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 My advice to everyone: treat the old APU rumours as being out of date.

 

[Kaluan]

Anyone else seen AllTheWatts' cryptic emoji tweets on Dragon and Phoenix? 😂

The jist for me was Dragon = hot and blazing fast
Phoenix = very cool and very efficient

Phoenix at 15-28W is prolly gonna be a blow out agaist previous gens.

I don't know about what AMD will show us. N31 didn't leave a good impression for me to have high expectations.
I personally expect more from N32 than N33.

RX 6850M XT managed 11762 points, so I am a bit skeptical 7600M would manage the same score or that 7700 would manage >13,000 points.
The good news is that we only have to wait 3 weeks to find out. At least I hope we will see some benches.

RDNA3 launch drivers are a (performance) mess and upcoming RDNA3 product stack will likely fare better at both the silicon level and launch day software.
I wouldn't base anything off of N31 just yet, especially since it has very little in common with N33.

Still...



Suffice to say, I don't think there are any educated guesses that can be very accurate about the future right now. I think I have more questions about RDNA3 now than I had before the launch.


BTW on paper 6850M XT has better specs than 6750 XT, yet 6750 XT is at least 15% faster (again, in TS graphics). 165W vs 250W (& better cooling and power delivery for such wattage...)
So clearly none of this is apples to apples, v/f curves and other efficiency characteristics matter. This ia why a 237mm2 chip barely sees any gains with 80% more power fed into it (laptop vs desktop) but it's bigger sibling sees bigger gains with just 50%.
I think N33 at (up to) 120W will pack quite a punch. The more aggressive SKU-name-to-die scheme and TGP brackets allude to that as well.
But eh, maybe I'm wrong.

 

[jpiniero]

Lenovo will announce Phoenix based IdeaPad Pro 5 14 and 16 at CES 2023.
You can choose from R9 7840HS or R5 7640HS

https://news.lenovo.com/pressroom/press-releases/new-consumer-devices-next-gen-performance-versatility-convenience/

There's also an AMD IdeaPad Slim... but it uses Barcelo. The Pro model also has Rembrandt as an option. Also of note the AMD Pro model uses LPDDR5/X and not DDR5.

 

[Saylick]

Anyone else seen AllTheWatts' cryptic emoji tweets on Dragon and Phoenix? 😂

The jist for me was Dragon = hot and blazing fast
Phoenix = very cool and very efficient

Phoenix at 15-28W is prolly gonna be a blow out agaist previous gens.

Dragon Range being simply mobile Raphael means it's going to suffer from the same misconceptions as Raphael, i.e. that it runs hot but that's just the thermal target. If the user desires, they can easily set a lower thermal target and voila, it no longer runs hot at pretty insignificant cost to boost clocks. Also, hot or high temps =/= high energy usage.

Phoenix being very cool is likely because it has a much lower TDP. 28W is much easier to cool than the 65W or however high Dragon Range is supposed to go up to.

 

[TESKATLIPOKA]

Phoenix-U will be limited to 15-28W.
Phoenix-HS will be 35W and up to 54W maybe?
What will be interesting is Phoenix-HS versus 8C Dragon Range.

BTW All The Watts!! updated Phoenix SKUs. According to him It should look like this:

Ryzen 9 7940HS 8c 12 CU.
Ryzen 7 7840HS 8c 12 CU.
Ryzen 5 7640HS 6c 6 CU.

Ryzen 7 7840U 8c 12 CU
Ryzen 5 7640U 6c 6 CU

PHX2.
Ryzen 7 7740U 6c 4 CU.
Ryzen 5 7540U 6c 4 CU.
Ryzen 3 7340U 4c 2 CU.

I don't think a 6C PHX2 is faster than 6C PHX, yet naming says otherwise.
I hope the marketing department didn't receive any Christmas bonuses, I certainly wouldn't give them any after this BS naming scheme.

 

[jpiniero]

I don't think a 6C PHX2 is faster than 6C PHX, yet naming says otherwise.
I hope the marketing department didn't receive any Christmas bonuses, I certainly wouldn't give them any after this BS naming scheme.

It could boost higher and/or be a lower TDP range.

 

[TESKATLIPOKA]

PHX2 is supposedly 15-28W.

 

[Kaluan]

Dragon Range being simply mobile Raphael means it's going to suffer from the same misconceptions as Raphael, i.e. that it runs hot but that's just the thermal target. If the user desires, they can easily set a lower thermal target and voila, it no longer runs hot at pretty insignificant cost to boost clocks. Also, hot or high temps =/= high energy usage.

Phoenix being very cool is likely because it has a much lower TDP. 28W is much easier to cool than the 65W or however high Dragon Range is supposed to go up to.

