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 My advice to everyone: treat the old APU rumours as being out of date.

 

[moinmoin]

LPDDR is not faster due to technical aspects but since official JEDEC specs are updated more often. That happens since demand for such is much larger in the mobile space.

Actually the main point is z-axis space savings in laptops.

There is another approach ODMs love using for saving z-axis space: just solder it. And LPDDR so far is only available soldered. So that's essentially CAMM's competition.

 

[Exist50]

LPDDR is not faster due to technical aspects but since official JEDEC specs are updated more often. That happens since demand for such is much larger in the mobile space.

No, it's faster due to technical aspects as well. We'll probably be seeing LPDDR5X-8500 by year's end. Those are speeds even overclocking struggles to reach today.

 

[moinmoin]

No, it's faster due to technical aspects as well. We'll probably be seeing LPDDR5X-8500 by year's end. Those are speeds even overclocking struggles to reach today.

Well yes, the updated specs do allow for faster speeds as well. But that doesn't mean such speed would be inherently impossible on desktop, it would just need new specs (and respective IMCs) there as well. The turn around time for new specs is much faster in the mobile space due to demand and as such naturally can accommodate new development that also affects speed.

 

[Exist50]

But that doesn't mean such speed would be inherently impossible on desktop, it would just need new specs (and respective IMCs) there as well.

The limiting factor isn't JEDEC arbitrarily defining a new speed tier. If that was the case, we'd see far higher overclocks. It's primarily the electrical characteristics where LPDDR is more stringent than DDR, and thus supports much higher speeds.

 

[moinmoin]

The limiting factor isn't JEDEC arbitrarily defining a new speed tier. If that was the case, we'd see far higher overclocks. It's primarily the electrical characteristics where LPDDR is more stringent than DDR, and thus supports much higher speeds.

What I mean is in the end JEDEC's LPDDR specs reflect the compromise what's technically reliably achievable for both the memory manufacturers and the high end IMCs available at the given time. In mobile space unlike desktop this is a very tight coupling since there is demand for such. Desktop is more driven by long term stability so fundamental specs (as in DDR3, 4, 5...) aren't updated often.

 

[FlameTail]

Are the new chips that use LPDDR5 using single channel or dual channel? I have seen mixed info. Single channel would be such a waste. Dual channel would serve the iGPU well.

LPDDR channels are 16-bit in size.

So while a 128b DDR5 is Dual Channel, a 128b LPDDR5 is 8 channel.

 

[Doug S]

Supporting sockets limits achievable frequency. Supporting multiple sockets per channel moreso.

 

[FlameTail]

Phoenix has RDNA3. May I ask,

1) what is the IPC increase going from an RDNA2 CU to an RDNA3 CU (or WGP for that matter) ?

2) Also AMD mentioned a 54% performance per watt uplift when unveiling the Radeon 7000 cards last year. What does this 54% apply to? Can we expect the same 54% Performance per Watt uplift in the Phoenix iGPU?

 

[soresu]

1) what is the IPC increase going from an RDNA2 CU to an RDNA3 CU (or WGP for that matter) ?

Hard to tell realistically at this point.

I'm not sure if the GFX11 drivers just need time to mature or what, but they performance is certainly not as expected.

2) Also AMD mentioned a 54% performance per watt uplift when unveiling the Radeon 7000 cards last year. What does this 54% apply to? Can we expect the same 54% Performance per Watt uplift in the Phoenix iGPU?

Theoretically it should be better for Phoenix Point as it uses a more advanced process (N4), but then the Infinity Cache is an issue as we have no indication (as yet) of any such addition to the chip, and from Rembrandt we have no reason to expect it so far.

Hopefully potential silicon bugs from early GFX11 implementations were scrubbed before Phoenix design was finalized - combined with the driver being more mature at the point Phoenix silicon arrives we should see a nice efficiency increase over Rembrandt.

