Question AMD Phoenix/Zen 4 APU Speculation and Discussion

[uzzi38]

I can finally make this thread.

https://twitter.com/x/status/1519669375283957760

Phoenix is indeed RDNA3. My advice to everyone: treat the old APU rumours as being out of date.

 

[jamescox]

Okay, I am back after a good Sleep..!

This is very interesting on the Dragon Range side.

View attachment 61039

So Full Fat 64 MiB L3 and 16C/32T on a Mobile Gaming CPU?

If it isn’t out until 2023, then it seems like it could be a stacked device in some manner. I don’t think they would go beyond 16 cores, unless it uses Zen 4c die or something similar. Has there been any rumors on what it actually is? “Dragon Range” is a very odd name.

 

[soresu]

If it isn’t out until 2023, then it seems like it could be a stacked device in some manner. I don’t think they would go beyond 16 cores, unless it uses Zen 4c die or something similar. Has there been any rumors on what it actually is? “Dragon Range” is a very odd name.

I doubt 16C is likely for a gaming APU while game consoles remain 8C max for the PS5/XSX generation.

IMO if Dragon Range differs significantly from the Phoenix hardware spec beyond LPDDR5 -> DDR5 support I imagine that it would be in adding stacked Infinity cache and/or greater GPU WGP count.

The former should be especially helpful for APUs sharing system memory and alleviating the bandwidth constraints this causes.

The LPDDR5 limitation of Phoenix noted on the slide makes me think it is possibly destined for a standalone VR HMD at some point, or at the very least being marketed to attract interest to this end.

 

[nicalandia]

I doubt 16C is likely for a gaming APU while game consoles remain 8C max for the PS5/XSX generation.

Current Mobile Alder Lake tops at 14C/20T, Raptor Lake will likely exceed that and AMD Knows that. So how are they supposed to have the highest Core/Thread/Cache count if not with 16C/32T?

 

[uzzi38]

Current Mobile Alder Lake tops at 14C/20T, Raptor Lake will likely exceed that and AMD Knows that. So how are they supposed to have the highest Core/Thread/Cache count if not with 16C/32T?

View attachment 61110

Raptor Lake and Meteor Lake-P are both 6+8. The HX parts (if they exist) are likely to exceed 16c, but not the regular parts.

 

[nicalandia]

Raptor Lake and Meteor Lake-P are both 6+8. The HX parts (if they exist) are likely to exceed 16c, but not the regular parts.

My point is still valid. How is AMD going to say they have the Highest Core Count with only 8C/16T? When even Alder Lake HK exceeds that?

 

[AtenRa]

I bet AMD will return to monolithic dies for desktop with ZEN4.
Using two monolithic dies for both Desktop and Mobile will solve many problems at once for them in order to be able to compete against Intel Raptor Lake in sub $200 segment.
They are unable to compete in the sub $200 segment with chiplet SKUs and Intel RaptorLake Core i5 sub $200 will see an increase in core/Thread count. (Core i5 13400F will be 6+4)

example,

Small die with 8C 16T + iGPU , use this one for 4C 8T all the way up to 8C 16T SKUs both for desktop and Mobile. Die could be close to 130-140mm2 at 5nm
This die will have bigger iGPU for Mobile use.

Large die with 16C 32T + iGPU, use this one for 10C 20T all the way up to 16C 32TT SKUs both for desktop and Mobile. Die could be close to 200-220mm2 at 5nm
This die will have smaller iGPU to keep die size at around 200mm2 and secondly because we dont need that much iGPU performance for those high end SKUs both in Desktop and Mobile as the vast majority will also use a dGPU.

For desktop , they will launch the high-end SKUs with the large die first in early Q4 2022 , example = Ryzen 9 7950X and 7900X.
In Q1 2023 they will release Ryzen 7 (Big Die with perhaps 10C 20T) and then in Q2 2023 they will release Ryzen 5 8C 16T from the small die. Ryzen 3 could also launch later in the same quarter using the small die.

For Mobile, in Q1 2023 they will release both Dragon Range (big die) and Phoenix (small die before the desktop)

Desktop could see the introduction of Ryzen 6000 on the AM5 launch so they will have more than two SKUs for the platform in 2022 (both Ryzen 7000 with ZEN4 and cheaper Ryzen 6000 with ZEN3+).

