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 My advice to everyone: treat the old APU rumours as being out of date.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

Again you've demonstrated your poor reading comprehension.

View attachment 73977

The 9885 score is on the first run.

View attachment 73978

Presumably this is also the first run.

So, yeah ~36% performance advantage for Lenovo.

AMD always uses gimped configurations for the competitor when it comes to laptop comparisons. Earlier in last year's CES, they used the 12 W MX450 to compare against the 680M iGPU.

Pro tip, before you mouth off on other people, look at your house first.

Here's another tip, 1 data point does not a conclusion make. And like someone else said, laptop hardware is incredibly finnicky and hard to apples-to-apples bench to begin with.

In your confirmation bias bonanza you forgot to see the forest for the trees... we also have M1 Pro data in the same slide. And you know... there's a ton of consistent data on M1 Pro performance.

M1 Pro - ~12400 CB R23 nT
1280P - ~11500 as median
If 1280P/11500 is the baseline, then 8% over is exactly around 12400 for the M1 Pro (30W). Which would put the 7940HS at ~15600, presumably running at typical out of the box PL values, 35/42W or 45/54W.

But you'd rather believe first that AMD is so misleading that Phoneix/Zen4 actually performs identical to Rembrandt/Zen3+ somehow, before you'd venture to think "maybe I'm wrong about some things."

And now, circling back to the "your house" part, you basically do nothing but complain, nitpick and bash everytime you "pitch in" on a AMD thread, and you're not even right more than half the time.


So maybe sit the name calling and frothing at the mouth behaviour part out, eh?

 

[TESKATLIPOKA]

Where is that stated that it wasn't IGP?

According to note 7, it's AMD integrated 780M vs Intel Iris Xe

AMD Extends its Leadership with the Introduction of its Broadest Portfolio of High-Performance PC Products for Mobile and Desktop

— AMD Ryzen™ 7000X3D Series Desktop processors bring the power of AMD 3D V-Cache™ technology to gamers and creators — — AMD Ryzen 7000…...

ir.amd.com

I was definitely hoping for more than 20%.

Isn't that actually a mistake or something? The same as 40MB cache?

I just checked that HP laptop vs R7 6800U and IGP performance looks like this:
Scores are for high setting at Full HD.

	HP EliteBook 840 G9 i7-1280P	Lenovo Yoga Slim 9i i7-1280P	Lenovo Yoga 7 14 G7 R7 6800U
Cyberpunk 2077	13 FPS (100%)	---------------	19 FPS (146%)
Grand Theft Auto V	23.4 FPS (100%)	27.1 FPS (116%)	49.5 FPS (146%)
Strange Brigade	35.9 FPS (100%)	37.9 FPS (106%)	41.8 FPS (116%)

We can't make any conclusions based on this, because I don't know how performance looks in the rest of games mentioned in that endnote, but It would be pretty weird If AMD choose those where the difference is the smallest.
Still, in these 3 games Phoenix should perform better than 6800U at least because It has higher IGP frequency and TDP.

 

[AtenRa]

Well we dont who at what TDP was that 780M performance vs G7.

 

[TESKATLIPOKA]

Well we dont who at what TDP was that 780M performance vs G7.

AMD should want the best performance, so I am thinking It was 45W.

 

[Glo.]

AMD should want the best performance, so I am thinking It was 45W.

It doesn't matter at what TDP they tested it.

If the Footnotes are correct, Phoenix is 40% slower than Rembrandt in iGPU performance.

 

[TESKATLIPOKA]

It doesn't matter at what TDP they tested it.

If the Footnotes are correct, Phoenix is 40% slower than Rembrandt in iGPU performance.

You don't really know how Rembrandt performs in those 12 games.
I found results for only 3 games from that list. 2 were 46% and one only 16% faster for 6800U. Average is 36%, but there are 9 games missing.
Don't take me wrong, I also find that increase very low, but we actually don't know how Rembrandt performs in those games to make a comparison.

 

[Kaluan]

Well we dont who at what TDP was that 780M performance vs G7.

Possibly 15W, same for both configurations. If the endnote was about that slide.

Either way, I think it's more likely it's a mistake and the numbers were for +eGPU or something.

 

[TESKATLIPOKA]

Possibly 15W, same for both configurations. If the endnote was about that slide.

Either way, I think it's more likely it's a mistake and the numbers were for +eGPU or something.

I have a similar opinion about It being about dGPU, although that endnote being wrong is not impossible, It would be just sad making such a big mistake.
I don't see a reason why would they mention IGP performance when they said practically nothing about the IGP of Phoenix.

