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 My advice to everyone: treat the old APU rumours as being out of date.

 

[Joe NYC]

Instead of messing around with sod infinity caches or expensive DDR5 modules,
Why can't AMD simply put a larger memory controller in Phoenix Point? What is stopping them? Apple already does this with their M chips and the results are marvellous.

256-bit or even 192-bit would suffice. Pair that with LPDDR5 (which i am sure is cheaper than standard DDR5, and you get plenty of bandwidth).

I have been puzzled about this as well for some time.

It's like leaving money and performance on the table - for the sole benefit that OEMs can pair garbage memory with AMD systems.

 

[itsmydamnation]

I have been puzzled about this as well for some time.

It's like leaving money and performance on the table - for the sole benefit that OEMs can pair garbage memory with AMD systems.

its really simple , the APU's dont really need more bandwidth. So if you double the bandwidth what are you going to do with it.

1. clock things higher ,TDP issues in the primary markets.
2. build more things ( CPU cores ,GPU cores ) which requires more die space.

 

[soresu]

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.

If it came with HBM on package (or GDDR on a BGA mobo) and/or infinity cache to address the BW issue I would 100% go for it.

Having a large PC has always felt cumbersome to me, and yet I dislike tethering myself to a more restrictive regime like console gaming and their closed systems.

Of course I could just buy a Rembrandt NUC for casual PC use, but then that's even more space taken up in my smol room and extra bother to manage changing HDMI connections on the TV display when switching between it and my gaming/work rig.

 

[soresu]

And of course they're not going to try selling CDNA anything to us. I wouldn't think.

That is what the new ML acceleration units are for in RDNA3 CUs 😁

This covers AMD to some extent for less hardcore use as and when ML becomes a greater necessity in the consumer space for ML assisted gfx techniques such as path guiding in path tracing and content creation tools related to ML ala nVidia's recently announced NeuralVDB.

 

[Tigerick]

AMD has hinted on upcoming 7000G APU for desktop system, which is essential desktop version of Phoenix Point. We can make a guess on specs and prices:

	Initial Prices	Current Prices	CPU Cores	GPU	Shaders	Clock	Timespy
5600G	$259	$129	6	Vega 7 CU	448	1.9GHz	1345
5700G	$359	$199	8	Vega 8 CU	512	2 GHz	1468
7600G	$199?		6	RDNA3 8 CU	512 SP (1024 ALU)	?	?
7700G	$299?		8	RDNA3 12 CU	768 SP (1536 ALU)	?	3500 ?

5000G are made by TSMC N7 process with 180mm2 die size and 10.7 billion transistors. What I am interested is how much more transistors being add to 7000G series? My guess would be at least 50% extra which is around 16 billions. We shall see...

 

[TESKATLIPOKA]

AMD has hinted on upcoming 7000G APU for desktop system, which is essential desktop version of Phoenix Point. We can make a guess on specs and prices:

	Initial Prices	Current Prices	CPU Cores	GPU	Shaders	Clock	Timespy
5600G	$259	$129	6	Vega 7 CU	448	1.9GHz	1345
5700G	$359	$199	8	Vega 8 CU	512	2 GHz	1468
7600G	$199?		6	RDNA3 6/8 CU ?	384/512 SP ?	?	?
7700G	$299?		8	RDNA3 12 CU	768 SP (1536 ALU)	?	3500 ?

5000G are made by TSMC N7 process with 180mm2 die size and 10.7 billion transistors. What I am interested is how much more transistors being add to 7000G series? My guess would be at least 50% extra which is around 16 billions. We shall see...

R7 7700 and R5 7600 should cost $229 and $329. I don't think G series will cost less, unless AMD lowers prices once more.

 

[BorisTheBlade82]

ATM I am not so sure, how AMD will position Phoenix Point in the Desktop space. As Raphael now also has a tiny iGPU that solves the most Office/Web usecases, they might want to place it in OEM SFF/AiO and embedded first and foremost.

Regarding costs:
We can expect Phoenix Point at maybe 190mm², as a lot of the components scale pretty badly at 5nm. At the same defect density this means 224 good dies (76,5% yield) vs. 719 good dies (90,3% yield) for Raphael. Or an increase of costs per die and low-cost-SKU of 321%. That is a pretty large margin to fit a 6nm cIOD and some basic packaging into.

 

[Glo.]

