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 My advice to everyone: treat the old APU rumours as being out of date.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

If you want to play with any settings, then the only option is Universal X86 Tuning Utility.

As it builds upon RyzenAdj that should work as well. So it's a matter of preference if one prefers GUI or CLI.

 

[Panino Manino]

Just a simple question, how fast is a Zen 4C?
As fast as Zen 2, Zen 3?

I was wondering, why AMD still didn't make a custom APU for these Steam Deck "clones" with Zen 4C cores only?

 

[Abwx]

Just a simple question, how fast is a Zen 4C?
As fast as Zen 2, Zen 3?

As fast as a Zen 4.

 

[Panino Manino]

As fast as a Zen 4.

But the average clock is lower, so in practice it should perform more like last gen Zen 3, no?

 

[Abwx]

But the average clock is lower, so in practice it should perform more like last gen Zen 3, no?

At 3.55GHz it is supposed to perform like a Zen3@4GHz, and set apart a few 105W TDP/142W PPT desktop parts there was no Zen 3 based SKU that did boost to 4GHz with all cores, so the 3.55 max frequency is not an issue for mobile parts.

 

[trivik12]

We still dont have any laptop with 7545U phoenix 2 chip? It was announced a while ago(November if I am not wrong?)

 

[SteinFG]

Just a simple question, how fast is a Zen 4C?
As fast as Zen 2, Zen 3?

I was wondering, why AMD still didn't make a custom APU for these Steam Deck "clones" with Zen 4C cores only?

Because the price AMD (and therefore, OEMs) would pay to make it far outweighs the benefits. Tens of millions of dollars for a piece of silicon that can already be matched by Phoenix1 which is produced at scale for all laptops/desktops? Not worth it, just stick to Phoenix.

 

[Det0x]

Have pretty much completed my benchmark run with the 8700G, lots of Norwegian flags

If you guys want me to run some other cpy/memory benchmarks please let me know

Please try to test in two configurations: highest bandwidth and lowest latency.

Here are my two 24/7 daily memory configurations 👍

2:1 8800MT/s + 2550mhz FCLK


2:1 8800MT/s 2200mhz FCLK "synced"


Memory timings for both profiles are 100% stable, tested in the following:

• 40k Karhu @ 333.45mb/s
• 17.5k GB3 memoryscore
• 23 cycles testmem5 1usmus (conflict with hwinfo made it stop running)
• 1 hour y-cruncher all memtests
• Small Linpack Xtreme powercheck

 

[igor_kavinski]

If you guys want me to run some other cpy/memory benchmarks please let me know

I was trying out Street Fighter 6 benchmark on my Ice Lake CPU at the highest settings with 4 to 5 fps ( ). Unfortunately it keeps crapping out with D3D errors. Can you please try and see if it finishes and post the score?

https://www.streetfighter.com/6/benchmark/en/

And MemMaxx2 screenshots for your both daily mem configurations. Thanks!

And if you still have time to spare, 3DMark11 score (it's free in UL legacy benchmarks page) and the combined CPU/GPU benchmark score with settings cranked all the way to max.

 

[Det0x]

I was trying out Street Fighter 6 benchmark on my Ice Lake CPU at the highest settings with 4 to 5 fps ( ). Unfortunately it keeps crapping out with D3D errors. Can you please try and see if it finishes and post the score?

https://www.streetfighter.com/6/benchmark/en/

And MemMaxx2 screenshots for your both daily mem configurations. Thanks!

And if you still have time to spare, 3DMark11 score (it's free in UL legacy benchmarks page) and the combined CPU/GPU benchmark score with settings cranked all the way to max.

I'm not using the IGPU, thats why i said cpu/memory benchmarks
I can ask a mate of mine to run them tho

In the meantime, here are some more recent 3dmark's he have run on the 8500G with IGPU

 

[igor_kavinski]

I'm not using the IGPU

So it needs to be disabled for these high speeds?

 

[Det0x]

So it needs to be disabled for these high speeds?

No i can run these settings just fine with IGPU, but i have a watercooled 3090 connected... dont want the hazzle of removing it.
And i added a few pictures to my post above 👍

 

[igor_kavinski]

No i can run these settings just fine with IGPU, but i have a watercooled 3090 connected... dont want the hazzle of removing it.

