Speculation: Ryzen 4000 series/Zen 3

[Shivansps]

AMD is definitely considering sticking with Vega for APUs. RDNA(2) has caused headaches for them. And the demand for powerful internal graphics is unclear.

I can tell you with certainty they do not have a decision right now. This could be a lategame call.

The demand for powerful internal graphics is unclear? What about the 3400G that now is MIA?

 

[uzzi38]

The demand for powerful internal graphics is unclear? What about the 3400G that now is MIA?

It's not really MIA because of demand as it is supply.

~2 months until Picasso is entirely obsolete.

 

[DrMrLordX]

I respectfully disagree.
By all means rdna2 is ready since its confirmed for consoles and graphics cards. So next gen apu should almost certainly have rdna2 as gpu part.

I do not question its readiness. I question which uarch AMD wants in their APUs moving forward. If they are still working towards hybrid CPU/GPU compute, they'll stick with CDNA.

The demand for powerful internal graphics is unclear? What about the 3400G that now is MIA?

Are you sure the market demands the same things you think it does? You seem to be confusing your own perspective for that of others.

@soresu

I do not think AMD is going to spend the resources to create different APUs based on RDNA and CDNA-family iGPUs. I think they will stick to one design for both the consumer and Pro market. And that design will likely be CDNA. I could be wrong, but if I were in their shoes, I might do the same thing.

What are y'all talking about Zen 4 with Zen 3 coming next first (and this thread being about Zen 3)?

I don't know, someone started talking Zen4 with undue optimism. Zen3's schedule definitely makes those Zen4 wishes seem improbable. At this point there's nothing left to do but hope for some leaks or at least get a launch date from AMD.

 

[maddie]

I do not question its readiness. I question which uarch AMD wants in their APUs moving forward. If they are still working towards hybrid CPU/GPU compute, they'll stick with CDNA.



Are you sure the market demands the same things you think it does? You seem to be confusing your own perspective for that of others.

@soresu

I do not think AMD is going to spend the resources to create different APUs based on RDNA and CDNA-family iGPUs. I think they will stick to one design for both the consumer and Pro market. And that design will likely be CDNA. I could be wrong, but if I were in their shoes, I might do the same thing.



I don't know, someone started talking Zen4 with undue optimism. Zen3's schedule definitely makes those Zen4 wishes seem improbable. At this point there's nothing left to do but hope for some leaks or at least get a launch date from AMD.

Agree with the CDNA in APUs. Should be much more useful for general purpose software. Want to game well? Slap in a GPU, also better for their bottom line overall.

 

[piokos]

I thought we were told it's a whole new family?

I mean the actual properties, not how the CPU is made.
Sure, the architecture may be new, but it only affects core performance as such.
From the leaks we've seen, there is nothing new feature-wise, it'll be similarly efficient, it'll have the same instruction set. Even the IGP will still be Vega.

What I mean is: if the only gain from Zen3 will be 10-15% more performance, they may as well keep the Zen2 lineup and just add a new halo variant on top.
Just shift the prices: 3950X->3900X, 3900X->3800X and so on.
They could save a ton of money and focus more on next 2 generations (which bring a lot of new stuff).

 

[inf64]

What I mean is: if the only gain from Zen3 will be 10-15% more performance, they may as well keep the Zen2 lineup and just add a new halo variant on top.
Just shift the prices: 3950X->3900X, 3900X->3800X and so on.
They could save a ton of money and focus more on next 2 generations (which bring a lot of new stuff).

I think you do not understand how hard is to get 15% IPC improvement, from gen to gen. Take Zen1->Zen2 for example. This change made possible for AMD to completley overtake intel in IPC, until Icelake arrived. intel managed to get similar IPC uplift with their Icelake (vs Skylake) core which was not an easy task at all. Now AMD will overtake intel yet again with their Zen3 core and will likely have higher clocking part vs Willow Cove as well.

Getting 15% IPC is not easy at all, especially given how much thought it takes for new cores to maintain power efficiency when staying at the same or similar node (like Zen2->Zen3).