 

[FlameTail]

Oh, alright! I was wondering what kind of performancr we can expect IF the Exynos 2300 has an RDNA3 GPU. Although initial rumours said the Exynos 2300 was cancelled, some rumours now say that it does, indeed exist. It will reportedly power the Galaxy S22 FE and Tab S8 FE

 

[TESKATLIPOKA]

Nah I mean chipset iGPUs added just $5-7. The CPU iGPUs add $30 plus the cost of the faster chip. So $30 if you were going to buy the CPU anyway, but the real cost might be $50-100. On mobile it might be even higher since you often need to buy an i7 to get the top of the line iGPU.

When Iris devices with eDRAM were being made, it was a $200-300 adder on top of the maxed out i7 config!

And $30 is a HUGE amount in terms of production cost. If it adds $30 to the MSRP of the CPU sure, but $30 is likely the cost of the entire chip.

You don't want a backup display GPU that expensive.

There was a reason why It was so cheap.

Both chipset with IGP and CPU were dirt cheap to produce at that time.
Last chipset with IGP from AMD was 890GX with HD 4290 for AM3 socket. Die size was 73 mm2 using 55 nm process.

The cost difference is:
I will use wafer prices from this: 65nm($1937) vs 5nm($16988), let's say 55nm and 4nm have similar price. I will use for Phenom II the price of 40nm $2274.
890GX 73mm2 55nm: good dies(all dies) -> 749(806) = $2.4-2.6
Phenom II X6 1055T 346mm2 45nm: 110(154) = $14.8-20.7
For comparison:
73mm2 chip using 4nm: good dies(all dies) -> 749(806) = $21.1-22.7
346mm2 chip using 4nm: good dies(all dies) -> 110(154) = $110-154.4
Just because of increased price per wafer the $/mm2 is already ~7.45-8.8x higher.

Now back to Phoenix:
Phoenix 178mm2 4nm: good dies(all dies) -> 266(317) = $53.6-63.9
Packaging + other associated costs with production will be $20.

This is from Locuza and IGP size was calculated based on this. IC is L3 size at 4nm, also from Locuza.

The whole 12CU Phoenix IGP is ~63mm2 and I will remove It.
115mm2 4nm: good dies(all dies) -> 449(503) = $33.8-37.8
Of course, this is kinda pointless because you would need a dGPU then.
If I use just a 2CU IGP It would be ~45mm2, so I will add It back.
160mm2 4nm: good dies(all dies) -> 303(355) = $47.9-56.1
Now I will add a 24CU Phoenix IGP with ~88mm2 + 24mm2 IC or 67mm2 more than the 2CU IGP.
227mm2 4nm: good dies(all dies) -> 195(244) = $69.6-87.1

Phoenix with 24CU IGP would cost $22-31 more, depending on how much dies can be salvaged from a wafer compared to a Phoenix with a 2CU IGP.

Now I will compare margins and profit.
Let's say AMD sells R9 7940HS for $299 to manufacturers.
Product margin= (selling price – cost of product) / selling price.
Product margin: [$299-($54 or $64)-$20]/$299 = 72-75%
Profit: [$299-($54 or $64) -$20] = $215-225
If AMD sold this hypothetical Phoenix without an IGP for $219 to manufacturers.
Product margin: [$219-($34 or $538-$20]/$219 = 74-75%
Profit: [$219-($34 or $38) -$20] = $161-165
If AMD sold this hypothetical Phoenix with 2CU IGP for $269 to manufacturers.
Product margin: [$269-($48 or $56)-$20]/$269 = 72-75%
Profit: [$269-($48 or $56) -$20] = $193-201
If AMD sold this hypothetical Phoenix with 24CU IGP+IC for $369 to manufacturers.
Product margin: [$369-($70 or $87)-$20]/$369 = 71-76%
Profit: [$369-($70 or $87) -$20] = $262-279

Difference in selling price could be $50 between no IGP and 2CU IGP for comparable margins, but profit would be $32-36 higher on the one with an IGP.
Difference in selling price could be $100 between 2CU and 24CU IGP for comparable margins, but profit would be $69-78 higher on the big one.