There is one more alternative, they will release Phoenix in Q1 2023 both for Mobile and Desktop and never release Ryzen 6000 for desktop AM5.


ps. Purely my speculation.

 

[jpiniero]

I doubt 16C is likely for a gaming APU while game consoles remain 8C max for the PS5/XSX generation.

IMO if Dragon Range differs significantly from the Phoenix hardware spec beyond LPDDR5 -> DDR5 support I imagine that it would be in adding stacked Infinity cache and/or greater GPU WGP count.

Dragon Range is literally Raphael BGA. Small change it's Zen 4 + Zen 4c but otherwise identical.

My point is still valid. How is AMD going to say they have the Highest Core Count with only 8C/16T? When even Alder Lake HK exceeds that?

So either 16/32 or 24/48.

 

[DisEnchantment]

They are unable to compete in the sub $200 segment with chiplet SKUs and Intel RaptorLake Core i5 sub $200 will see an increase in core/Thread count. (Core i5 13400F will be 6+4)

If you read the earnings report, they are not really focused on the budget market. They don't have the supply to do so. That is stated many times in the earnings call.
At best, older chips like Zen3 will address that market. The target audience for Zen4 won't mind the higher DDR5 costs either.
Lower end chips comes much later when they have harvested enough defective chips.

 

[AtenRa]

If you read the earnings report, they are not really focused on the budget market. They don't have the supply to do so. That is stated many times in the earnings call.
At best, older chips like Zen3 will address that market. The target audience for Zen4 won't mind the higher DDR5 costs either.
Lower end chips comes much later when they have harvested enough defective chips.

Its not only the supply, the cost of producing CPUs with chiplets is higher vs monolithic in the sub 200$ segment. So they prefer to use the chiplet dies for higher-end SKUs.
But they cannot compete in the sub $200 by using older dies, we can see this even today with Ryzen 5500 and Ryzen 4000 in desktop which all of them come from the mobile monolithic parts and not the Chiplets.

A monolithic ZEN4 8C 16T + RDNA3 iGPU APU at 5nm will be very small, close to 130-140mm2. In comparison Ryzen 6000 ZEN3+ (8C 16T) with RDNA2 iGPU is close to 210mm2 and the chiplet ZEN 3 8C 16T with IO die is 80mm2 + 125mm2 + interposer.
If im not mistaken, Core i3 12100 die is 165mm2.

So you can see that AMD with a small monolithic 8C 16T APU at 5nm could be able to have both a competitor to core i3 and to Core i5 13600K at a way smaller die (higher volume) and way higher iGPU performance. They cannot do that with chiplets or a single big 200mm2 APU they have today.

Also to note they need that small die for both Desktop and Mobile.

edit: So to sum up my speculation, I believe AMD will only use chiplets in Server this time and not in desktop.

 

[jpiniero]

Its not only the supply, the cost of producing CPUs with chiplets is higher vs monolithic in the sub 200$ segment. So they prefer to use the chiplet dies for higher-end SKUs.
But they cannot compete in the sub $200 by using older dies, we can see this even today with Ryzen 5500 and Ryzen 4000 in desktop which all of them come from the mobile monolithic parts and not the Chiplets.

Is that really a problem though.

 

[AtenRa]

Is that really a problem though.

Not being able to compete in the sub $200 ??? absolutely

They need higher volume not only for them but for the Motherboard makers.

 

[jpiniero]

Not being able to compete in the sub $200 ??? absolutely

They need higher volume not only for them but for the Motherboard makers.

In case you haven't noticed AMD is all about the ASP now. They are so short on Epyc demand that you almost wonder how much more money they could be making if they had more wafers. Mid range and lower desktop is obviously not a priority. I think the dumping ground is all you are going to get there.

 

[nicalandia]

In case you haven't noticed AMD is all about the ASP now. They are so short on Epyc demand that you almost wonder how much more money they could be making if they had more wafers.

Also the market share % should have been higher if they had more wafers. They are constrained as they are and still gaining YoY

 

[AtenRa]

In case you haven't noticed AMD is all about the ASP now. They are so short on Epyc demand that you almost wonder how much more money they could be making if they had more wafers. Mid range and lower desktop is obviously not a priority. I think the dumping ground is all you are going to get there.

