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

[DisEnchantment]

Speculate at will

 

[igor_kavinski]

@StinkyPinky , are you able to put your NPU to good use with this? https://www.asus.com/proart/software-solutions/musetree/

 

[Hitman928]

Unless it's an M3 Max 16-core. The idiot didn't specify in the graph.

I watched the video and the top most point for the M3 is indeed an M3 Max. I won't be too harsh on them as this was part of a live stream review, but they really should clean up their presentation, even if you explain it in the video. Basically, the data points with lines in them means they could directly control the max power draw (M3 being the exception). The individual dots means that they couldn't control the power draw and are relying on the power profiles to set the power draw, which means that it's a bit more of an estimate. The biggest difference is that the M3 data is really the M3 family data and each data point is a separate SoC with more cores as you go up in power consumption, which explains the seemingly impossible perf/w curve.

Edit: Here's the graph with added M3 CPU info (I'm guessing on the M3 Pro variants).

 

[Abwx]

Unless it's an M3 Max 16-core. The idiot didn't specify in the graph.

If all cores are functional at all powers it s still impossible, even if it were perfect transistors the scaling should be such that perf increase as a square root of power, so 2x the score mean at least 4x the power, 1.41x the perf mandate 2x the power.

FI at NBC the Asus Pro Art score 41% better in CB R23 than the Zenbook wich is around 30W, so the former use something like 65-70W.

Edit : On a side note :

Was Ryzen 9000 delayed over a simple typo? AMD Ryzen 9 9700X spotted - VideoCardz.com

Ryzen 9000 delay, it was a typo all along? Ryzen 9 9700X instead of Ryzen 7 9700X. AMD has not directly confirmed the reason for the Ryzen 9000 launch delay. Officially, the company cited quality issues that did not meet their expectations. However, it seems the issue might have been a simple...

videocardz.com

 

[Philste]

FI at NBC the Asus Pro Art score 41% better in CB R23 than the Zenbook wich is around 30W, so the former use something like 65-70W.

Everything is really weird. In NBCs Zenbook S16 Review, there is a Cinebench 2024 result (1099 points) of ProArt, which isn't in the ProArt test itself. 1099 is only 10% higher than Computerbase result of 33W Version.

 

[jdubs03]

Something funky going on with that int score looks like (the HXL dude pointed this out):


Looks more normal there:

 

[Rheingold]

Edit: Here's the graph with added M3 CPU info (I'm guessing on the M3 Pro variants).

I think you can remove the question marks. For the 12-core it fits, for the 11-core it's a bit low. Basically the strange shape of the curve is due to their low 11-core result. But there are no other models, ore are there?

 

[FlameTail]

Apparently that "M3" more or less means the whole family of M3 silicon.

Holy bizarre cow.

 

[igor_kavinski]

Ah, finally a use for the Ryzen AI NPU! https://videocardz.com/newz/amd-lau...zen-and-radeon-features-xdna-super-resolution

 

[SarahKerrigan]

I wonder, now that the Strix results are out, if folks will perhaps use a tiny bit more skeptical thinking when someone on a forum starts claiming Zen 6 will offer +85% iso-clock perf or whatever.

 

[techjunkie123]

I wonder, now that the Strix results are out, if folks will perhaps use a tiny bit more skeptical thinking when someone on a forum starts claiming Zen 6 will offer +85% iso-clock perf or whatever.

Orrrrr you just ride the HYPE train till launch and then cope with the retail performance...

Zen6 50% PPC incoming ??!!

 

[gdansk]

I wonder, now that the Strix results are out, if folks will perhaps use a tiny bit more skeptical thinking when someone on a forum starts claiming Zen 6 will offer +85% iso-clock perf or whatever.

We already have the Zen 6 estimate of 10% from the same slide that had Zen 5 right so there is little chance we can get a good hype train going even if the usual suspects try.

 

[igor_kavinski]

so there is little chance we can get a good hype train going even if the usual suspects try.

We still got Strix Halo fantasies to cling to, bro

 

[Abwx]

Everything is really weird. In NBCs Zenbook S16 Review, there is a Cinebench 2024 result (1099 points) of ProArt, which isn't in the ProArt test itself. 1099 is only 10% higher than Computerbase result of 33W Version.

Guess that they are messing with the power profiles, if around 30W the score is about 1000 then at roughly twice the power it should be at 1400, or possibly that NBC made their CB 2024 run while the laptop was still very hot, wich could explain the 10% only better perf than Cmputerbase.

 

[SarahKerrigan]

Guess that they are messing with the power profiles, if around 30W the score is about 1000 then at roughly twice the power is should be at 1400

Based on what?