I don't think It makes a sense to have a CPU without any IGP.
You will have to pay $100 more for a 24CU IGP.
I think that's a fantastic price for that performance.
The problem would be OEM. Manufacturers would newer put this 24CU Phoenix in a cheap laptop, so It would probably end up like Rembrandt where a laptop with faster dGPU would cost the same or less.

edit: I forgot to add 32MB IC. Facepalm. I will recalculate It.

 

[TESKATLIPOKA]

Phoenix has RDNA3. May I ask,

1) what is the IPC increase going from an RDNA2 CU to an RDNA3 CU (or WGP for that matter) ?

2) Also AMD mentioned a 54% performance per watt uplift when unveiling the Radeon 7000 cards last year. What does this 54% apply to? Can we expect the same 54% Performance per Watt uplift in the Phoenix iGPU?

1) IPC is "supposedly" 17.4% better, clocks should be also a bit higher. From IPC comparison so far this IPC is a lot different between games, In some games no IPC increase and in some>20% increase.

2) This 54% was a nonsense. AMD was playing with TDP settings on cards.
As @soresu mentioned Phoenix IGP I will add just this.
TDP and CU stayed the same with Phoenix, but boost is 25% higher. If you lowered the frequency to 2.4GHz then you should normally see more than 25% improvement in efficiency.
The problem is, you are also using CPU during gaming.
You would need to check some monitoring tool which shows only IGP power draw.

 

[IntelUser2000]

There was a reason why It was so cheap.

No the primary reason is that CPU iGPUs are easier to monetize, therefore more profit for the manufacturer.

Intel was stagnant on it's earnings until 2010, when they introduced Arrandale and Clarkdale, the first CPU with GMCH on die. You needed to buy the high end Core i5 to get the full GPU config. Another reason they increased profits is because of the Core ix branding. It's very effective. By 2011 Intel said Core products were the highest volume sellers, which was stunning to me, since Celerons were always the volume leaders.

The second year of increased profits was in 2011 with Sandy Bridge, because that product was pretty great. But also a further knock on effect of the Core ix branding.

There's a reason semiconductors are called silicon gold rush. Because it really is that profitable. If you just look at BOM and assume they'll price it entirely that way, to be blunt you are being naive. They always increase prices way more than they say bill of material increases are.

Profits are all about control. The reason ix branding works is because the manufacturer takes the control back. When it didn't use model numbers customers were fully aware of the product and bought what they needed. With branding, they lose that control. And that's also why segmentation works brilliantly from a profit perspective too. When you need it you buy it.

We are in record government, personal, and business debt. Reckless spending is the reason for some, but effective marketing is another.

 

[TESKATLIPOKA]

There's a reason semiconductors are called silicon gold rush. Because it really is that profitable. If you just look at BOM and assume they'll price it entirely that way, to be blunt you are being naive. They always increase prices way more than they say bill of material increases are.

I just calculated It this way so margins will stay comparable across different chip sizes.
If they have a good enough marketing to sell this imaginary 24CU Phoenix for more than what I calculated and there is demand, then why not.
I think this profit is still good for AMD and the final price is also very reasonable, but of course investors want the highest possible margins.

 

[FlameTail]

2) This 54% was a nonsense. AMD was playing with TDP settings on cards.

I certainly suspected something like this. 54% is a number way too high.

Oh, alright! I was wondering what kind of performance we can expect IF the Exynos 2300 has an RDNA3 GPU.

Follow up question: Is it even technically possible for Samsung to integrate an RDNA3 GPU within the current timeframe?

The Exynos 2300 should debut in Febraury, ~4 months after the first RDNA3 product was released, The Exynos 2200 which integrated RDNA2 was launched more than an year after RDNA2 debuted.

I am aware reading through these forums, that RDNA3 has some kind of hardware bug in the silicon, in addition to the the software driver issues. Is it possible that these issues could be resolved in a mobile RDNA3 implementation in the Exynos 2300?

 

[TESKATLIPOKA]

I certainly suspected something like this. 54% is a number way too high.

