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 My advice to everyone: treat the old APU rumours as being out of date.

 

[IntelUser2000]

DDR5 on Zen 4 is already well stacked with 4x32 bit channels.

Which is nothing special since that's 128-bit and we've been 128-bit since Willamette Pentium 4 days back in 2002. DIMMs are 64-bits and Willamette used dual channel DDR, hence 128-bit.

 

[itsmydamnation]

Which is nothing special since that's 128-bit and we've been 128-bit since Willamette Pentium 4 days back in 2002. DIMMs are 64-bits and Willamette used dual channel DDR, hence 128-bit.

except that the concurrency is doubled so for scalar IPC its more like having 4 ddr channels.

 

[Kaluan]

The small and normal looks realistic for next year to me.
Strix Point could be an interesting product, but I am not waiting for It. I am not even sure we will see It at CES 2024.

Angstronomics seems to think Strix Point is confirmed to be announced at CES 2024. So there's that.

There's also rumblins about EPYC Turin in Q2 2024. AMD seems to be doing everything but slowing down. Good on them but I hope they don't just fixate on being on schedule while releasing unfinished and/or paper launch products. Like RDNA3 in the former or Rembrandt (and possibly PP) in the latter.

 

[TESKATLIPOKA]

Angstronomics seems to think Strix Point is confirmed to be announced at CES 2024. So there's that.

There's also rumblins about EPYC Turin in Q2 2024. AMD seems to be doing everything but slowing down. Good on them but I hope they don't just fixate on being on schedule while releasing unfinished and/or paper launch products. Like RDNA3 in the former or Rembrandt (and possibly PP) in the latter.

Even If Angstronomics things announcement is planned at CES 2024 It's still one year away. It could end up delayed.
I think I read somewhere that some limited number of Phoenix laptops should be available in Q1 of this year, more to come in Q2.
Still, I am from a pretty small country, and we have here a very limited number of Rembrandt laptops, so I fear It will be the same with Phoenix.

 

[jpiniero]

Like RDNA3 in the former or Rembrandt (and possibly PP) in the latter.

Rembrandt's issue I am assuming was more demand... they greatly prefer Barcelo. We'll see about Phoenix.

 

[IntelUser2000]

For those believing in super large iGPUs:
-Not a new idea or a belief, it's been hoped for over 15 years
-The only iGPU vendor now has a dGPU lineup(Intel)
-The only dGPU vendor is very profitable(Nvidia)

It can't be "all or nothing". They can all exist. dGPUs will continue to exist probably 20+ years into the future. It's all a good thing. Computers continue to proliferate in all shapes, sizes, and power envelopes.

 

[soresu]

For those believing in super large iGPUs:
-Not a new idea or a belief, it's been hoped for over 15 years
-The only iGPU vendor now has a dGPU lineup(Intel)
-The only dGPU vendor is very profitable(Nvidia)

You missed the rather sizable obvious example.

AMD - all 3 PS4 SoCs, all 3 XB1 SoCs, the PS5 SoC and the XSX SoC.

The PS5 Pro and XSX refresh are not confirmed yet, but we can safely assume they likely have even bigger GPUs on board.

To say nothing of at least 1 other semi custom APU for Subor.

IMHO the only reason AMD have not leveraged this potential yet on PC is an agreement with their semi custom console partners to not step on their toes as such a chip in the PC market would inevitably make decently powered console form factor PCs much more viable and impact the console market.

Once nVidia and Intel inevitably take this path though it will become impossible for AMD to avoid answering them in kind to maintain marketshare.

 

[moinmoin]

AMD - all 3 PS4 SoCs, all 3 XB1 SoCs, the PS5 SoC and the XSX SoC.

The PS5 Pro and XSX refresh are not confirmed yet, but we can safely assume they likely have even bigger GPUs on board.

To say nothing of at least 1 other semi custom APU for Subor.

You arguably can also include Steam Deck with Van Gogh, as well as XSS. All those APUs are all more comparable to Apple's M series in that those are not chips sold on the open market but requested for specific end products. So there it makes perfect sense to spend the initial cost for a better simplified overall package then sold unchanged in significant amounts.

