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 My advice to everyone: treat the old APU rumours as being out of date.

 

[Heartbreaker]

More details here and in the link, FWIW the iGPU was clocking at 2.8ghz for the 54w tdp run.
View attachment 75858
https://t.bilibili.com/758309109477933061?spm_id_from=333.999.0.0

Videocardz reporting on this link. Handy because easier to see full size images. Slower memory, so graphics scores are held back more, closer to 680M:

AMD Radeon 780M RDNA3 integrated graphics have been tested in 3DMark - VideoCardz.com

AMD Radeon 780M tested Chinese tech outlet “Golden Pig Upgrade” shares first performance figures of AMD RDNA3 integrated GPU. The final driver for integrated RDNA3 graphics is probably still a few weeks away. However, some preliminary results have now been shared on Bilibili platform featuring...

videocardz.com

 

[TESKATLIPOKA]

Erm, that would be shared between the GPU and CPU cores unless something has changed. Huge improvement over Cezanne.

I know It's shared. Against Cezanne It looks pretty good, but not against Rembrandt.
CPU part is in my opinion the interesting one, IGP is not that great so far.

 

[TESKATLIPOKA]

Videocardz reporting on this link. Handy because easier to see full size images. Slower memory, so graphics scores are held back more, closer to 680M:

AMD Radeon 780M RDNA3 integrated graphics have been tested in 3DMark - VideoCardz.com

AMD Radeon 780M tested Chinese tech outlet “Golden Pig Upgrade” shares first performance figures of AMD RDNA3 integrated GPU. The final driver for integrated RDNA3 graphics is probably still a few weeks away. However, some preliminary results have now been shared on Bilibili platform featuring...

videocardz.com

It wasn't just 54W, but 64W according to HWiNFO.
Limited to 25W the max IGP clock was 2362MHz.

 

[BorisTheBlade82]

First CBR23 score of a R7 7840HS.WIthout knowing the thermal constraints these numbers are pretty meaningless 🤷🏽‍♂️

AMD Ryzen 7 7840HS "Phoenix" 8 Core CPU Benchmarks Show Up To 25% Faster Performance Versus Ryzen 7 6800H

The first benchmarks of AMD's Ryzen 7 7840HS "Phoenix" CPU which is designed for mainstream laptops have been leaked & show a nice gain.

wccftech.com

 

[lopri]

What's the point of hiding ST score while revealing MP ratio? Am I missing something?

It comes out to be 1588 simply by dividing the MT score by MP ratio.

 

[BorisTheBlade82]

What's the point of hiding ST score while revealing MP ratio? Am I missing something?

It comes out to be 1588 simply by dividing the MT score by MP ratio.

I'd say this guy might just not have a lot of experience in the "leaking game".

 

[TESKATLIPOKA]

What's the point of hiding ST score while revealing MP ratio? Am I missing something?

It comes out to be 1588 simply by dividing the MT score by MP ratio.

And you think 1588 is correct?
Zen 4 7700 has >1900 points in ST.
Even Rembrandt scores higher than that.
That multiplier is clearly not correct.

 

[Abwx]

And you think 1588 is correct?
Zen 4 7700 has >1900 points in ST.
Even Rembrandt scores higher than that.
That multiplier is clearly not correct.

The multiplier include SMT 31% gain per thread, also MT and ST frequencies are vastly different.

Edit : Assuming 31% SMT gain the ST score within MT is 1212, also assuming that MT frequency is 3.5GHz and ST at 5.4GHz then ST score is 1212 x 54/35 = 1869 pts.

 

[TESKATLIPOKA]

The multiplier include SMT 31% gain per thread, also MT and ST frequencies are vastly different.

That doesn't matter. That MP ratio is wrong.

You can divide MT score by MP ratio to get ST score, or multiply ST score by MP ratio to get MT score.

 

[Abwx]

That doesn't matter. That MP ratio is wrong.

You can divide MT score by MP ratio to get ST score, or multiply ST score by MP ratio to get MT score.

I edited my most to explain the numbers, that seems accurate within what is at disposal.

 

[insertcarehere]

And you think 1588 is correct?
Zen 4 7700 has >1900 points in ST.
Even Rembrandt scores higher than that.
That multiplier is clearly not correct.

Trust Wccftech to post without context:

https://t.bilibili.com/759004791124787200?spm_id_from=333.999.0.0
The original leaker covered up the ST scores because he didn't think those would be in any way representative given the botched frequencies, "normal" ST scores should be in the 1800s.

