Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[Win2012R2]

Anybody seen 3d chiplet clocks for 9950x3d? AMD's site just show usual "up to 5.7", without breaking down by chiplet type.

 

[gdansk]

Anybody seen 3d chiplet clocks for 9950x3d? AMD's site just show usual "up to 5.7", without breaking down by chiplet type.

They never listed them for the 7950X3D either. Not sure if incompetent or deliberately misleading. For AMD either seems plausible.

 

[Win2012R2]

Not good look, base is up only 4.2 to 4.3, guess will have to wait for testing - from AMDs result does not feel like 3d chiplet is fast enough for scheduling to be not a problem, and if it again comes to disabling one chiplet then might as well get 9800x3d - unless they give higher clocks for 3rd chiplet, one would have thought it would be in their interest to do so and sell more of 9950x3ds...

 

[MS_AT]

https://www.nvidia.com/en-us/project-digits/ seems like a direct competitor for Halo in compact desktop workstations. Still the availability of Halo in only 3 products, 2 business focused doesn't give me hopes I will be able to buy one before they will become obsolete.

Do I get it right, that Z2 Extreme is based on cut down Strix Point (3 big cores, 5 dense cores) and Krackan gets a new die (4 big cores, 4 dense cores)?

 

[gdansk]

Do I get it right, that Z2 Extreme is based on cut down Strix Point (3 big cores, 5 dense cores) and Krackan gets a new die (4 big cores, 4 dense cores)?

Oh you're right. It's 4+4 and Z2E is a weird Strix Point chop.

 

[igor_kavinski]

Still the availability of Halo in only 3 products, 2 business focused doesn't give me hopes I will be able to buy one before they will become obsolete.

You will get your chance when Minisforum releases it for $699 or less.

 

[eek2121]

valve representative just denied this, no z2 steam deck

Did you expect them to admit this? lol! Of course they did. They are working on new hardware.

They never listed them for the 7950X3D either. Not sure if incompetent or deliberately misleading. For AMD either seems plausible.

From what I gather, the boost clocks are 5.7 ghz for the 3D chiplet. This is from the rumor mill, however, so we will see at launch.

 

[HurleyBird]

https://www.nvidia.com/en-us/project-digits/ seems like a direct competitor for Halo in compact desktop workstations. Still the availability of Halo in only 3 products, 2 business focused doesn't give me hopes I will be able to buy one before they will become obsolete.

Everything being equal, the Nvidia solution will be appealing to a more pro audience, while the AMD solution will be appealing to a more amateur audience. Everyone would rather have the Nvidia GPU given the choice, but one is also a broadly useful x86 computer and the other is an ARM thing.

I imagine a lot will come down to the price difference between 128GB SKUs.

 

[yuri69]

Still the availability of Halo in only 3 products, 2 business focused doesn't give me hopes I will be able to buy one before they will become obsolete.

IMO this is still better than expected. I expected they would cancel it.

 

[StefanR5R]

[AMD Ryzen AI Max PRO 380... OEM-only entry part]

I am unaware of AMD disabling any L3 cache on a chiplet since Zen 3 debuted in 2020

They have done it before in some EPYC parts. Examples:
EPYC Milan 7453 has got 4 CCXs (on 4 CCDs) with 7 cores/CCX and 16 MB L3$/CCX.
EPYC Genoa 9224 has got 4 CCXs (on 4 CCDs) with 6 cores/CCX and 16 MB L3$/CCX.
I am not aware of an EPYC Turin 9005 being cut down this way, at this time.

 

[StefanR5R]

Zen 5c has worse efficiency at the same clock rate

Perhaps you have been misled by measurements which did not show per core power but something else. Furthermore, 3rd party measurements of Zen 5c have only happened for Strix Point yet, on which such measurements need to take the large difference of last-level cache size per core between classic and dense cores into account.

They're worse per AMD's own statements at ISSCC 2024.

This area optimization enables “Zen 4c” to deliver more than 25% improvement in performance/mm2, and 9% improvement in performance/W on SPECrate®2017_int_base [3], as compared to “Zen 4.” “Zen 4c” can operate up to 3.1GHz in frequency in a server configuration, while delivering more power-efficient performance.

Pretty sure that wasn't iso $$.

Pretty sure there is just some solidified confusion on his part.

AMD just can't do 4+4 like everyone else.

Nevermind that they can. Yet when "everyone else" is doing 4+4 it's 4x large µarch + 4x small µarch, while when AMD is doing either 4+4 or 3+5, it's 8x large µarch.

 

[GTracing]

any list showing the laptops with HALO and FIRE RANGE 9955hx? the models matter

only found a few

HALO: Asus rog flow x13 , which else?

9955HX: Asus rog Strix G18 , MSI vector A18 , both with nvidia 5000 blackwell
both models are mid

Zephyrus G14 is Strix Point 370HX + nvidia 5000

way too many arrow lake laptops......

They showed three Strix Halo devices. I don't get why they'd put it in a miniPC.

 

[Abwx]

They showed three Strix Halo devices. I don't get why they'd put it in a miniPC.

View attachment 114396

Possibly that they gave a one or two months exclusivity to Asus for gaming laptops
and HP for whatever WS usage, they did the same with Strix Point whch was Asus exclusive for one month, and previously for Phoenix with Lenovo as exclusive for 1-2 months.

 

[igor_kavinski]

I don't get why they'd put it in a miniPC.

Much easier to cool that way and it can run at full power without any heat related impediments.