 

[Abwx]

View attachment 73995

Go on. Try to make yourself look even more stupid than you really are.

From NBC in the same article :

35 watts long-term (PL1) in the Lenovo Yoga Slim 9i and, for 40 seconds, even 52 watts (PL2). In intelligent cooling mode, PL2 is even set to 64 watts but only maintained for 15 seconds in that case.

64W, in intelligent mode of course, for whom want to be on point, others will keep polluting this thread with their incompetence....

 

[Exist50]

From NBC in the same article :




64W, in intelligent mode of course, for whom want to be on point, others will keep polluting this thread with their incompetence....

They explicitly said that wasn't the testing mode they used.

 

[Abwx]

They explicitly said that wasn't the testing mode they used.

They explicitely stated this :

All performance-related benchmarks we conducted with the highest performance mode enabled.

 

[TESKATLIPOKA]

@tamz_msc @Abwx @Exist50
Why are you still debating about those power limits? What's the point?
We already know a chiplet based 8C16T Raphael at 45W performs better than i7-1280P. A monolithic Phoenix at better process shouldn't perform worse even If It has only 16MB L3 cache. Phoenix will be faster, not by 34% but still faster.

edit: If that's not enough then Apple M1 Pro scores 12370 points and Phoenix is 24% faster, meaning we are talking about 15339 points! This would put It ahead of i7-1280P in that Lenovo by 13.6%. @Kaluan already wrote the same thing.

 

[desrever]

I think IGP performance is going to be limited by ram until they add infinity cache. Likely we will only see APUs in the very low end in the future with chiplets advancement. We will likely see them attaching a GPU chiplet to a CPU rather than having these large IGPUs integrated into 1 chip.

780M probably not much faster than 680M. The 680M isn't even that much faster than 660M despite having 2x the cores enabled and higher clocked.

 

[Glo.]

I think IGP performance is going to be limited by ram until they add infinity cache. Likely we will only see APUs in the very low end in the future with chiplets advancement. We will likely see them attaching a GPU chiplet to a CPU rather than having these large IGPUs integrated into 1 chip.

780M probably not much faster than 680M. The 680M isn't even that much faster than 660M despite having 2x the cores enabled and higher clocked.

The logic will be monolithic.

Chiplet based: Infinity Cache+Memory Controller, just like in Navi 31.

 

[Exist50]

They explicitely stated this :

"Highest performance" is a specific power profile in the chart linked for you, with 35W/52W PL1/PL2. I fail to see the ambiguity.

 

[Abwx]

@tamz_msc @Abwx @Exist50
Why are you still debating about those power limits? What's the point?
We already know a chiplet based 8C16T Raphael at 45W performs better than i7-1280P. A monolithic Phoenix at better process shouldn't perform worse even If It has only 16MB L3 cache. Phoenix will be faster, not by 34% but still faster.

edit: If that's not enough then Apple M1 Pro scores 12370 points and Phoenix is 24% faster, meaning we are talking about 15339 points! This would put It ahead of i7-1280P in that Lenovo by 13.6%. @Kaluan already wrote the same thing.

My point was that the 7040 is 2x more efficient than a 1280P at same throughput, even if the Lenovo one use only 52W this would put the 7040 at 2.17x the perf/watt assuming the 7040 is using 35W in AMD s tests.

We ll soon get some better numbers but seems that there s people who foolishly believe that Intel s offering could be competitive in thoses conditions, hence some desperation from the usual suspects...

 

[Heartbreaker]

I think IGP performance is going to be limited by ram until they add infinity cache. Likely we will only see APUs in the very low end in the future with chiplets advancement. We will likely see them attaching a GPU chiplet to a CPU rather than having these large IGPUs integrated into 1 chip.

780M probably not much faster than 680M. The 680M isn't even that much faster than 660M despite having 2x the cores enabled and higher clocked.

Sadly yeah. Bandwidth remains the Achilles heel.

 

[insertcarehere]

Pretty sure those numbers are analogus to these from the 7045 chips:
View attachment 73985

If they were showing more Thin & Light specific stuff (they weren't, since 7040U is MIA for now), I think they would've actually covered the iGPU.
But something also makes me think they're waiting on the driver team here. If you look at whatever endnotes are available, a lot of the Phoneix testing dates as far back as August last year, I doubt they had any sort of well working RDNA3 drivers back then.

The 7045 series is Dragon Point, which uses the same chip as Raphael, so it is extremely unlikely they were talking about iGPU performance in that slide.