It was already done with RDNA2 when It used the same process, so It's not that high. The problem is that N31 didn't end up performing as was expected and TBP was also pretty high.

Follow up question: Is it even technically possible for Samsung to integrate an RDNA3 GPU within the current timeframe?

The Exynos 2300 should debut in Febraury, ~4 months after the first RDNA3 product was released, The Exynos 2200 which integrated RDNA2 was launched more than an year after RDNA2 debuted.

I am aware reading through these forums, that RDNA3 has some kind of hardware bug in the silicon, in addition to the the software driver issues. Is it possible that these issues could be resolved in a mobile RDNA3 implementation in the Exynos 2300?

They already announced and are producing Phoenix, so Exynos 2300 with RDNA3 is not impossible.
Exynos is made on a different process in different factories, so It can't have the same bug If there really is a bug.
It could be RDNA2 based.

 

[insertcarehere]

The whole 12CU Phoenix IGP is ~63mm2 and I will remove It.
115mm2 4nm: good dies(all dies) -> 449(503) = $33.8-37.8
Of course, this is kinda pointless because you would need a dGPU then.
If I use just a 2CU IGP It would be ~45mm2, so I will add It back.
160mm2 4nm: good dies(all dies) -> 303(355) = $47.9-56.1
Now I will add a 24CU Phoenix IGP with ~88mm2 + 24mm2 IC or 67mm2 more than the 2CU IGP.
227mm2 4nm: good dies(all dies) -> 195(244) = $69.6-87.1

I don't find the estimate that a 2 CU IGP will take 45mm^2 on 4nm credible, given that, as per AMD, the Zen 4 IO die is 3.4B transistors, which is "only" 1.31B over the Zen 2/3 IO die and already includes a 2CU IGP within the additional budget, among other things, so using 4nm density it's probably closer to 15mm^2.

But the more prescient point is that even being conservative with a 70-80mm^2 difference between hypothetical 2CU vs 24CU APU, that is already more than the size of a Zen 4 CCD (on 5nm no less). 2 complete CCDs + a IOD would already be selling at $600+ (7950X), while a single CCD + IOD wouldn't be less than ~$300. This puts the "value" of a single full CCD in potential revenue likely no less than $200. Can AMD obtain $200+ in price premium between putting a iGPU that can boot up a computer vs one that can maybe also play games moderately well?

 

[Tigerick]

I certainly suspected something like this. 54% is a number way too high.




Follow up question: Is it even technically possible for Samsung to integrate an RDNA3 GPU within the current timeframe?

The Exynos 2300 should debut in Febraury, ~4 months after the first RDNA3 product was released, The Exynos 2200 which integrated RDNA2 was launched more than an year after RDNA2 debuted.

I am aware reading through these forums, that RDNA3 has some kind of hardware bug in the silicon, in addition to the the software driver issues. Is it possible that these issues could be resolved in a mobile RDNA3 implementation in the Exynos 2300?

Is there a reason that you are interested on Exynos 2300? Cause as far i can tell, they are made by Samsung 4LP process; even Samsung upcoming flagship S23 series are using SD8 Gen 2

 

[TESKATLIPOKA]

I don't find the estimate that a 2 CU IGP will take 45mm^2 on 4nm credible, given that, as per AMD, the Zen 4 IO die is 3.4B transistors, which is "only" 1.31B over the Zen 2/3 IO die and already includes a 2CU IGP within the additional budget, among other things, so using 4nm density it's probably closer to 15mm^2.

Zen4 IOD 6nm: 3,400,000,000/122 = ~27 MTr/mm2
Zen3 IOD 12nm: 2,090,000,000/125 = ~16.7 MTr/mm2
7nm Zen 3 CCD has ~51.5 MTr/mm2.
5nm Zen 4 CCD has ~91 MTr/mm2.
6nm RDNA3 MCD has ~55.4 MTr/mm2. Why does It have higher density than Zen3 CCD?