 

[DrMrLordX]

@inf64

Zen3 isn't going to just be 15% better IPC. It should also bring higher clockspeeds in most circumstances, even if single-core boost won't go up by that much.

 

[moinmoin]

I do not think AMD is going to spend the resources to create different APUs based on RDNA and CDNA-family iGPUs. I think they will stick to one design for both the consumer and Pro market. And that design will likely be CDNA. I could be wrong, but if I were in their shoes, I might do the same thing.

Isn't the CDNA line supposed to eschew all the parts needed for actual graphical rendering for more specialized compute parts? That would obviously make it unsuitable as iGPU. So I'd hope APUs actually see a mix of both.

 

[DrMrLordX]

Isn't the CDNA line supposed to eschew all the parts needed for actual graphical rendering for more specialized compute parts? That would obviously make it unsuitable as iGPU. So I'd hope APUs actually see a mix of both.

I hadn't heard that, though it's possible. I don't think that's true though since I don't see AMD using RDNA2 for workstation video cards.

 

[RetroZombie]

Isn't the CDNA line supposed to eschew all the parts needed for actual graphical rendering for more specialized compute parts? That would obviously make it unsuitable as iGPU. So I'd hope APUs actually see a mix of both.

Agree with the CDNA in APUs. Should be much more useful for general purpose software. Want to game well? Slap in a GPU, also better for their bottom line overall.

I do not think AMD is going to spend the resources to create different APUs based on RDNA and CDNA-family iGPUs. I think they will stick to one design for both the consumer and Pro market. And that design will likely be CDNA. I could be wrong, but if I were in their shoes, I might do the same thing.

Following all your reasons, wouldn't make more sense to integrate/pair cdna with the zen cores and the rdna with the apus?

 

[DrMrLordX]

Following all your reasons, wouldn't make more sense to integrate/pair cdna with the zen cores and the rdna with the apus?

No. AMD won't create that many different APU designs. At least they haven't signaled an interest in doing so. Outside of semi-custom stuff.

 

[RetroZombie]

No. AMD won't create that many different APU designs. At least they haven't signaled an interest in doing so. Outside of semi-custom stuff.

The cdna units even if just a few, would be inside the zen chiplet or even part of the ccx.
If i understand correctly, cdna units lack tmus and display interfaces and maybe api support and other stuff.

 

[DrMrLordX]

If i understand correctly, cdna units lack tmus and display interfaces and maybe api support and other stuff.

Then I don't really know what AMD plans on doing, because I'll go back to what I said a few minutes ago in that I do not expect AMD to use RDNA2 in Pro cards designed for the workstation market. I also don't expect them to integrate CDNA into Zen chiplets. But who knows? I can say we shouldn't expect that in Vermeer, and probably not other Zen3 products like Cezanne.

 

[Shivansps]

It's not really MIA because of demand as it is supply.

~2 months until Picasso is entirely obsolete.

If it were a supply issue the 3200G would be in problems as well, and for now the stock is holding even trought it has to be suffering from a massive increase in demand right now.

Are you sure the market demands the same things you think it does? You seem to be confusing your own perspective for that of others.

The market made the most GPU powerfull APU to disappear, while the smaller one, the 3200G is still avalible. We cant really know why this happened, but we just cant dismiss it. And is really not easy to explain why the 3200G is still avalible.
BTW, the 2400G also dissapeared way before the 2200G did, after they where replaced by Picasso.

-And to give you an idea, since the 3400G is gone, we are selling an averange of 25 3200G Computers, PER DAY- We are not Amazon but thats a lot.

But there is one problem here, faster GPU in APUs will start eating too much into GPU sales and AMD does not want that. If AMD decides to stick with Vega is because they have no competition, not because "just go and buy a dGPU", THAT to me is projecting your own perpective over others.
Also im not sure who said that GPU are only for gaming, but a fair amount of production software that really needs more than 4 cores also uses GPU for acceleration for specific stuff.

 

[amd6502]

If it were a supply issue the 3200G would be in problems as well, and for now the stock is holding even trought it has to be suffering from a massive increase in demand right now.