PC being so customizable leads most OEMs to just-in-time put together their configurations so dGPUs are preferable for them. APUs with bigger iGPUs will be available as soon as there's significant demand. Currently going by the amount of models available it seems open market demand even for Rembrandt is not really there.

 

[Glo.]

For those believing in super large iGPUs:
-Not a new idea or a belief, it's been hoped for over 15 years
-The only iGPU vendor now has a dGPU lineup(Intel)
-The only dGPU vendor is very profitable(Nvidia)

It can't be "all or nothing". They can all exist. dGPUs will continue to exist probably 20+ years into the future. It's all a good thing. Computers continue to proliferate in all shapes, sizes, and power envelopes.

Well, its funny that you say this, but... IF Intel would not be venturing into powerful SOCs we would not have this conversation in the first place .

Intel specifically develops dGPUs and their drivers, because they want to integrate their dGPUs into their SOCs.

dGPU business in itself is not profitable enough, for Intel, you need to have good reason to go into it, considering the amount of money and time it consumes.

And yes, Intel will be one of companies which will deliver large iGPU that will be competing in performance with dGPUs. All about money and marketshare, and changing the form factor of computers.

And no, dGPUs will not die. But integrated GPU market will consume EVERYTHING up to 106 dies from Nvidia, including the 106 die. After that - we will still have dGPUs. Its simply market segments will shift. Midrange core i5-12600K/3060-type of a PC performance will simply be eaten by APUs/SOCs. If we will get 256 Bit buses on SOCs/APUs I expect that maybe even i9-12900/3060 Ti market segment could be fulfilled by this, but there will be little to no financial incentive to do it, for desktop. For mobile - that is a completely different story.

 

[Glo.]

IMHO the only reason AMD have not leveraged this potential yet on PC is an agreement with their semi custom console partners to not step on their toes as such a chip in the PC market would inevitably make decently powered console form factor PCs much more viable and impact the console market.

The reason why AMD hasn't leveraged their position in the market this way is because there was no market for it on PC side.

I mean, how many of you would even consider a product that has 8P/16E cores, 32 CU dGPU with very high clocks, and 256 bit bus RAM, comes soldered to the mITX board with RAM soldered in?

Not many of us.

And yes, CAMM allows us to have different RAM configurations, including 256 bit, but that was not available untill recently.

 

[TESKATLIPOKA]

The reason why AMD hasn't leveraged their position in the market this way is because there was no market for it on PC side.

I mean, how many of you would even consider a product that has 8P/16E cores, 32 CU dGPU with very high clocks, and 256 bit bus RAM, comes soldered to the mITX board with RAM soldered in?

Not many of us.

And yes, CAMM allows us to have different RAM configurations, including 256 bit, but that was not available untill recently.

Why should we limit this only to desktop? In laptops "everything" is already soldered to the board, so this would make much more sense there than in desktop, where you would prefer a possibility for a later upgrade.

I don't even know why you bring up CAMM, that's not even a JEDEC standard, only Dell's proprietary tech.
Yes, it can be scaled to quad channel, but you can also solder memory on board and make 256-bit from that.
Apple did the same thing, they used 8 memory chips for 1024-bit wide LPPDR5 bus for M1 Ultra.

Still, slapping a 64MB 3d cache on top would be the cheapest option. It's only 41mm2 on a cheaper process.

 

[Tigerick]

It is funny that people are arguing something without technical info. Here comes the comparison chart to clear confusion and why I think big iGPU is not feasible for PC market, both in technical and marketing:

GPU	Memory Bus	Type	Memory BW	Timespy
GF 3060	192-bit 14 Gbps	6GB GDDR6	336 GB/s	9200 (140W)
Radeon 7600M XT (4096 ALU)	128-bit 18 Gbps	8GB GDDR6	280 GB/s	13500? (120W)
CPU + iGPU (4096 ALU)	256-bit 8533	16GB LPDDR5X	273 GB/s	< 100W?

Lets focus on notebook platform cause there is bigger market to conquer. Even with LPDDR5X support, 256-bit memory bus won't have enough bandwidth to feed both CPU and GPU. That's why AMD design custom APU with GDDR6 support; PS5 has total memory bandwidth of 448GB/s. Apple can get away with it cause Macs are not designed with gamers in mind, PC market are different, we need performance as much as possible. Of course competition between Intel and AMD only heats thing up.