 

[Abwx]

Idle power of Strix Halo look quite good given the gigantic SoC design, HP state 5.32W for the full laptop with IPS screen.

the internal battery measures 74.5 Wh.
HP promises up to 14 hours of battery life, though this probably only applies to the IPS option.

Small powerhouse with Strix Halo: HP ZBook Ultra 14 G1a launches with Ryzen AI Max Pro

Strix Halo goes ZBook: HP is one of the launch partners of AMD for its new Ryzen AI Max Platform. The new HP ZBook Ultra 14 G1a is a compact workstation that may compete well against the Apple MacBook Pro 14.

www.notebookcheck.net

 

[Josh128]

Idle power of Strix Halo look quite good given the gigantic SoC design, HP state 5.32W for the full laptop with IPS screen.

Small powerhouse with Strix Halo: HP ZBook Ultra 14 G1a launches with Ryzen AI Max Pro

Strix Halo goes ZBook: HP is one of the launch partners of AMD for its new Ryzen AI Max Platform. The new HP ZBook Ultra 14 G1a is a compact workstation that may compete well against the Apple MacBook Pro 14.

www.notebookcheck.net

Yeah, it was part of the Halo design. Supposedly it has 2 low power Zen 5 cores IN THE GPUIOD for this purpose, according to past rumored specs. It was needed. Zen 5 dual CCD desktop has terrible idle consumption. They couldnt do that with a serious mobile dual CCD setup.

 

[Josh128]

I guess I don't understand. The URL is amazon, how could it be a fraud ??

Remember my 9900X for $309 post the other day? Same thing happened to me until someone here pointed it out. Just searched 9900X in Amazon at it came up. The listing disappeared in less than an hour and came up as $439 shipped by Amazon, lol.

 

[MS_AT]

I don't get why they'd put it in a miniPC

I personally find miniPC form factor the most interesting for Halo. Since it has efficiency enhancements and seems to be capable of running inference of usable LLMs at usable speed, it sounds like a good candidate for efficient home server (if you would like to run LLMs locally).

On the other hand using it on this 13inch tablet capped to 32GB of RAM seems like a wasted opportunity.

 

[Kolifloro]

I can understand why Zen 5 is on N4 , but ... ... from my infinite ignorance ... I wouldn't have thought RDNA4 would ALSO be manufactured on 4 nanometers ...

GPUs ... those power hungry devices. It would make perfect sense to miniaturize them to contain those ultra high power demands ...

I wonder whether next AMD GPU iteration will be manufactured on 2 nanometers ...

Would it be ready for Medusa integrated GPUs ?????

 

[gdansk]

Pretty sure there is just some solidified confusion on his part.

Spoiler

It's pretty much as I remember?

 

[jpiniero]

I can understand why Zen 5 is on N4 , but ... ... from my infinite ignorance ... I wouldn't have thought RDNA4 would ALSO be manufactured on 4 nanometers ...

GPUs ... those power hungry devices. It would make perfect sense to miniaturize them to contain those ultra high power demands ...

$$$$$$$$

 

[Joe NYC]

Prices are made by the OEMs and retailers, and so far there s was no competition in the Strix Point segment, that will change with Intel releasing some mobile SKUs.

As for KRK they are branding it as entreprise and everyday users dedicated, that s not exactly a high price segment, and the much cut down GPU is an indication that it s not a HPoint replacement, at the start itshould be in 800-1200€ laptops as long as the Zen 4 mobile inventory has not dried up, wich will take at least one year.

AMD Krackan (Point): Günstigere Langläufer-APUs mit Zen 5(c), RDNA 3.5 & 50 TOPS

Unter der Strix-Point/Halo-Familie ist noch Platz: AMD Ryzen AI 7 350 und Ryzen AI 5 340 debütieren hier, Codename Krackan (Point).

www.computerbase.de

BTW, looking at the summary table at computerbase, specifically cache sizes (L2+L3) being 24 MB for Kraken, of which, 8 MB is L2, implies 16 MB L3.

Which hopefully means unified, single ring bus serving both 4c full Zen 5 and 4c of dense Zen 5c.

(It could still be 8+8, which would really stink...)

 

[StefanR5R]

It's pretty much as I remember?

Here is a graph directly for Zen 4 APUs, comparing a 6c/12t Phoenix 2 with a 6c/1t Phoenix, made by AMD:

Spoiler

source: https://www.computerbase.de/2023-11...-setzt-nun-auch-auf-die-kleinen-zen-4c-kerne/

So you are right as far as that there is a cross-over of power-efficiency at about 18 or 19 W, above which the classic-cores-only Ryzen 5 7540U starts to perform slightly more efficient than the 2 classic + 4 dense cores of Ryzen 5 7545U (in Cinebench 23, that is). Default TDPs of both parts are 15 W¹, I don't know whether that's also their default PPT limit.

________
¹) edit, my info on default TDP was incorrect, see below

 

[gdansk]

Default TDPs of both parts are 15 W.

That's not true, it's 28W. Which is what vendors also generally ship. Of course, you can change it so you're still not stuck with it.
But overall it was confusion on my part - AMD did go 4+4 like everyone else and so it should be fine having 4x Zen5c cores.

 

[Joe NYC]

Did you expect them to admit this? lol! Of course they did. They are working on new hardware.

From what I gather, the boost clocks are 5.7 ghz for the 3D chiplet. This is from the rumor mill, however, so we will see at launch.

That would be quite out of this world, if true. It would, theoretically, outperform 9800x3d by nearly 10%, based on clock speeds.

I am a little skeptical. I think AMD would mention it / demonstrate it.