You don't have to remove it by the way. Just set the iGPU as the default GPU for the SF6 and 3DMark11 executables.

 

[Abwx]

Some numbers from AMD about Hawk Point, there s other slides for AI numbers in the link but i post only the ones of interest CPU wise :

AMD Ryzen Pro: 8000er Serie soll nicht nur Intel, sondern auch Apple schlagen

AMD Ryzen Pro 8000 soll alle vom Platz fegen: Nicht nur Intel Meteor Lake, sondern auch Apples M3 Pro. Alles ist super, zeigt AMD. Oder?

www.computerbase.de

 

[igor_kavinski]

@FlameTail

Look at the battery life slide

 

[uzzi38]

Battery life there shouldn't be too surprising, Teams is the literal best case workload for AMD vs their competitors and they're comparing against M3 Pro, which does sport worse battery life than the base M3. Across an average of multiple tests Hawk Point wouldn't look that good, but I mean, it's still a solid showing that they can compete in any metric (that's not power under full load) I guess.

 

[Abwx]

Battery life say that for a same amount of work MTL used in average 11W while HPoint used 7.7W, this is due to the fact than not only AMD use a more efficient process but also MTL idle power being higher once the complete chip is fired.

This is correlated by the MSI Claw review that pointed to around 7W idling when all parts of the CPU are under use but with only the OS as load, in real usage Intel can no more use irrelevant figures to artificialy boost the battery life numbers such as local videoplay with only the two e cores running and the rest of the chip gated off.

For full load the CPU Vray test point to 3% better perf for Hawk Point@15W relatively to MTL@28W, so the figure would be no different than in the battery life slide, that is, roughly 45-50% better battery life at full load and at about same perf.

 

[Hitman928]

Anyone have the end notes to those AMD slides? There's no way Zen 4 has that much of a lead against MTL in office productivity work, all else being equal.

 

[Abwx]

Anyone have the end notes to those AMD slides? There's no way Zen 4 has that much of a lead against MTL in office productivity work, all else being equal.

These are in the Computerbase article but anyway here they are if you have the patience to read through such a wall of text :

AMD Ryzen Pro: 8000er Serie soll nicht nur Intel, sondern auch Apple schlagen

AMD Ryzen Pro 8000 soll alle vom Platz fegen: Nicht nur Intel Meteor Lake, sondern auch Apples M3 Pro. Alles ist super, zeigt AMD. Oder?

www.computerbase.de

 

[Hitman928]

These are in the Computerbase article but anyway here they are if you have the patience to read through such a wall of text :

AMD Ryzen Pro: 8000er Serie soll nicht nur Intel, sondern auch Apple schlagen

AMD Ryzen Pro 8000 soll alle vom Platz fegen: Nicht nur Intel Meteor Lake, sondern auch Apples M3 Pro. Alles ist super, zeigt AMD. Oder?

www.computerbase.de

Thanks, I don’t see anything unusual in there except that some of their MTL testing was done at launch so maybe immature firmware played a role but either way, those office tests do not represent reality today.

 

[SteinFG]

Looking at 8400F specs, it's a downgraded 8600G, not 8500G

8400F has more PCIe lanes than 8500G; it has no c cores (unlike 8500G), and has a high base clock (unlike 8500G)

On performance, it's probably on par with Ryzen 5600, but I didn't find any conclusive tests

 

[Abwx]

Looking at 8400F specs, it's a downgraded 8600G, not 8500G

8400F has more PCIe lanes than 8500G; it has no c cores (unlike 8500G), and has a high base clock (unlike 8500G)

On performance, it's probably on par with Ryzen 5600, but I didn't find any conclusive tests

It should be 14% faster than a 5600X and match a 5700X.

AMD Ryzen 8700G & 8600G im Test: Gaming, Multimedia, Apps: App-Benchmarks, Multimedia und Stromverbrauch

Ryzen 7 8700G & 5 8600G (OC) im Test: App-Benchmarks, Multimedia und Stromverbrauch / Testmethodik für Anwendungen

www.computerbase.de

 

[marees]

The follow up for Mendocino is Sonoma Valley, it's already been rumoured thanks to a leaked diagram at some point back around the time rumours for Mendocino started popping up.