GF 3060 is two years old GPU, designing something as big that still perform slower in fast pace PC market is not really feasible as explained by other. Too bad NV has to clear GF3060 this year, if not you will see GF4060 with similar performance as Radeon 7600M XT. The upcoming AMD Dragon Range 8C with Radeon 7600M XT which should cost around US$1,200 in China - of course with higher TGP. The only advantage of big iGPU is lower power consumption and thus lower performance, and they are always few steps behind full GPU performance. And next year, GPUs are getting GDDR7 support, want to guess how much bandwidth with just 128-bit interface? And no news on LPDDR6?

 

[TESKATLIPOKA]

It is funny that people are arguing something without technical info. Here comes the comparison chart to clear confusion and why I think big iGPU is not feasible for PC market, both in technical and marketing:

GPU	Memory Bus	Type	Memory BW	Timespy
GF 3060	192-bit 14 Gbps	6GB GDDR6	336 GB/s	9200 (140W)
Radeon 7700M (4096 ALU)	128-bit 20 Gbps	8GB GDDR6	384 GB/s	13500 (140W)
4096 ALU	256-bit 8533	16GB LPDDR5X	277 GB/s

Lets focus on notebook platform cause there is bigger market to conquer. Even with LPDDR5X support, 256-bit memory bus won't have enough bandwidth to feed both CPU and GPU. That's why AMD design custom APU with GDDR6 support; PS5 has total memory bandwidth of 448GB/s. Apple can get away with it cause Macs are not designed with gamers in mind, PC market are different, we need performance as much as possible. Of course competitions between Intel and AMD only heat up the systems.

GF 3060 is two years old GPU, designing something as big that still perform slower in fast pace PC market is not really feasible as explained by other. Too bad NV has to clear GF3060 this year, if not you will see GF4060 with similar performance as Radeon 7700M. The upcoming AMD Dragon Range 8C with Radeon 7700M which should cost around US$1,200 in China - of course with higher TGP. The only advantage of big iGPU is lower power consumption and thus lower performance, and they are always few steps behind full GPU performance. And next year, we are going to get GDDR7, want to guess how much bandwidth with just 128-bit interface? And no news on LPDDR6?

BW is 320GB/s for 7700M. 128/8*20=320

 

[Glo.]

Why should we limit this only to desktop? In laptops "everything" is already soldered to the board, so this would make much more sense there than in desktop, where you would prefer a possibility for a later upgrade.

I don't even know why you bring up CAMM, that's not even a JEDEC standard, only Dell's proprietary tech.
Yes, it can be scaled to quad channel, but you can also solder memory on board and make 256-bit from that.
Apple did the same thing, they used 8 memory chips for 1024-bit wide LPPDR5 bus for M1 Ultra.

Still, slapping a 64MB 3d cache on top would be the cheapest option. It's only 41mm2 on a cheaper process.

Because what I believe will happen for mainstream desktops, is that everything up to Core i5/Ryzen 5 will become soldered, and only the highest end, of highest end DYI desktop will be DYI, still.

And CAMM may have to play a role in that.

Remember key sentence that Raja Koduri in recent interview said: "Changing the form factor of a PC".

 

[Thibsie]

Read a bit about CAMM.
Mmmm seems to me it sounds much like scam but maybe it's just me.

 

[TESKATLIPOKA]

Read a bit about CAMM.
Mmmm seems to me it sounds much like scam but maybe it's just me.

It's not a scam, but I also don't see It as a great option. Height is smaller than SO-DIMM, but size is larger and even that changes based on the amount of RAM on that PCB.

 

[Glo.]

It's not a scam, but I also don't see It as a great option. Height is smaller than SO-DIMM, but size is larger and even that changes based on the amount of RAM on that PCB.

You'd rather just straight up soldered to MoBos?

 

[TESKATLIPOKA]

You'd rather just straight up soldered to MoBos?

Didn't you notice how much space It requires?
BTW laptops already have soldered RAM and for desktop what's the point of this? There is not enough space for this on ITX boards, unless It's placed vertically and not horizontally like in Dell laptop.

 

[Glo.]

Didn't you notice how much space It requires?
BTW laptops already have soldered RAM and for desktop what's the point of this? There is not enough space for this on ITX boards, unless It's placed vertically and not horizontally like in Dell laptop.