5nm is 2x more dense vs 7nm, you can increase volume with the same number of wafers vs 7nm.
By not using Zen3 7nm chiplets for Desktop as much as they did in 2021 and 1H 2022 due to the ZEN4 release, they will have higher volume of chiplets left and IO dies to use for EPYC SKUs until they will release a new ZEN4 EPYC (end of 2023 start of 2024 ??).

As far as ASP goes, a small 130-140mm2 ZEN4 8C 16T APU could reach up to $300 price segment in desktop (against Core i5 13600K). That will increase margins over a 80mm2 chiplet + 125mm2 IO + interposer (Ryzen 5000) and will also increase volume and will be able to use the same die for both Desktop and Mobile. Win win win situation

 

[desrever]

5nm is 2x more dense vs 7nm, you can increase volume with the same number of wafers vs 7nm.

The under $200 market is served by lower end cut down parts. AMD doesn't care if they barely sell anything in that market segment because silicon is expensive now. 5nm cost per transistor is probably the same if not higher than 7nm, if they can fab x number of transistors given their wafer supply. Why would they care about sub $200 when basically all their chips will sell for more?

Most of the demand for lower end chips are in laptops and prebuilts anyways and those are served by APUs.

 

[nicalandia]

5nm is 2x more dense vs 7nm, you can increase volume with the same number of wafers vs 7nm.

Yet the Zen4 Chiplets remain about the same size than the Zen3, the density gained is used in more performance, not in making smaller CPUs

 

[AtenRa]

The under $200 market is served by lower end cut down parts. AMD doesn't care if they barely sell anything in that market segment because silicon is expensive now. 5nm cost per transistor is probably the same if not higher than 7nm, if they can fab x number of transistors given their wafer supply. Why would they care about sub $200 when basically all their chips will sell for more?

Most of the demand for lower end chips are in laptops and prebuilts anyways and those are served by APUs.

The same die will be used for both Mobile AND Desktop, so by using a single die they will serve both markets.
Due to single die used for both markets, it will decrease cost of development, will decrease the time of development, will decrease manufacturing cost, will decrease assemble and packaging cost, will decrease assemble and packaging time thus raise end product volumes, will increase margins etc etc.

You have to understand here that Im not saying sub $200 market is the number ONE priority, what im saying is that they will also be able to compete in the sub $200 market this way something they cannot do today with Ryzen 5000.

 

[jpiniero]

The same die will be used for both Mobile AND Desktop, so by using a single die they will serve both markets.

If anything the AMD slide suggests that Rembrandt may not be released on desktop any time soon, and would only be done so to dump dies that won't work in mobile.

Phoenix might be the same way.

 

[Mopetar]

The same die will be used for both Mobile AND Desktop, so by using a single die they will serve both markets.
Due to single die used for both markets, it will decrease cost of development, will decrease the time of development, will decrease manufacturing cost, will decrease assemble and packaging cost, will decrease assemble and packaging time thus raise end product volumes, will increase margins etc etc.

You have to understand here that Im not saying sub $200 market is the number ONE priority, what im saying is that they will also be able to compete in the sub $200 market this way something they cannot do today with Ryzen 5000.

Maybe plans changed, but there was supposed to be a low-end part (probably more like $100 market) that was being made with older technology on one of Global Foundries newer nodes. Maybe it's something that will target the Asian market more than US/Europe, but it would probably be the new Athlon CPUs which are the bottom of the barrel as far as AMDs branding has gone.

Unless they scraped that, I have a suspicion that the newer TSMC-based products will not occupy the low end of the market. The wafers are too valuable to be used for that market segment and the Zen cores are small enough that AMD could make a single APU with up to 16 cores and bin aggressively to have a mix of products available.

 

[Glo.]

If you read the earnings report, they are not really focused on the budget market. They don't have the supply to do so. That is stated many times in the earnings call.
At best, older chips like Zen3 will address that market. The target audience for Zen4 won't mind the higher DDR5 costs either.
Lower end chips comes much later when they have harvested enough defective chips.

Rembrandt will be used for low end AM5, for the time being.

Even after Phoenix lands on desktop in 2H 2023, thats my belief.

 

[LightningZ71]

There will certainly be a market for Rembrandt dies that have broken iGPUs. It just won't exist until AM5 hits the scene.