Regarding costs:
We can expect Phoenix Point at maybe 190mm², as a lot of the components scale pretty badly at 5nm. At the same defect density this means 224 good dies (76,5% yield) vs. 719 good dies (90,3% yield) for Raphael. Or an increase of costs per die and low-cost-SKU of 321%. That is a pretty large margin to fit a 6nm cIOD and some basic packaging into.

PHX is smaller than 190 mm2.

 

[BorisTheBlade82]

PHX is smaller than 190 mm2.

What is your assumption and on what are you basing it?

 

[TESKATLIPOKA]

What is your assumption and on what are you basing it?

Rembrandt is ~206-208mm2. Phoenix at 190mm2 would be only 9% smaller, and we are talking about 4(5)nm vs 6nm.
I think we can expect smaller size than that, there is no extra cache, more cores or CUs compared to Rembrandt.
Ok, L2 is bigger.

 

[BorisTheBlade82]

Rembrandt is ~206-208mm2. Phoenix at 190mm2 would be only 9% smaller, and we are talking about 4(5)nm vs 6nm.
I think we can expect smaller size than that, there is no extra cache, more cores or CUs compared to Rembrandt.
Ok, L2 is bigger.

Exactly, and Phoenix Point like it's predecessors has a lot of IO and caches. So don't expect miracles.

https://twitter.com/x/status/1477015065022930946

But even if it is only 160mm2, my point still stands: From a pure production cost perspective I do not expect it to be cheaper than a lowly-binned Raphael.

 

[TESKATLIPOKA]

Exactly, and Phoenix Point like it's predecessors has a lot of IO and caches. So don't expect miracles.

https://twitter.com/x/status/1477015065022930946

But even if it is only 160mm2, my point still stands: From a pure production cost perspective I do not expect it to be cheaper than a lowly-binned Raphael.

No one expect miracles. The only thing I and @Glo. did was correct you about die size.

Does It matter which one is cheaper to make? I as a customer care only about price, I don't care If AMD has more profit from this or that. I will buy what I think provides the better value. Furthermore, I don't think we will see Phoenix in desktop anytime soon.

 

[Glo.]

What is your assumption and on what are you basing it?

Last time I heard the die size was 168 mm2.

But that was long time ago, and a lot of rumors turned out to be incorrect about AMD - then - upcoming technology.

 

[BorisTheBlade82]

@TESKATLIPOKA
Well, I think I made clear in my first post, that the 190mm² were a ballpark-estimation. It does not really matter if it is 170 or 195mm². And production costs might not be of interest to you. But regarding market segmentation - which was my initial point - it is of high importance.

 

[TESKATLIPOKA]

@TESKATLIPOKA
Well, I think I made clear in my first post, that the 190mm² were a ballpark-estimation. It does not really matter if it is 170 or 195mm². And production costs might not be of interest to you. But regarding market segmentation - which was my initial point - it is of high importance.

If Phoenix cost more to produce, It can be priced higher than Raphael.
8C Phoenix and 8C Raphael in desktop are not really competitors. Phoenix is intended for people who want a stronger IGP, for them Raphael is useless with 2CU RDNA2 IGP. It's not like there is any alternative to Phoenix, so a bit higher price because of higher production cost is in my opinion acceptable.

It's not like the higher production cost would make a significant difference on the final price.
Just an example:
5nm 190mm2 => 324 dies per wafer => $15,000 / 324 = $46
5nm 170mm2 => 365 dies per wafer => $15,000 / 365 = $41
5nm 70mm2 => 930 dies per wafer => $15,000 / 930 = $16
6nm 125mm2 => 506 dies per wafer => $7,000 / 506 = $14
So basically we are talking about $10-20 difference in production cost. I don't consider this as a significant problem for market segmentation.

 

[Kaluan]

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

Also isn't the max TGP 120W not 140W? Was the case with N23 and All The Watts also alleged that, irrespective of their flip-flopping on it being RX 7600 or 7700.

AMD has hinted on upcoming 7000G APU for desktop system, which is essential desktop version of Phoenix Point. We can make a guess on specs and prices:

	Initial Prices	Current Prices	CPU Cores	GPU	Shaders	Clock	Timespy
5600G	$259	$129	6	Vega 7 CU	448	1.9GHz	1345
5700G	$359	$199	8	Vega 8 CU	512	2 GHz	1468
7600G	$199?		6	RDNA3 6/8 CU ?	384/512 SP ?	?	?
7700G	$299?		8	RDNA3 12 CU	768 SP (1536 ALU)	?	3500 ?