Back of the Mobo. For both: soldered RAM, or CAMM module connection.

Or:

PRO HM570TI-B I526

PRO series motherboards with 11th Gen Intel® Core™ i5-1126H Processors, tuned for better performance by USB 3.2 Gen 1, TPM header

www.msi.com

Where the SO-DIMM slots are.

 

[TESKATLIPOKA]

Back of the Mobo. For both: soldered RAM, or CAMM module connection.

Not a bad idea, but that's applicable only in desktop.

 

[Exist50]

It's not a scam, but I also don't see It as a great option. Height is smaller than SO-DIMM, but size is larger and even that changes based on the amount of RAM on that PCB.

To my understanding, the Dell module you're using as a reference is comparable to the largest JEDEC-proposed form factor, and replaces 2x SO-DIMM with one single-sided module. There will be other, smaller modules.

Imo the far more interesting benefit of CAMM would be the ability to have socketed LPDDR. It could make laptops much more upgradable/repairable, and even make LPDDR viable for desktops. It is faster, after all, so if the soldered limitation is removed, why not switch almost everything to it?

 

[Glo.]

Imo the far more interesting benefit of CAMM would be the ability to have socketed LPDDR. It could make laptops much more upgradable/repairable, and even make LPDDR viable for desktops. It is faster, after all, so if the soldered limitation is removed, why not switch almost everything to it?

Considering that mainstream desktops would be just "immobile" laptops, with mobile Parts, that are in the case of Intel from P lineup - consider what you have written a very plausible possibility .

A next gen desktop computer with Arrow Lake-P Core i5-15600, soldered to Mobo, with LPDDR5X 8400 MHz, 6P/8E/384EU PC, that is no larger than 15x15x5 cm PC.

Didn't you notice how much space It requires?
BTW laptops already have soldered RAM and for desktop what's the point of this? There is not enough space for this on ITX boards, unless It's placed vertically and not horizontally like in Dell laptop.

Forgot to mention. Indeed, for laptops, soldering of RAM is not a problem. But on desktop space - you still want some upgradeability. So here would CAMM come to the rescue, you would get all of the benefits of having soldered memory, but without all of its drawbacks. You would still have upgrade paths.

 

[Exist50]

Considering that mainstream desktops would be just "immobile" laptops, with mobile Parts, that are in the case of Intel from P lineup - consider what you have written a very plausible possibility .

Oh I wasn't even talking about soldered chips. There's no technical reason why we can't have our typical, socketed CPUs, just paired with CAMM modules rather than typical DDR5 DIMMs. Really the only thing you'd lose is max capacity from a 2 DIMMs per channel implementation, and that's dying anyway.

 

[DrMrLordX]

Once nVidia and Intel inevitably take this path though it will become impossible for AMD to avoid answering them in kind to maintain marketshare.

Mi300 is already a direct shot across nVidia's bow. They want to get cute with Grace? AMD has something loaded in the barrel and ready to go. The real question is: how long will it take for something like that to trickle into the consumer space? And of course they're not going to try selling CDNA anything to us. I wouldn't think.

The reason why AMD hasn't leveraged their position in the market this way is because there was no market for it on PC side.

I mean, how many of you would even consider a product that has 8P/16E cores, 32 CU dGPU with very high clocks, and 256 bit bus RAM, comes soldered to the mITX board with RAM soldered in?

Depends, how often can I update this platform, and at what price? Can I just bolt a full-cover waterblock to it, crank up the clocks and go to town? The dGPU situation right now is absolutely miserable. A change might be in the cards.

 

[MadRat]

A high quality brand using soldered in memory for a good all around value is going to sell. The problem isn't that the memory is soldered in, its the track record of OEMs with selling dodgy kit. My wife is going on 5 years with the same desktop, and the only change in it was going to a solid state drive. Performance was never bad but the solid state drive made her very happy with near instant load times. And it came with two empty memory slots if it needs the extra memory. But I already know it won't ever get that extra memory because newer technology is cheaper. OEMs can deliver a base of soldered in memory that is very good quality and leave a single CAMM slot for expansion, and most people would be happy. And most people would probably never use the CAMM slot, because most people simply never upgrade. But they want the illusion that its possible.