It's a matter of binning as well as higher than anticipated demand of the 11CU chips. The higher bins previously have gone to mobile ryzen 7 and the 3400g. The 3400g was out of stock but checked amazon today and they now have a few of them again ($150).

Renoir should ease the demand for mobile 12nm Picasso. That means more of the high bin Picasso can now easily go to AM4.

Even so, an Picasso "AF version" of of 2400g with an overclockable iGPU, or a new sku like a 10CU 3290g (with cpu frequencies like that of 3200g) would very much fit the bill for budget 12nm APUs with a capable iGPU.

 

[soresu]

Isn't the CDNA line supposed to eschew all the parts needed for actual graphical rendering for more specialized compute parts? That would obviously make it unsuitable as iGPU. So I'd hope APUs actually see a mix of both.

Yes, AMD specifically outlined that in CDNA the gutted fixed function rasterization hardware would be replaced by tensor units which work super efficiently for ML operations - presumably the new ray intersection and traversal acceleration functionality added in RDNA2 would also be completely absent from CDNA.

They have yet to specifically outline whether CDNA will be wave64 or not, I'm inclined to think it is, at least for now - though that could change in the future, especially given that they are not beholden to console compatibility with CDNA, which I believe has not only shackled GCN/SI uArch development but will continue to do so long past RDNA2, and whatever uArch iteration is used by the mid cycle refresh..

 

[soresu]

The cdna units even if just a few, would be inside the zen chiplet or even part of the ccx.
If i understand correctly, cdna units lack tmus and display interfaces and maybe api support and other stuff.

Despite being more closely termed as a compute accelerator, it's still a completely different architecture from x86 even if it can coherently share the same memory space.

I wouldn't strictly say it's impossible for them to mesh them togehter as a compute uArch superset in the future, but I think it's far more likely that it would remain a separate chiplet on the interposer for quite some time, at best another tightly coupled second layer in a 3D logic stack if thermal problems can be solved (think hundreds of thousands of TSV's at least per layer).

They would probably end up giving it a new name if they fused the uArch's together in such a fashion - at the earliest it would be the next major Zen/CPU core iteration on the far horizon, as Zen3 is to Zen1 according to AMD.

I do wonder just how large a CDNA CU would be minus all that raster HW - I estimate RDNA1 CU's at 5.625 mm2 which is quite a bit larger than a Zen2 core, almost 2x the size, though as a cluster of 2 or 4 inside a whole 8 core CCD that would not matter as much.

 

[Geranium]

This is for the pundits who doesn't follow AMD's APU :
There is no APU with first gen GPU architecture. There is no APU with Terascale1, GCN1 (except for custom console one with 2nd gen feature) or first iteration of Vega. Llano used Terascale2, first GCN APU like Kabini, Kaveri used GCN2, first Zen based APU uses second iteration of Vega architecture.

Following the same tradition there is no RDNA1 based APU. Also CPU/APU and GPU is developed by two separate team lead by two different leader. The CPU/APU teams needs some time to optimise the architecture to function properly in a single die configuration.

 

[DrMrLordX]

The market made the most GPU powerfull APU to disappear, while the smaller one, the 3200G is still avalible. We cant really know why this happened, but we just cant dismiss it.

No offense, but I'm quite prepared to dismiss it. The 3200G and 3400G are some of the least-exciting products in AMD's lineup right now. I don't know how long it will be before AMD finally decides to launch AM4 variants of Renoir as a part of the 4000-series, but I do know that desktop Picasso products are mostly-irrelevant in the grand scheme of things. You see things from the perspective of your shop and your customers which is not everyone else's . . .

 

[TESKATLIPOKA]

Hello everyone!

If CDNA is capable to render games why not, but I am not sure If It's worth It instead of just using RDNA2. With LPPDDR5 I think we will see an Increase in CU by 25-50% -> 10-12CU and considering It's a better architecture compared to Vega I expect a significant increase in performance.

I do wonder just how large a CDNA CU would be minus all that raster HW - I estimate RDNA1 CU's at 5.625 mm2 which is quite a bit larger than a Zen2 core, almost 2x the size, though as a cluster of 2 or 4 inside a whole 8 core CCD that would not matter as much.