5000G are made by TSMC N7 process with 180mm2 die size and 10.7 billion transistors. What I am interested is how much more transistors being add to 7000G series? My guess would be at least 50% extra which is around 16 billions. We shall see...

I think 3500 TSG is a bit too low for a full fat desktop Phoenix part. Considering hypothetical desktop Rembrandt would already hover around 2800-3000.

 

[Glo.]

If Phoenix cost more to produce, It can be priced higher than Raphael.
8C Phoenix and 8C Raphael in desktop are not really competitors. Phoenix is intended for people who want a stronger IGP, for them Raphael is useless with 2CU RDNA2 IGP. It's not like there is any alternative to Phoenix, so a bit higher price because of higher production cost is in my opinion acceptable.

It's not like the higher production cost would make a significant difference on the final price.
Just an example:
5nm 190mm2 => 324 dies per wafer => $15,000 / 324 = $46
5nm 170mm2 => 365 dies per wafer => $15,000 / 365 = $41
5nm 70mm2 => 930 dies per wafer => $15,000 / 930 = $16
6nm 125mm2 => 506 dies per wafer => $7,000 / 506 = $14
So basically we are talking about $10-20 difference in production cost. I don't consider this as a significant problem for market segmentation.

APUs are supposed to be under the desktop highest end products, and are supposed to replace, at some point, entry level CPUs from AMD. They are, supposedly, planned to become Ryzen 5 lineup, at best, and above that - chiplet based CPUs, with small iGPU. So Ryzen 3 and Ryzen 5 - all based on APUs.

And yes, production costs will play huge role in viability of desktop offerings for Zen APUs.

Thats the general idea. If it will happen soon enough, I don't know, and I don't even think this is possible. With Strix Point - thats more possible.

 

[BorisTheBlade82]

If Phoenix cost more to produce, It can be priced higher than Raphael.
8C Phoenix and 8C Raphael in desktop are not really competitors. Phoenix is intended for people who want a stronger IGP, for them Raphael is useless with 2CU RDNA2 IGP. It's not like there is any alternative to Phoenix, so a bit higher price because of higher production cost is in my opinion acceptable.

It's not like the higher production cost would make a significant difference on the final price.
Just an example:
5nm 190mm2 => 324 dies per wafer => $15,000 / 324 = $46
5nm 170mm2 => 365 dies per wafer => $15,000 / 365 = $41
5nm 70mm2 => 930 dies per wafer => $15,000 / 930 = $16
6nm 125mm2 => 506 dies per wafer => $7,000 / 506 = $14
So basically we are talking about $10-20 difference in production cost. I don't consider this as a significant problem for market segmentation.

First of all, your calculation looks weird - what kind of defect density did you use? I used 0.14 as this gives 90% for the Zen4 CCD - which sounds reasonable. Did you use 300mm wafers as is the industry standard?
So going by the numbers of my previous post and applying your wafer costs the difference is 67 USD vs. 20,87 USD, or 46 USD or 320%. So factoring in your 14 USD for the IOD there are still 32 USD left per unit. After packaging still around 20.
And if you think that 20 bucks per unit of pure variable production costs are not significant, then I honestly don't know what. to. say. These costs define the lowest possible price for which it is worth for AMD to produce and sell a unit. And all I am saying is that for me Raphael as a basis for future low cost SKUs in the Desktop is more likely than Phoenix Point.

 

[jpiniero]

I think 3500 TSG is a bit too low for a full fat desktop Phoenix part. Considering hypothetical desktop Rembrandt would already hover around 2800-3000.

NBC did test a SFF with the 6900HX and it got 2735 with DDR5-4800. So maybe?

https://www.notebookcheck.net/Beelink-GTR6-Ryzen-9-6900HX-mini-PC-review-Faster-than-most-Intel-NUCs.668899.0.html

 

[TESKATLIPOKA]

First of all, your calculation looks weird - what kind of defect density did you use? I used 0.14 as this gives 90% for the Zen4 CCD - which sounds reasonable. Did you use 300mm wafers as is the industry standard?
So going by the numbers of my previous post and applying your wafer costs the difference is 67 USD vs. 20,87 USD, or 46 USD or 320%. So factoring in your 14 USD for the IOD there are still 32 USD left per unit. After packaging still around 20.
And if you think that 20 bucks per unit of pure variable production costs are not significant, then I honestly don't know what. to. say. These costs define the lowest possible price for which it is worth for AMD to produce and sell a unit. And all I am saying is that for me Raphael as a basis for future low cost SKUs in the Desktop is more likely than Phoenix Point.