 

[lopri]

What's the point of hiding ST score while revealing MP ratio? Am I missing something?

It comes out to be 1588 simply by dividing the MT score by MP ratio.

And you think 1588 is correct?
Zen 4 7700 has >1900 points in ST.
Even Rembrandt scores higher than that.
That multiplier is clearly not correct.

1588 is obviously correct for that run. I don't know why it's even a question.

Whether that run is a fair representation of the hardware is a different question. That obviously depends on many variables.

 

[TESKATLIPOKA]

What's the point of hiding ST score while revealing MP ratio? Am I missing something?

It comes out to be 1588 simply by dividing the MT score by MP ratio.

1588 is obviously correct for that run. I don't know why it's even a question.

Whether that run is a fair representation of the hardware is a different question. That obviously depends on many variables.

1588 is correct for that run, but It's not representative of the actual capability of this APU.
This is the likely reason why he hid It.
He probably didn't know (didn't notice) MP ratio could be used to calculate ST score.

 

[deasd]

Phoenix leaks got removed altogether with All_The_Watts twitter account deletion.

https://twitter.com/All_The_Watts

 

[MrTeal]

Need a Star Wars "It's still going to be 24CU, right?" meme.

 

[DisEnchantment]

Looks like UIF 1.1 is unifying all the AMD portfolio of Accelerators with one single frontend.

GitHub - amd/UIF

Contribute to amd/UIF development by creating an account on GitHub.

github.com

For now Linux Only. But the upcoming ROCm release would apparently be supporting Windows. Similarly Xilinx Run Time XRT upcoming release will support Windows as well.

 

[Thibsie]

Sometimes I wonder if the real big thing Xilinx will bring to AMD is, rather than FPGAs or network ICs, would be their software engineering.

They don't seem to have the problem AMD often has.

 

[Kepler_L2]

IIRC larger caches are only partly true for Phoenix, but only with N31/N32 parts you get the full enhanced config. N33 and Phoenix, except IIRC L0, should have the same cache sizes as RDNA2.

All RDNA3(+) products get 2x larger L0/L1 caches, N31 and N32 also get 50% larger VGPR.

 

[moinmoin]

Sometimes I wonder if the real big thing Xilinx will bring to AMD is, rather than FPGAs or network ICs, would be their software engineering.

They don't seem to have the problem AMD often has.

During the merger former Xilinx taking the leadership of essentially all of AMD's software efforts was stated to be one of the main synergies for the overall company. It's still up to see which areas this will actually cover and how it will fare in each, but this has been an exciting prospect for some time already.

 

[Thibsie]

During the merger former Xilinx taking the leadership of essentially all of AMD's software efforts was stated to be one of the main synergies for the overall company. It's still up to see which areas this will actually cover and how it will fare in each, but this has been an exciting prospect for some time already.

Oh exciting it is !

 

[IronLynx]

Hey guys, a bit lost: HX series are rebranded/fine tunned Desktop parts to use lower power?

There are some nice results with 7745HX(8C/16T) in videocardz: 1828 and 18606 points in single and multi-core using CB23.

 

[leoneazzurro]

Hey guys, a bit lost: HX series are rebranded/fine tunned Desktop parts to use lower power?

There are some nice results with 7745HX(8C/16T) in videocardz: 1828 and 18606 points in single and multi-core using CB23.

ATM there are not so many details about the implementation of these parts, but for sure we know Dragon Range uses the same Core Complex Dies as in Raphael, almost probably the same I/O die but a new socket. Dies are almost probably binned for low consumption and there may be features not enabled on desktop for further enhancing power consumption at idle.

 

[Tigerick]

AMD has updated the specs of 7040 Phoenix series by dropping all GPU speed by 200MHz. That's mean the top range 7940HS has 780M graphics of 2.8GHz speed.

ASUS has launched TUF series (called Tian Xuan 4) in China. There is one family with Ryzen 9 7940HS bundled with RTX4050,4060 and 4070. Waiting for official prices though but should be below USD1500...

 

[HoveringStyle]

That makes it 4.3 TFLOPS of FP32 right? or does the dual-issue thing change the equation? I ask because I noticed that some websites (like Techpowerup) have it at 9 TFLOPS.

 

[jpiniero]

AMD has seemingly cut the IGP frequency on the H line too. Though there really isn't any difference between the H and HS... and AMD says the H is only going to be released in China.