One CU can't be so big. 40CU Navi 10 die size is 251mm2 and with your estimate only CU would account for 225mm2. It would leave 26mm2 for everything else and that's impossibly small area for ROPs, controller, ACE, Geomerty engine, L2 cache, Graphic command engine, display, multimedia engine, PCIe4 and so on. CU should be a lot smaller than your estimate.
I found this picture of RDNA dieshot.
Navi 10 dieshot
Original Dieshot

 

[Shivansps]

No offense, but I'm quite prepared to dismiss it. The 3200G and 3400G are some of the least-exciting products in AMD's lineup right now. I don't know how long it will be before AMD finally decides to launch AM4 variants of Renoir as a part of the 4000-series, but I do know that desktop Picasso products are mostly-irrelevant in the grand scheme of things. You see things from the perspective of your shop and your customers which is not everyone else's . . .

Since when the perpective of a single person (that lives in another country) makes the stock of a sku dissapear from stock in Amazon, newegg, etc? And as i said that can be used BOTH ways. Also is not only my perpective here, i have import numbers for my country and Picasso is, by far the top imported CPUs and it has been for months. Motherboards is A320->H310->B450.

But i do agree that Picasso is irrelevant just now with Renoir and Zen 3 coming up. But that dosent explain why AMD understimated the demand.

 

[DrMrLordX]

Since when the perpective of a single person (that lives in another country) makes the stock of a sku dissapear from stock in Amazon, newegg, etc?

It is not that the perspective of one person makes a product disappear. It is that one person acts as though the disappearance actually matters much. It doesn't.

 

[piokos]

I think you do not understand how hard is to get 15% IPC improvement, from gen to gen.

I certainly don't understand, because I'm not a semiconductor engineer. Maybe you are.

But I absolutely believe improving CPUs is very difficult and costly. This is exactly what I'm talking about.

Take Zen1->Zen2 for example. This change made possible for AMD to completley overtake intel in IPC, until Icelake arrived. intel managed to get similar IPC uplift with their Icelake (vs Skylake) core which was not an easy task at all.

Ryzen 3000-series were revolutionary because of the new fab node - making double the core count possible overnight. That's a big change worth investing in.

Zen3 isn't bringing anything like that. It's just a bit faster.
Consumer looks at the performance vs price curve. And increase in core performance is great. But increasing core count is the second best thing.
AMD could keep selling Zen2 at lower prices. It could still give us +15% perf/$.

AMD is leading anyway. Intel is unlikely to catch up this year.
Imagine the massive savings possible if they skipped a generation and focused on future already (DDR5, PCIe5.0, new instructions etc).

 

[Shivansps]

It is not that the perspective of one person makes a product disappear. It is that one person acts as though the disappearance actually matters much. It doesn't.

Yet on Steam the only AMD GPUs doing better than Vega 8 are the RX570 and RX580. And that is desktop only, mobile ones have a diferent name. Combined Vega 11 and Vega 8 is around 0.77% while RX570 is 1.13%. We can argue all day about steam numbers, but the fact remain that you can plug a RX570 on any system and you can only get Vega IGP by owning an APU and that along the fact Intel is absolutely destroying AMD in those same stadistics for CPU usage seems to support that people saying "go and buy a gpu, no one uses those IGP for gaming" is wrong and APUs for gaming are far more important than it looks like in places that buying an RX570 is pocket change. A failure of perpective as you would say.

 

[DrMrLordX]

We can argue all day about steam numbers

Steam's hardware survey is a). unscientific b). useless for determining what is AMD's current product priority.

AMD has forced desktop APUs to take a back seat to nearly every other part of their product lineup since 2017.

People saying "go and buy a gpu, no one uses those IGP for gaming" is wrong.

Nobody said that here. What I'm saying is that AMD is not prioritizing desktop APUs right now. They may never again. Such products are mostly a low-margin afterthought. There isn't much that can be inferred by stock levels of 3400G CPUs. You may as well be reading tea leaves.