And like, this is the testing notes for the iGPU :

The new AMD Ryzen 7040HS Series Mobile processors offer:

Up to 34% faster multithreaded performance over the competition9
Up to 21% faster gaming performance over the competition7
7 - Based on testing by AMD as of 12/23/2022. Testing results demonstrated in Borderlands 3, Cyberpunk 2077, Rainbow Six Siege, Assassin's Creed: Valhalla, World of Tanks Encore, League of Legends, Far Cry 6, Grand Theft Auto V, Shadow of the Tomb Raider, F1 2021, Strange Brigade, Total War: Three Kingdoms Battle. Ryzen™ 9 7940HS system: AMD reference motherboard configured with 4x4GB LPDDR5, Samsung 980 Pro 1TB SSD, Radeon 780M Graphics, Windows® 11 64-bit. Core i7-1280P system: HP Elitebook 840 G9 configured with 16GB DDR5-4800, 1TB SSD, Intel Iris Xe, Windows 11 64-bit. System manufacturers may vary configurations, yielding different results. PHX-9

Seems pretty unambiguous that they are comparing iGPU vs iGPU and tested it recently enough. I doubt AMD would have bothered showing performance here if the drivers were still broken.

 

[tamz_msc]

Pro tip, before you mouth off on other people, look at your house first.

Here's another tip, 1 data point does not a conclusion make. And like someone else said, laptop hardware is incredibly finnicky and hard to apples-to-apples bench to begin with.

In your confirmation bias bonanza you forgot to see the forest for the trees... we also have M1 Pro data in the same slide. And you know... there's a ton of consistent data on M1 Pro performance.

M1 Pro - ~12400 CB R23 nT
1280P - ~11500 as median
If 1280P/11500 is the baseline, then 8% over is exactly around 12400 for the M1 Pro (30W). Which would put the 7940HS at ~15600, presumably running at typical out of the box PL values, 35/42W or 45/54W.

But you'd rather believe first that AMD is so misleading that Phoneix/Zen4 actually performs identical to Rembrandt/Zen3+ somehow, before you'd venture to think "maybe I'm wrong about some things."

And now, circling back to the "your house" part, you basically do nothing but complain, nitpick and bash everytime you "pitch in" on a AMD thread, and you're not even right more than half the time.


So maybe sit the name calling and frothing at the mouth behaviour part out, eh?

Before you have your own shot at defending AMD, carefully read through their press material and slide decks to be absolutely certain that you are defending something that is actually defensible.

Here is the slide in question at CES:

Clearly, it is stated that 7940HS is up to 34% faster than the 1280P configuration in multiprocessing, that is Cinebench R23 nT, right?

Now look at the press release:

AMD Extends its Leadership with the Introduction of its Broadest Portfolio of High-Performance PC Products for Mobile and Desktop

— AMD Ryzen™ 7000X3D Series Desktop processors bring the power of AMD 3D V-Cache™ technology to gamers and creators — — AMD Ryzen 7000…...

ir.amd.com

Footnote 9 states:



So the press release on the other hand states that the "up to 34% faster multithreaded performance" is obtained with a composite test suite consisting of the aforementioned apps.

Now which do you believe is true? Would you now modify your defense of AMD's data presentation accordingly?

 

[Tigerick]

View attachment 70476

Holy Mackerel, density in N31 GCD based on N5 is around 132 MTr/mm2, just about the same as an older gen iPhone SoC, even higher than AD102 on N4.

If this is what AMD is going for in Phoenix, they can fit 18 Billion transistors in less than 140 mm2. ( 6000 series@13.1 Billion XTor and 5000 series@10.7 Billion XTor)
Even a more conservative 120 MTr/mm2 would make an 18 Billion transistor chip come in at 150 mm2.
Not even considering if they might gain additional density from N4 migration too.

Bean counters must be twiddling their thumbs now waiting for the numbers to come in.

Turn out Phoenix has die area of 178mm2 with 25 B, pretty impressive.

I have a question, since you mentioned about opportunistic dual Issue of N33/Phoenix; can you explain what is the difference between opportunistic dual Issue of N33 and dual issue of N31?

 

[FlameTail]

Thanks for sharing with us that for you even an older SoC with ~409.6 GFLOPs is good enough.

Here is something more usable for comparison purposes. Got It from notebookcheck.