What I wrote are dies per 300mm wafer, so both bad and good ones, that's why I wrote "Just an example".
If you want to say that It's flawed, then yes It is, but yours is too, because you calculate only the good dies. Where did you leave the bad ones? There are still many usable dies as cut down versions. So your calculation is also flawed. I think at least 1/2 of them can be reused.

I honestly don't understand why are you nitpicking about $10-20 difference in production cost. It's still not more than $60, while the APU would sell for >$200. AMD would still have very healthy margins on them.
BTW, It's not like I said Phoenix Point has to be cheaper than comparable Raphael. Why should It be?
Phoenix has a 12CU RDNA3 IGP, which should perform similarly or better than GTX 1650 or RX 6400 4GB and those cards cost ~ €175-190 in my country.

I don't think It would be a problem even If Phoenix Point cost $50 more than comparable Raphael, which would easily cover the higher production cost.
It would still be a lot cheaper than Raphael + GTX 1650(RX 6400).
If someone doesn't need the strong IGP in Phoenix Point or want even faster dGPU, then just buy the cheaper Raphael or Intel + dGPU If needed.

 

[IntelUser2000]

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

Console hardware LOSES money every console sold, or at best it's even. They use software to make it up.

It does not happen in PCs, where they profit off the hardware.

 

[Shivansps]

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.

Both AMD and Intel are going to start feeling the pressure from ARM offerings in the next 2 years, they have no choice. The fact that OpenGL ES got killed by Vulkan and now ARM gpus are targeting the same api that works for mobile and desktop OS is not a small detail.

I saw what the RK3588/RK3588S can do, and its the first time i can truly say i found a ARM chip that could be definately be used for desktop/notebook usage, and thats includes console gaming emulation and even light native gaming. And thats a basic 4 core A76 with 4 A55 with a Mali G610 x4. Its really old at this point. They cant sleep with GPU performance or they will start lossing market in notebooks.

 

[Mopetar]

Console hardware LOSES money every console sold, or at best it's even. They use software to make it up.

It does not happen in PCs, where they profit off the hardware.

I'm not certain they really lose money on the hardware. There's a lot of up front cost that goes into development of the console and the hardware in it, but once you spread that out over tens or hundreds of millions of units over a five year period, it means they're making money on the hardware as well.

Console manufacturers certainly would like you to believe they're cutting their own throats though. They obviously don't need to make a huge profit given they get a cut from each game sold, but that whole model falls flat on its face if the hardware doesn't sell, so better not to eat a loss on the hardware if you can help it. But by all means let the customer think you are.

 

[IntelUser2000]

I'm not certain they really lose money on the hardware. There's a lot of up front cost that goes into development of the console and the hardware in it, but once you spread that out over tens or hundreds of millions of units over a five year period, it means they're making money on the hardware as well.

That might be true in the long term and 5 years into the future but that's different from the 50%+ margins merchant chip vendors get and at launch!

The same hardware in PC will cost upwards of 50% or more.

And thats a basic 4 core A76 with 4 A55 with a Mali G610 x4. Its really old at this point. They cant sleep with GPU performance or they will start lossing market in notebooks.

What? And you think they are anywhere PC level? They don't play outside of laws of physics. And unlike CPUs where "smarts" and innovations are required to stand out, GPUs are mostly having more of them.

And please, the mobile ones throttle after few mins, because that's all they need. Not to mention the FP16 vs FP32 difference that's still widespread responsible for 30% gap difference alone. And it's all about practical reasons. Graphics is all about taking shortcuts, that's how the original Voodoo got it done and even how RTX's ray tracing is done. When you have a 4-inch screen you can skimp on details and no one will notice.

Then in the end it's all about power targets for GPU.

 

[Doug S]

Console makers sell the "same" hardware for 5+ years. Historically things like DRAM and storage become a lot cheaper for the same fixed amount over such a long period of time, and CPUs/GPUs get two process shrinks making them much cheaper. Even with the lowered prices over time they may start out unprofitable in year one but be making a tidy profit in year five and beyond.

That model may be breaking down though, as you aren't getting two full shrinks in five years at near constant per wafer costs as used to be the case so CPUs/GPUs aren't costing 1/4th of what they did originally. You might be better off keeping them on the same process and relying on getting cheaper wafers from a fully depreciated fab in year five. DRAM and NAND are getting cheaper for constant capacity, but not at the rates they used to.