX a great console would also make it a great SFF gaming PC, and still a great value as PC with a 50% markup.

So I definitely get the desire for something like this. But I don't let my desire for something, translate into a belief that it will happen. We need to wary of wishful thinking.

The realistic take, is that AMD will just keep making a generic APU for wide usage, ignoring the console like APU niche, unless some OEM asks (pays) for it.