If the number of transistors for a 2CU IGP is really less than 1.3B, let's say 1.2B, then with N31's density of ~150MTr/mm2 I end up with only 8mm2.
Then the whole IGP part of Phoenix is probably only ~35-40mm2. CPU part is ~50mm2.
178-35-50=93 or 178-40-50=88
Then I don't understand what is 88-93mm2 for? Why is It still so big?

But the more prescient point is that even being conservative with a 70-80mm^2 difference between hypothetical 2CU vs 24CU APU, that is already more than the size of a Zen 4 CCD (on 5nm no less). 2 complete CCDs + a IOD would already be selling at $600+ (7950X), while a single CCD + IOD wouldn't be less than ~$300. This puts the "value" of a single full CCD in potential revenue likely no less than $200. Can AMD obtain $200+ in price premium between putting a iGPU that can boot up a computer vs one that can maybe also play games moderately well?

I don't think they would sell a lot for $499, but $429 looks pretty reasonable to me.
Why would they even need to ask $200+ for this?
I understand that die size difference can be used for another CCD, but would they actually sell It? Has TSMC insufficient wafers for AMD to order more?
I see a problem with this logic.
Let's say there is a buyer for this big APU, but AMD didn't design It because It's not as profitable as 7950X, so he can't buy It.
What will he do?

For comparison.
7600X costs only $249 on Amazon. CCD + IOD
7700X costs only $349 on Amazon. CCD + IOD
7950X costs only $590 on Amazon. 2xCCD + IOD
Cheapest 6600XT costs $275 on Amazon.

Will he buy 7950X + 6600XT instead? Unlikely. He wanted only an 8C16T 24CU APU, this has 16C32T which he likely doesn't need, and you have to pay $965 for this.
Will he buy a 7700x + 6600XT? The same core count and similar GPU performance, combined price is $624, that's $195 difference.
Will he buy a 7600x + 6600XT? Lower core count and similar GPU performance, combined price is $524, that's only $95 difference. In this case, I have to wonder If AMD would actually earn more money from this combo or from the big APU.
All of them have higher price, so this big APU would offer better FPS/$ ratio to the buyer.

Then there are still 2 worse situations for AMD.
The buyer will buy AMD CPU + Nvidia GPU or Intel CPU + Nvidia GPU and AMD won't earn anything in the worst case.

Conclusion: I can see a reason for such an APU, but the question is If there would be a big enough market for this.
I have to wonder how many millions of CPUs would they need to sell to be willing to design this. This is also a reason why I wanted a chiplet GPU in Strix Point, because It would cost less to design small chiplets than two big monoliths. At least I think so.

 

[Glo.]

If the number of transistors for a 2CU IGP is really less than 1.3B, let's say 1.2B, then with N31's density of ~150MTr/mm2 I end up with only 8mm2.
Then the whole IGP part of Phoenix is probably only ~35-40mm2. CPU part is ~50mm2.
178-35-50=93 or 178-40-50=88
Then I don't understand what is 88-93mm2 for? Why is It still so big?

I think you can draw your own answers from this.

 

[TESKATLIPOKA]

I think you can draw your own answers from this.

That AI engine looks like It's only ~15mm2. Still 73-78mm2 is left. For memory controller+ PHY, PCIe+PHY and some other IO It looks pretty big, but maybe It's just not that dense.

 

[IntelUser2000]

If they have a good enough marketing to sell this imaginary 24CU Phoenix for more than what I calculated and there is demand, then why not.
I think this profit is still good for AMD and the final price is also very reasonable, but of course investors want the highest possible margins.

I think this is the hoped point of view. The reality most of the time is they'd likely price it in a way comparable to CPU + Equivalent dGPU config. And what is that price adder?

Sure it would be nice if they priced it just a little more. But they'd be losing "potential profits" by doing that. Also losing dGPU sales in the process since going the large iGPU way would be decisively better at something like 1/2 or 1/3rd of the price.

The best way I see a good iGPU is raising the minimum bar.