Yeah I myself have had to explain many times that temperature does no equal high power usage.

What I took from the Phoenix thingamajig is that it can hit close to it's performance plateau at it's rated up to 28W & 35W. By contrast Rembrandt needs 45/54W to get there (which still put them ahead of 12th gen mobile but that's besides the point). With big diminishing returns for both when go up from their respective "sweet spots".

Phoenix-U will be limited to 15-28W.
Phoenix-HS will be 35W and up to 54W maybe?
What will be interesting is Phoenix-HS versus 8C Dragon Range.

BTW All The Watts!! updated Phoenix SKUs. According to him It should look like this:


I don't think a 6C PHX2 is faster than 6C PHX.

TBH I don't think our leaker friend has foolproof data on these, these new posts also serve as a correction to their previous info. PHX2 is allegedly 6 months away, why would there be SKU naming out already?

 

[TESKATLIPOKA]

TBH I don't think our leaker friend has foolproof data on these, these new posts also serve as a correction to their previous info. PHX2 is allegedly 6 months away, why would there be SKU naming out already?

It looks like he has access to OEM reference data. What he originally posted was the old one and now he updated It.
Naming for PHX2 could be a placeholder or something.
It will be interesting to see what PHX2 will show us.

 

[Kaluan]

RX6600M = 28CU:1792SP:128TMU:64ROP; boost: 2416 MHz -> 8.66 TFlops

Mobile N33 will have comparable specs to N23, but boost is ~2700MHz or ~12% higher, the rest of performance has to come from architecture.

Phoenix can be paired with both N32 and N33 or Nvidia, It has enough performance.
Mobile N33 supposedly performs like desktop RX 6750XT, but we don't know at what resolution. Performance increase should be at least 25%.

Just noticed you used opportunistic/ideal real-world 6600M (cut down chip, less silicon to power up) boosting clocks in contrast to alleged full N33 clocks.
I think that's not a good comparison.

A better framing would be 6650M XT (120W) or 6800S (up to 105W), both full fat N23. Advertised (game) clocks of 2177 and 1975MHz (AMD doesn't seem to advertise boost clocks for mGPUs). 6650M XT never saw any design wins AFAIK, so zero data on that SKU. But here's some 6800S clock data from KitGuru in a default power preset ("Performance"):

All in all, throtling or ideal power/cooling scenarios aside, I believe AMD's "game clocks" are a much better baseline for gen-on-gen comparison.

And to that end, I think it's more like ~2150MHz vs (alleged) ~2700MHz (for full dies @ 120W TGP). Or at the very least greater than 12%.

No chance on these being very affordable but I would be in the market for such a Phoenix+N33 design in late 2023.
Less than 2 weeks to go and I'll know if there's less or more to be excited about.

 

[scineram]

It looks like he has access to OEM reference data. What he originally posted was the old one and now he updated It.
Naming for PHX2 could be a placeholder or something.
It will be interesting to see what PHX2 will show us.

Interestingly MLID is now backtracking on his smol Phoenix leak due to this. But I still believe his more than this Zen 4c BS.

 

[TESKATLIPOKA]

Just noticed you used opportunistic/ideal real-world 6600M (cut down chip, less silicon to power up) boosting clocks in contrast to alleged full N33 clocks.
I think that's not a good comparison.

A better framing would be 6650M XT (120W) or 6800S (up to 105W), both full fat N23. Advertised (game) clocks of 2177 and 1975MHz (AMD doesn't seem to advertise boost clocks for mGPUs). 6650M XT never saw any design wins AFAIK, so zero data on that SKU. But here's some 6800S clock data from KitGuru in a default power preset ("Performance"):
View attachment 73303
All in all, throtling or ideal power/cooling scenarios aside, I believe AMD's "game clocks" are a much better baseline for gen-on-gen comparison.

And to that end, I think it's more like ~2150MHz vs (alleged) ~2700MHz (for full dies @ 120W TGP). Or at the very least greater than 12%.

No chance on these being very affordable but I would be in the market for such a Phoenix+N33 design in late 2023.
Less than 2 weeks to go and I'll know if there's less or more to be excited about.

I didn't make a mistake.
I compared boost against N33 because I didn't think 2.7GHz for N33 was gaming frequency.
RX 6650M XT - 32CU, gaming frequency: 2162 MHz
RX 6600M - 28CU, gaming frequency: 2177 MHz
AMD doesn't mention boost for mobile chips, but according to TPU database, both of them have the same 2416MHz boost clock and gaming frequency is almost the same.


Link

If that leak was about gaming frequency, then the difference in clock speed will be larger, but that's something we don't know yet.

 

[FlameTail]

I want to see a 6-core Phoenix with 12 CUs. That probably might not happen, would it?