 

[Abwx]

Based on what?

Based on mosfets transistors physics and observations of the actual power vs frequency curves of TSMC s process.

 

[SarahKerrigan]

Based on mosfets transistors physics and observations of the actual power vs frequency curves of TSMC s process.

LOL.

Man, I wish it was that easy.

 

[Abwx]

LOL.

Man, I wish it was that easy.

That s not that difficult when one is aware of the subjacent laws of physics, but sure that a few of those who dont understand those matters can be skeptical and rely on an "argument" like "lol" even if they are completely cluless about the thing.

 

[Rheingold]

Holy bizarre cow.

Yeah, it's really strange. He even talks how "the ones with the lines" were laptops where they tested all power levels on the same device, and then doesn't notice that this puts the white line in a strange spot. Later, he explains how with Apple you basically can't change the power consumption, but you buy another device if you want more performance. So yes, the white dots absolutely are the fixed values for different M3 chips.

That being said, @Hitman928 could you perhaps not only remove the question marks but also the white line between the white dots? 🤣

 

[Timorous]

We already have the Zen 6 estimate of 10% from the same slide that had Zen 5 right so there is little chance we can get a good hype train going even if the usual suspects try.

It did say 10%+ so. 100% IPC at 5.5Ghz here we goooooooo!!!

 

[Hitman928]

Something funky going on with that int score looks like (the HXL dude pointed this out):
View attachment 104056

Looks more normal there:
View attachment 104057

From the tests so far, it seems like the Zenbook model doesn't allow the HX 370 to sustain its single core boost. Add to that, the 7940HS has a 5.2 GHz boost versus 5.1 GHz of the HX 370, and you are looking at ~7% IPC increase, assuming the 7940HS holds its single core boost and the HX 370 at Anandtech is ~5 lower than boost clock (according to other ST tests from other outlets). I think this actually matches the previous ES Strix measurements done for Spec Int IPC.

 

[SarahKerrigan]

That s not that difficult when one is aware of the subjacent laws of physics, but sure that a few of those who dont understand those matters can be skeptical and rely on an "argument" like "lol" even if they are completely cluless about the thing.

The "spherical cows in a vacuum" energy here is amazing

 

[Hitman928]

Yeah, it's really strange. He even talks how "the ones with the lines" were laptops where they tested all power levels on the same device, and then doesn't notice that this puts the white line in a strange spot. Later, he explains how with Apple you basically can't change the power consumption, but you buy another device if you want more performance. So yes, the white dots absolutely are the fixed values for different M3 chips.

That being said, @Hitman928 could you perhaps not only remove the question marks but also the white line between the white dots? 🤣

Since you asked nicely, lol. It's quick and ugly but here you go. I also added a fitted line for the HX 370 and X Elite, though it's just eyeballed and the lower power levels for the X Elite numbers are weird. I'm guessing it's different SoCs at the bottom, similar to the M3 issue.



Edit: redid the fit lines.

 

[Abwx]

The "spherical cows in a vacuum" energy here is amazing

I would suggest a start here, that will spare some empty debates, to me and all the community that is not interested in tit for tat as an occupation to fill some...void..?.

MOSFET - Wikipedia

en.wikipedia.org

 

[Philste]

Guess that they are messing with the power profiles, if around 30W the score is about 1000 then at roughly twice the power it should be at 1400,

There's a problem with that Math and that Problem is ZEN5c. Clocks of 8 of the 12 Cores are Limited to 3.7GHz. Computerbase got an avg core clock of 3.43GHz at 33W. So starting at 40W, there's only 4 of 12 Cores that can scale further. So basically at that point every scaling is gone. Even if those 4 Cores got to 5.0 GHz that's only a 11-12% uplift over 40W. No matter how much power you throw in there. 1400 isn't possible.

It's even worse for 365 SKU, where ZEN5c is limited to 3.3GHz. I suspect limitations starts at about 25W here. So scaling is dead starting at 25W, however it's 4 of 10 Cores that continue scaling, so it could gain 20% at Max power compared to 25W.

 

[jdubs03]

From the tests so far, it seems like the Zenbook model doesn't allow the HX 370 to sustain its single core boost. Add to that, the 7940HS has a 5.2 GHz boost versus 5.1 GHz of the HX 370, and you are looking at ~7% IPC increase, assuming the 7940HS holds its single core boost and the HX 370 at Anandtech is ~5 lower than boost clock (according to other ST tests from other outlets). I think this actually matches the previous ES Strix measurements done for Spec Int IPC.

Shouldn’t that affect the fp score as well? Seems particular to just int.

~5, percent?