As for what it is though, not much is publicly known for now.

As per leaks, sonoma valley is 4x zen 5c on samsung 4nm

Anyone knows the gpu arch for the same ? (Has to be RDNA 3.5 or RDNA 2 would be my guess)

 

[marees]

As per leaks, sonoma valley is 4x zen 5c on samsung 4nm

Anyone knows the gpu arch for the same ? (Has to be RDNA 3.5 or RDNA 2 would be my guess)

As per this post it is 4 CU RDNA 3.5

Highly Confident:

When Reuter reported NV and AMD are going to launch ARM SoC for PC:


My first reaction is NV yeah possible, AMD no way. With full lineup of mobile APU coming namely Sonoma Valley, Kraken Point, Strix Point and Sarlak. What is the point? All the above SoCs almost covers all the performance (from 64-bit to 256-bit LPDDR5x) and price points.

However, when I started a thread for upcoming LPDDR6 here, I think AMD might create new lineup for ARM SoC with LPDDR6, more info below:

I contacted my source, turned up there might have some truths there. Apparently, MS Surface which presumably will upgrade Surface Pro X to Snapdragon X Elite this year. Next year 2025 and later, NV and AMD have won the contract for upcoming Surface X series. PS: And I believe here is the source of Reuter's leaks. If Microsoft going to launch new Surface in 2025, the specs must be better than upcoming Surface Pro X w/ X Elite.......or not? See below:



Highly Speculation

My source insisted that AMD will create custom off-shelf ARM core for Microsoft. They won't be used on standard PC lineup, I actually have different opinion. Below is my speculated lineup:

Let's compare Hawk Point APU with my speculated ARM SoC with 64-bit memory bus:

	Hawk Point	ARM 64-BIT MC	X Elite	Apple's M3	Lunar Lake
Node	N4	?	N4P	N3B	N3B + N6
Die Size	178 mm2	?	~171 mm2	~146 mm2	?
Memory Supported	128-bit LPDDR5X-7500	64-bit LPDDR6-12800	128-bit LPDDR5X-8533	128-bit LPDDR5-6400	128-bit LPDDR5X-8533
Memory BW	120 GB/s	102 GB/s	136 GB/s	100 GB/s	136 GB/s
CPU	Zen 4 8P/16T	Cortex 6-8 P?	12P Oryon	4P + 4E	4P + 4E
GPU	RDNA3 12CU	RDNA 8CU ?	Adreno	1280 ALU	ARC 8XE
FP32	4.3 TF	~ 4 TF	4.6 TF	4.1 TF	3.8 TF
AIE (TOPS)	16	?	45	18	45

• I don't know how the core arrangement be. Let's assume 6-8 cores with RDNA 12CU.
• The memory bandwidth of 64-bit LPDDR6 is a bit slower than Hawk Point, that's why I estimate core count to between 6-8 cores. Remember Zen4 comes with HT while ARM likely won't support HT.
• And why don't AMD create x86 version of APU to fit the same memory bandwidth? Cause Hawk Point exist, AMD already has full lineup of solution from top to bottom. BTW, Kraken Point and Sonoma Valley will replace Hawk Point in 2026 for mobile market. Thus, there is no point to create another x86 SoC with similar bandwidth. And that's my reason why AMD would use ARM core to create new lineup with LPDDR6. Also, ARM's die area is smaller than Zen 5 (pending confirmation) and definitely more power efficient. It is possible AMD will support 9W fanless design with ARM SoC.