Right now, with Rembrandt, the 6-core is limited to the 6 CU Radeon 660M.

 

[BorisTheBlade82]

I want to see a 6-core Phoenix with 12 CUs. That probably might not happen, would it?

Yep, I would not expect that either.

 

[TESKATLIPOKA]

12CU will be limited to 8-core Phoenix.
I don't think N4 process has a high failure rate, so AMD will want to sell Phoenix as a full chip instead of a cutdown one, which will give them higher profit.

 

[FlameTail]

Infinity cache in RDNA3 according to leaks will be on an external die, so I don't see why It should be on die in Phoenix, but L3 would be separate.

Size [MB]	Bandwidth [GB/s]	Hit rate at Full HD [%]	Actual bandwidth [GB/s]
128 MB	1987	81	1609.5
64 MB	993.5	72	715
32 MB	497	55	273
16 MB	248	37	92

Only 16MB Infinity cache for a 6WGP(24CU) IGP would be small and slow.

The best thing would be a shared external cache of 64 MB for both CPU and IGP. Size would be still small, only one additional die and you could use a cheaper process for It. Main die would save a lot of space, which could be used for more WGP, CPU cores and so on.

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).

 

[FlameTail]

(which i am sure is cheaper than standard DDR5, and you get plenty of bandwidth).

LPDDR5 debuted in phones in early 2020. It's been almost 3 years since. We are starting to see LPDDR5 even in sub-$500 midrange phones, which is indicative that the costs have certainly come down significantly.

 

[moinmoin]

Why can't AMD simply put a larger memory controller in Phoenix Point? What is stopping them?

OEMs are not demanding such a product. More memory channels increase the complexity and cost of mainboards. AMD won't create a product OEMs are not asking for.

 

[Glo.]

OEMs are not demanding such a product. More memory channels increase the complexity and cost of mainboards. AMD won't create a product OEMs are not asking for.

Increase the complexity of mainboards, but lower also the overall complexity. Higher memory bandwidth allows you to increase the performance of CPUs, GPUs and AI, and you can integrate GPU into the CPU package, much more powerful GPU should be said.

Think of it as something like M1 Pro.

 

[Exist50]

Why can't AMD simply put a larger memory controller in Phoenix Point? What is stopping them?

Cost. Let me tell you a heuristic I once heard about platform cost. $0.01 is noise, $0.10 gets grumbling, $1.00 gets meetings between the companies, and $10 gets meetings between high-level executives.

More memory channels is certainly one way to provide the bandwidth needed to feed a big GPU, but it's an incredibly delicate balance. PCs aren't limited to iGPUs like Macs are, so if you increase costs too much, OEMs might just go for a dGPU anyway, especially with AMD's current gaps vs Nvidia. There's a very narrow band of high price, high (but not too high) performance, and constrained form factor that can support such a product. I wouldn't be surprised if AMD or even Intel eventually makes one, but it's no surprise they're not quick to make those tradeoffs for a mainstream product.

Pair that with LPDDR5 (which i am sure is cheaper than standard DDR5

LPDDR is moderately more expensive than DDR.

 

[FlameTail]

Phoenix is not a mere CPU with a measly iGPU slapped in just for the sake of it. Phoenix has an iGPU so powerful it renders low-end dGPUs irrelevant.

The 6CU models compete with the likes of the Geforce MX350 and the 12 CU models compete with the likes of RTX 3050, do they not? Phoenix effectively replaces combinations like an Intel i5 + Geforce MX350 or i7 + RTX 3050...

So an OEM can choose between going for such a traditional 'CPU + Low end dGPU' or Phoenix. Thing is, those low end dGPUs have their own memory buses with their own vRAM. The MX350 has 64- bit bus and the RTX 3050 has a 128 bit bus.

You can think of it as repurposing those memory channels which would have otherwise been used for a dGPU to the APU.

Since, a normal CPU has dual-channel (128-bit);

128 + 64 = 192 bit
128 + 128 = 256 bit.

192 bit for the 6CU model and 256-bit for the 12CU model.

I think this kind of configuration actually saves OEMs money. In place of a CPU and separate dGPU, now, with this APU their functionality is combined into one chip, so this would save manufacturing costs, would it not?

 

[Exist50]

The 6CU models compete with the likes of the Geforce MX350 and the 12 CU models compete with the likes of RTX 3050, do they not? Phoenix effectively replaces combinations like an Intel i5 + Geforce MX350 or i7 + RTX 3050...

Sure, but by the time you can actually buy Phoenix, Ada-based GPUs will probably be in the market, so that's still only the low end of the stack. That's worth something, for sure, but when we start talking about even larger GPUs and more memory channels, it's a whole 'nother ball game.