 

[tomatosummit]

If anything the AMD slide suggests that Rembrandt may not be released on desktop any time soon, and would only be done so to dump dies that won't work in mobile.

Phoenix might be the same way.

As mentioned by someone else the slide doesn't cover everything, only high end gaming stuff that desktop apus haven't been traditionally included in.

My bet is rembrandt is coming to am5 but maintrem oem focussed, although no as bad as renoir was for consumer availability.
Although am4 is going to stick around, am5 will be a big new shiney marketing program for oems so they'll want a nice big range of products to offer and hopefully remb is in the lower lineup somewhere. Especially with the current low end dgpu market.

 

[jamescox]

Its not only the supply, the cost of producing CPUs with chiplets is higher vs monolithic in the sub 200$ segment. So they prefer to use the chiplet dies for higher-end SKUs.
But they cannot compete in the sub $200 by using older dies, we can see this even today with Ryzen 5500 and Ryzen 4000 in desktop which all of them come from the mobile monolithic parts and not the Chiplets.

A monolithic ZEN4 8C 16T + RDNA3 iGPU APU at 5nm will be very small, close to 130-140mm2. In comparison Ryzen 6000 ZEN3+ (8C 16T) with RDNA2 iGPU is close to 210mm2 and the chiplet ZEN 3 8C 16T with IO die is 80mm2 + 125mm2 + interposer.
If im not mistaken, Core i3 12100 die is 165mm2.

So you can see that AMD with a small monolithic 8C 16T APU at 5nm could be able to have both a competitor to core i3 and to Core i5 13600K at a way smaller die (higher volume) and way higher iGPU performance. They cannot do that with chiplets or a single big 200mm2 APU they have today.

Also to note they need that small die for both Desktop and Mobile.

edit: So to sum up my speculation, I believe AMD will only use chiplets in Server this time and not in desktop.

You seem to be thinking 2 dimensionally. These are 2023 products when different stacking tech will be going mainstream (2.5D and 3D). The rumors on the gpu side seem to say that there will be an infinity cache base die with a compute die stacked on top; they may be 128 MB each. It looks like two of those (at least) might be connected with EFB. It is unclear how they scale to more (more EFB or IFOP style IF?). If the infinity cache die is a modular thing used all over the place, then it is plausible that they would make a device with two base die connected by EFB, one with a gpu chiplet and the other with cpu chiplet(s). Given the wording, it seems like it will be at least 16 cores. If the base die can hold 2 cpu chiplets, it may be plausible to do one high clock 8 core and one low power 16 core die. This may be wild speculation though.

AMD will almost certainly be using more APUs in the desktop space, but I think they will have a chiplet based product available. They will be making massive numbers of chiplets for Genoa, so using them in some high end desktop parts is very likely. They might only be in the Ryzen 9 series though. I have wondered if they will do something even stranger and design the desktop IO die to be connected together with a second die via an EFB link. This would allow them to make very cheap Threadripper and workstation parts with double the core count, memory channels, and IO vs AM5. That would preferably be in a new intermediate socket between AM5 and SP5. They wouldn’t have to use massive SP5 IO die on much cheaper products.

It may also be possible to connect two APUs with an EFB link. That is what the Apply M1 Ultra seems to be. If they have a big APU with 16 cores and a good sized gpu, then the high end version could just be 2 of them hooked together for ridiculous core counts never seen before in a mobile processor. Apple has 20 cores in the M1 ultra, right? That isn’t really a mobile chip yet. Although, If they have some kind of modular, stacked infinity cache based solution coming later, then such an in between product may not make sense.

If you start taking 3 dimensions into account, then you can come up with some really crazy speculation, but people don’t seem to realize that some of this has already been done (Apple M1 Ultra). It doesn’t seem that unlikely that AMD would do something similar.

 

[soresu]

Rembrandt will be used for low end AM5, for the time being.

Even after Phoenix lands on desktop in 2H 2023, thats my belief.

DDR5 RAM will be far more problematic as far as costs go for some time yet vs APU costs for system builders.

 

[Glo.]

DDR5 RAM will be far more problematic as far as costs go for some time yet vs APU costs for system builders.

I think people should leave the hope that prices will ever come back for DDR5, to best DDR4 levels.

I expect 100€ per 16 GB DDR5 to be the best possible price.

So its a matter: get used to it.