	Strix Point	ARM 128-BIT MC	Sarlak6 256-BIT
Node	N4P	N3P ?	N3P + N3E
Die Size	225 mm2	?	?
Memory Supported	128-bit LPDDR5X-8533	128-bit LPDDR6-12800	256-bit LPDDR6-12800
Memory BW	136 GB/s	205 GB/s	410 GB/s
CPU	4 x Zen5 + 8 x Zen5c	12-16 P ?	16 x Zen6
GPU	RDNA3.5 16CU	RDNA 24CU ?	RDNA5? 40CU + 32MB IC
FP32	~ 6.1 TF	~ 9.2 TF	?
AIE (TOPS)	48	?	48

View attachment 94158

Speculated Mobile Roadmap		Memory Bus	2026 DDR5/LPDDR5		2026 LPDDR6	Memory Bus
Ultimate Compute		128-bit DDR5	Zen 6 / N3P / 16C + X3D
Elite Experiences		256-bit LPDDR5X 192-bit LPDDR5X	Sarlak Zen 5? / N4P + N3E / 16C + RDNA3.5 40CU + 50 TOPS		Sarlak6 Zen 6 / N3P + N3E / 16C + RDNA5? 40CU + >50 TOPS	256-bit LPDDR6 192-bit LPDDR6
Premium		128-bit LPDDR5X	Strix Point Zen 5 / N4P / 12C + RDNA3.5 16CU + 50 TOPS		ARM 128-bitMC / 16P + RDNA 24CU + >50 TOPS ?	128-bit LPDDR6
Mainstream		128-bit LPDDR5X	Kraken Point Zen 5 / N4P / 8C + RDNA3.5 8CU + TOPS		ARM 64-bitMC / 10c + RDNA 12CU + >50 TOPS ?	64-bit LPDDR6
Everyday		64-bit LPDDR5X	Sonoma Valley Zen 5c / SF4X / 4C + RDNA3.5 4CU

Highly Confident

• Zen 6 is pretty much confirmed to use N3P and feature 16-core into one die area. And I believe 16-core Zen6 will pair with Sarlak iGPU to utilize LPDDR6 with higher memory bandwidth.
• 32-core Zen 6C (that's the term AMD refer) on N2 process. There's no difference in the architecture, only in the libraries and floorplan used. And Zen 6C would be used to counter upcoming NovaLake-S.
• Since DDR6 is delayed, NVL-S will support 128-bit DDR5. Based on timings, it should support up to DDR5-8000. And that might be the speed AMD would use to support Zen 6C, that's mean AM5 platform will live for another generation.

Model	Zen 4 Cores	Speed	TDP		GB6 1T	Perf Per Clock		GB6 MT	Perf Per Core
R9 7950X	16	5.7 GHz	170 W		2937	515
R7 7700X	8	5.4 GHz	105 W		2906	538
R7 7745HX	8	5.1 GHz	55 W		2686	527
R9 7940HS	8	5.2 GHz	35 - 54 W		2475	475		11670	1459
R7 7840U	8	5.1 GHz	28 W		2102	412		8762	1095
	Zen 5 Cores
R9 8950H ?	4P + 8E		35 W
R7 8950U ?	4P + 8E
Apple's M3	4P + 4E	4.06 GHz	20 W		3084	760		11564	1445
M2	4P + 4E	3.49 GHz	20 W		2634	755		9753	1219
X Elite	12P	4 GHz	23 W		2780	695		14029	1169

	Cortex-X3	Cortex-X4	Cortex-X5	Cortex-X6
SoC Launch Date	Q4 2022	Q4 2023	Q4 2024	Q4 2025
SoC	SD 8 Gen 2	SD 8 Gen 3	D9400	D9500
Node	N4	N4P	N3E	N3P
Clock Speed	2.0 - 3.2 GHz	3.3 GHz
L2 Cache	1 MB	2 MB ?
L3 Cache	8 MB	12 MB
ROB	320	384
Decode	6	10
Int ALU	6	8
Max Memory Speed	64-bit LPDDR5X-8400	64-bit LPDDR5X-9600		64-bit LPDDR6-12800 ?
Memory BW	67 GB/s	77 GB/s		102 GB/s
GB6 1T	~ 2043	~ 2277	~ 2700 ?

 

[SteinFG]

As per this post it is 4 CU RDNA 3.5

That's also just guesswork. It's most likely still 2 CUs:
1) 64-bit LPDDR5 bus is vary narrow, you won't game on it.
2) It's literally bottom of the barrel part (Samsung node, all C cores, 4 PCIe lanes). Anything that can be cut will be cut.