 

[insertcarehere]

I don't think they would sell a lot for $499, but $429 looks pretty reasonable to me.
Why would they even need to ask $200+ for this?
I understand that die size difference can be used for another CCD, but would they actually sell It? Has TSMC insufficient wafers for AMD to order more?
I see a problem with this logic.

It's not a question of "would they actually sell it", it's a question of opportunity cost and maximizing revenue per wafer, AMD is not going to be happy to sell ~80mm^2 at $50-60 when that 80mm^2 can be an additional CCD yielding ~$200+, that's Economics 101.

For comparison.
7600X costs only $249 on Amazon. CCD + IOD
7700X costs only $349 on Amazon. CCD + IOD
7950X costs only $590 on Amazon. 2xCCD + IOD
Cheapest 6600XT costs $275 on Amazon.

Will he buy 7950X + 6600XT instead? Unlikely. He wanted only an 8C16T 24CU APU, this has 16C32T which he likely doesn't need, and you have to pay $965 for this.
Will he buy a 7700x + 6600XT? The same core count and similar GPU performance, combined price is $624, that's $195 difference.
Will he buy a 7600x + 6600XT? Lower core count and similar GPU performance, combined price is $524, that's only $95 difference. In this case, I have to wonder If AMD would actually earn more money from this combo or from the big APU.
All of them have higher price, so this big APU would offer better FPS/$ ratio to the buyer.

Why does it matter that an APU may offer better FPS/$ for the buyer when its AMD who needs to make the numbers work to make such an APU worth it to them?

Setting aside my reservations on how 24CU RDNA3 might perform in an APU, 7600X is not a valid comparison here since its a binned part that doesn't need the whole CCD to function, which reduces costs, so the valid comparison here would be 7700X vs 7950X, there an additional CCD in area generates > $200 in potential revenue, so that's the sort of internal benchmark for the premium that an APU needs to get to make putting that as a GPU worth it.

 

[TESKATLIPOKA]

Why does it matter that an APU may offer better FPS/$ for the buyer when its AMD who needs to make the numbers work to make such an APU worth it to them?

If It had the same FPS/$ as CPU+dGPU including Vram, no one would want to buy It.
Do you know how much profit does AMD need to make on an APU to be worth It, because I don't?

It's not a question of "would they actually sell it", it's a question of opportunity cost and maximizing revenue per wafer, AMD is not going to be happy to sell ~80mm^2 at $50-60 when that 80mm^2 can be an additional CCD yielding ~$200+, that's Economics 101.
......
Setting aside my reservations on how 24CU RDNA3 might perform in an APU, 7600X is not a valid comparison here since its a binned part that doesn't need the whole CCD to function, which reduces costs, so the valid comparison here would be 7700X vs 7950X, there an additional CCD in area generates > $200 in potential revenue, so that's the sort of internal benchmark for the premium that an APU needs to get to make putting that as a GPU worth it.

This is not true for this case, because this big Zen4 APU is using N4 and Raphael is using N5, so different production lines.
What you wrote could happen only If Strix Point and Raphael's successor were both using the same N3 process and there were not enough wafers, but in that case I prefer a chiplet design which would cover both APUs and CPUs, where IGP would use a separate chiplet.

 

[TESKATLIPOKA]

I think this is the hoped point of view. The reality most of the time is they'd likely price it in a way comparable to CPU + Equivalent dGPU config. And what is that price adder?

Sure it would be nice if they priced it just a little more. But they'd be losing "potential profits" by doing that. Also losing dGPU sales in the process since going the large iGPU way would be decisively better at something like 1/2 or 1/3rd of the price.

The best way I see a good iGPU is raising the minimum bar.

You are talking about desktop, right?
Then this big APU doesn't need to be priced as CPU + Equivalent GPU.
That GPU's price is a lot higher than what AMD is selling that chip for to the manufacturers.
We would first need to know this price, then we can talk about potential lost profits.

If AMD made this APU, then I have to ask If AMD actually needs N33. This APU would pretty much cover this performance level.
Would a combination of AMD CPU+N33 sell more chips than this big APU especially If It cost less?