	Mali-G610 MP6	Xclipse 920	Mali-G710 MP10	Adreno 740	A16 GPU 5-Core	Rembrandt 680M	RTX 3050 Ti 45W
3DMark Wild Life Unlimited	5632	7890	9770	14023	12429	16072
3DMark Wild Life Extreme Unlimited	1366	1935	2650	3733	3265	3960	9340

The problem with this test is that It runs only for 1 minute, but in laptop these SoCs wouldn't be limited by heat dissipation.

Adreno 740 is very close to Rembrandt in this test. I checked Its specs, and It turns out It has 2560ALU vs 768ALU in 680M, but of course clockspeed is only 680-719MHz. TFlops are very similar for both of them.
Phoenix will be faster than this, and I am not sure Qualcomm will come with another significantly more powerful SoC this year.

This shows Qualcomm has a pretty powerful SoC Snapdragon 8 Gen 2, so where are laptops with this SoC?
The only one with ARM + Windows are Samsung Galaxy Book S or Microsoft Surface, and they have a weaker SoC, compatibility with x86 APPs is still a serious problem.
Here is the latest Microsoft laptop with ARM review. PCmag
Until they fix the compatibility issues and actually release some powerful Laptops, AMD or Intel don't need to worry.
Unless I can play my Steam library on an ARM chip, this is for me a dead platform and It won't matter how powerful It is.

2560 ALU in Adreno 740? Source? That's unbelievable.

Many leakers were saying 1536 ALU in Adreno 740, up from 1024 ALU in Adreno 730.

 

[Glo.]

If I calculate this correctly, Phoenix has 140 mln xTors/mm2 density.

 

[Kaluan]

Before you have your own shot at defending AMD, carefully read through their press material and slide decks to be absolutely certain that you are defending something that is actually defensible.

Here is the slide in question at CES:

Clearly, it is stated that 7940HS is up to 34% faster than the 1280P configuration in multiprocessing, that is Cinebench R23 nT, right?

Now look at the press release:

AMD Extends its Leadership with the Introduction of its Broadest Portfolio of High-Performance PC Products for Mobile and Desktop

— AMD Ryzen™ 7000X3D Series Desktop processors bring the power of AMD 3D V-Cache™ technology to gamers and creators — — AMD Ryzen 7000…...

ir.amd.com

View attachment 74036

Footnote 9 states:

View attachment 74037

So the press release on the other hand states that the "up to 34% faster multithreaded performance" is obtained with a composite test suite consisting of the aforementioned apps.

Now which do you believe is true? Would you now modify your defense of AMD's data presentation accordingly?

All i see is you conveniently ignoring the M1 Pro data and lecturing others about bias. Again.

Not to mention, you started all this nonsense debate in the first place. Own it and get a grip.

I'm not interested in derailing this thread over pointless minutiae. But you do you I suppose.

 

[tamz_msc]

All i see is you conveniently ignoring the M1 Pro data and lecturing others about bias. Again.

Not to mention, you started all this nonsense debate in the first place. Own it and get a grip.

I'm not interested in derailing this thread over pointless minutiae. But you do you I suppose.

M1 Pro data is irrelevant to the question whether AMD's presentation and data is free of errors. It isn't, and therefore any conclusion arrived at with said "data" is meaningless. Like you did when you replied to me previously. Oh, and user Abwx is 100% pro-AMD biased. That's irrefutable truth.

These forums are funny - pointing out glaring shortcomings with AMD press data is termed as fighting over "pointless minutiae".

 

[TESKATLIPOKA]

M1 Pro data is irrelevant to the question whether AMD's presentation and data is free of errors. It isn't, and therefore any conclusion arrived at with said "data" is meaningless. Like you did when you replied to me previously. Oh, and user Abwx is 100% pro-AMD biased. That's irrefutable truth.

These forums are funny - pointing out glaring shortcomings with AMD press data is termed as fighting over "pointless minutiae".

That chart is more than enough to know the performance they measured.
CB R23 nT
7940HS - 134% -> 12370/108*134 = 15348 points
M1 Pro - 108% -> This is a known score: 12370 points
1280P - 100% -> 12370/108*100 = 11453 points

Want to claim that chart is wrong or that press release? Ok, then give us some proof. Just because there are many mistakes, doesn't mean this is also wrong.

edit:
You should really pay more attention to what you use as proof.

That footnote 9 is only about AMD and Apple. Phoenix has 34% higher performance on average in those 5 tests than M1 Pro.




And If you looked hard at that chart Intel vs Apple vs AMD, then you would notice Phoenix is only 24%(134/108) faster in CB R23 nT than M1 Pro.
So both of them look correct.