I think this kind of configuration actually saves OEMs money. In place of a CPU and separate dGPU, now, with this APU their functionality is combined into one chip, so this would save manufacturing costs, would it not?

AMD would have to design a new chip/platform, which is far from cheap both for them and OEMs. There are certainly some potential advantages, but it's a very tricky product to position.

 

[TESKATLIPOKA]

Phoenix is not a mere CPU with a measly iGPU slapped in just for the sake of it. Phoenix has an iGPU so powerful it renders low-end dGPUs irrelevant.

The 6CU models compete with the likes of the Geforce MX350 and the 12 CU models compete with the likes of RTX 3050, do they not? Phoenix effectively replaces combinations like an Intel i5 + Geforce MX350 or i7 + RTX 3050...

Low end dGPUs won't become irrelevant, even if Phoenix would perform like them or better. You should realize Intel has a weak IGP, and they sell much larger volumes than AMD.

We still don't know how Phoenix IGP will really perform.
At best It will perform like the weakest 40W RTX 3050 based on leaks.
That would be very good performance for IGP, but that's It.
RTX 3050 at higher TDP will outperform Phoenix, and we still have the uncut RT 3050Ti. Let's not forget the new Ada based low-end, which should be significantly faster than GA107.

BTW performance is not the biggest problem for Phoenix.
What is the biggest disadvantage of Rembrandt laptops? The cost.
In my country, It looks like this:

Lenovo Yoga 7 - 6800U; 16GB LPDDR5; 1TB SSD; 14" OLED 1800p
Cost: €1,315

Lenovo Yoga Slim 7 Pro - 5800H; 16GB DDR4; 1TB SSD;
16" IPS 1600p; 3050 55W
Cost: €1,199
A faster laptop cost less. Phoenix most likely won't be cheaper than Rembrandt.

If I wanted to have a strong IGP+CPU then I would probably wait for Strix Point. I expect more than 8 cores and IGP should be 16CU RDNA3+.
I think we will finally see 16MB Infinity cache or some unified 32-48MB cache, which will help the IGP a lot.

 

[IntelUser2000]

LPDDR5 debuted in phones in early 2020. It's been almost 3 years since. We are starting to see LPDDR5 even in sub-$500 midrange phones, which is indicative that the costs have certainly come down significantly.

There's no doubt you can even on PCs. The Jasper Lake-based products use LPDDR4 and Gracemont based successor will likely move to LPDDR5.

The high cost is when you want to support greater than bog standard 128-bit width common in dual channel PC platforms. Then you need to double(if 256-bit) the amount of physical chips on the board, and you are also doubling the amount of wire traces on the PCB.

All that for potentially twice the upgrade in performance, which is not small, but will be a niche market when you consider outside of that the market is very very niche as you wouldn't want to pair it with a dGPU and make it more expensive than others.

Even if it's $5 more than a typical 128-bit setup, why would you use that? Then you need to take into account the beefed up iGPU on the SoC die.

Intel and AMD are merchant chip vendors, so are companies like Qualcomm. If they make chips with 256-bit memory support, then you are also selling more expensive chips.

+CPU cost(which will be greatly more expensive as you need a large GPU in addition to complex 256-bit memory support)
+Motherboard cost
+Memory chip cost

Apple sells it as a platform. You NEVER see their chips outside their branded computers. Hence the few $ increase may be seen as worth it by the company.

The rest, not so much.

(as a side this is also why I doubt we'll ever see server platform by Apple, other than vendor-locked version sold only by them)

 

[TESKATLIPOKA]

AMD would have to design a new chip/platform, which is far from cheap both for them and OEMs. There are certainly some potential advantages, but it's a very tricky product to position.

The easiest thing in my opinion would be 3d cache, maybe even in 2 sizes(32MB and 64MB) as a LLC.
This way your chip stays as small as possible and you don't need faster memory or wider bus for IGP, It will also help CPU.

edit: I forgot you can have only 2x more cache, so It could be 16MB+32MB or 32MB+64MB depending on how much you have inside die.

P.S. I don't expect such a thing for Phoenix. Strix Point is my bet.

 

[BorisTheBlade82]

@TESKATLIPOKA
For this to work AMD would need to unify cache access for CPU and GPU first. Intel is moving into the opposite direction because of disaggregation. My guess is that AMD will eventually follow suit with their Mobile lineup.

IMHO adding a unified cache by 3D stacking over/under multiple chiplets is cost prohibitive for the nearer future. Doing this with 2.5D would be rather expensive as you need a lot of bandwidth - so EFB or InFo-R to each consuming chiplet would be needed in order to keep consumption in check.
Not saying that it won't happen - but there is quite some way to go.