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

[DisEnchantment]

Speculate at will

 

[tsamolotoff]

IT would be above 170w if you had it DELID with diamond cpu-waterblock, underground salt water reservoir at -3C, with premium industrial heathing pumps, heatexchangers, two car radiators. According to some people.

I actually have a sort of car radiator (Bykski B-RD1080-TK-V2 in addition to two 360 rads) and i use LM instead of paste, but it wouldn't probably go above much this amount of watts even with delid because it's mostly voltage limited, not thermally limited. Of course, there is something that prevents proper heat transfer with x3d cores (just compare temps on my highest score screenshot), but even with that CPU could probably do more (normal z4 can do 5.5 ghz all core with 1.25v ish even without delid)

 

[igor_kavinski]

Perf/watt comparison 9950X ES vs. Det0x's optimized 7950X



Det0x's data source: https://forums.anandtech.com/thread...ranite-ridge-ryzen-9000.2607350/post-41255152

Past 100W, 7950X can't compete on perf/watt in this benchmark and basically game over after 210W PPT. But it does show that if you are allergic to watts, 7950X will offer you better performance for your electricity bill at lower power levels.

STILL, for most of us, it's

RIP 7950X.

 

[Det0x]

Think we will need a proper V/F curve for 16core Zen5 before we can draw any final conclusions in regards to the comparison above..
Dont think the V/F curve for the 9550X ES is optimized/finalized at the lower end of the spectrum (sub 160w)

After i get a retail 9950X in my hands, i can do a real apples to apples comparison

 

[igor_kavinski]

Think we will need a proper V/F curve for 16core Zen5 before we can draw any conclusions in regards to comparison above..

True but it will be very surprising if the retail sample does well at the lower power levels.

 

[tsamolotoff]

What is interesting how the leakers got their IPC increase numbers with epycs if this trend is also prevalent among those SKUs (and if TDP wasn't increased they should roughly get the same amount of watts per core like 9950x at 60W PPT or less)

 

[Abwx]

True but it will be very surprising if the retail sample does well at the lower power levels.

Your ES is fed with the right cores voltages but the VRMs voltage is kept at a fixed value whatever the cores effective voltage.

This way the CPU integrated regulators have very low efficency at lower cores voltage since their voltage drop is uselessly high, with a proper implementation the VRMs voltages will be reduced to lower values with the ES when the cores are fed with a very low voltage, as to get the integrated regulators voltage drop at an adapted value.

 

[Hemu2K]

48K @300W is insane, hopefully the temps are much easier to handle this gen.
Looking forward to crazy avx512 speeds!

 

[igor_kavinski]

48K @300W is insane, hopefully the temps are much easier to handle this gen.

Bit less efficient than a monster 128 thread 7702P but hey, it's almost delivering similar perf with 4x fewer threads!

Cinebench R23 Scores [Updated Results]

Cinebench R23 is the newest instalment of the most popular CPU-Rendering Benchmark Cinebench. We have the Score Results for all modern Processors.

www.cgdirector.com

 

[Timmah!]

Yep.

Not bad in the grand scheme of things. Same CCD die area (~70 mm2), >20% increase in density while maintaining clocks, +16% IPC uplift, and even a small boost to perf/W. Oh, and they managed to add AVX512 and allegedly running AVX512 doesn't reduce clocks. If you can't call that a success, Idk what does.

https://twitter.com/x/status/1813281978591944827

So no about 30GHz slower as was the case with Skylake-X? :-D

 

[majord]

just drilling down into the screenshots of haims run with identical scores at 60w ppt..

little tidbit

9950X: 35w core pwr /2686Mhz [23985cb]
7950X: 36.9w core pwr/3160Mhz [23780cb]

= 17.6% higher clocks

looking at the VId's it seems the ES is asking a tad more Volts per Mhz in these lower ranges.. have to plot it. but if that's the case, there may be some better perf/watt in retail if that can be brought down a bit. Not sure what the Vf curve would be expected to look like on this process on the same core ( obviously this isn't the same core, but it does have similar Vmax it seems)

 

[igor_kavinski]

So no about 30GHz slower as was the case with Skylake-X? :-D

Intel CPUs with AVX-512 be like, hey so much data incoming! Things getting wayyyy too hot in here so gotta clock down!

 

[igor_kavinski]

Hx 370

https://twitter.com/x/status/1813896531352953040

Seems about right. Closer to the 60W desktop score.

 

[igor_kavinski]

HX 370 per thread score: 970.916

9950X ES per thread score: 749.531

WOW. Gimme 48 threads of HX 370 please, AMD!

 

[majord]

View attachment 103381

Seems about right. Closer to the 60W desktop score.

but only 12 cores!

 

[Philste]

HX 370 per thread score: 970.916

9950X ES per thread score: 749.531

What? 9950X has 32 Threads and over 40k score, so way over 1000 per thread.

 

[leoneazzurro]

What? 9950X has 32 Threads and over 40k score, so way over 1000 per thread.

He's speaking about the scores at 55-60W. But we are comparing mobile monolithic vs desktop chiplet and my opinion is that there is still room for optimization on the 9950X ES.

 

[jdubs03]

HX 370 per thread score: 970.916

9950X ES per thread score: 749.531

WOW. Gimme 48 threads of HX 370 please, AMD!

Just for reference for R23:

AMD Ryzen 9 8945H: cb 17423 / 1088.9375 per thread
Prime95 Power Consumption - external Monitor * - 107.5w… 80 W / 70 W SKU.

AMD Ryzen 9 8945HS: cb 17097 / 1068.5625 per thread
Single-core for both was right around 1812.
Both Prime95 and R15 power consumption were 93.5w. Using the 65 W / 45 W SKU.

AMD Ryzen 9 7940HS: cb 16867 / 1054.1875 per thread
Single-core was 1788.
Cinebench R15 Multi Power Consumption - external Monitor * - 107.6w. A bit murky since there are so many PL varieties but the device closest to the 107.6w value was a 104.7w and did a cb 16285 / 1017.8125 per-thread, was 65 W / 60 W rated.

(all values are from notebookcheck.net)

Will be interesting to see the power draw for similar testing.

 

[Philste]

After doing some math I would say this is basically full throttle for strix, so more a score of a potential Desktop SKU where TDP is no limit.

How I got to this conclusion:

8700G has ~18k in multi, 1833 in single according to CPU monkey. 1833×8×1.25(for SMT) = 18330. Fits quite well, so I just assume 8700G is able to run allcore on specified Boost.

If we do this for Strix we get: 2010×4×1.25= 10.050 for ZEN5

1460(3.7GHz Boost for ZEN5c according to David Huang)×8×1.25 = 14.580

So I get 24.600 as max value strix can theoretically achieve, 8700G was 1.5-2% higher than reality so I would say 24.400 is the max points strix can achieve.

Cutdown strix: ~19600 points with same measurement.

 

[Josh128]

Thats the difference between monolithic design + Zen 5c cores vs MCM high power cores. Much, much more efficient at low power and much lower power at idle.

 

[IEC]

I wouldn't bother putting too much stock in a gray market ES.
1) It is an ES. Who knows what is different from retail.
2) "It's totally a 9950X! I swear it!" It wouldn't be the first time someone faked CB data...

 

[igor_kavinski]

I wouldn't bother putting too much stock in a gray market ES.

I don't think the ES user I know, obtained it illegally. If he did, I would've bought one from him too

 

[Nothingness]

I don't think the ES user I know, obtained it illegally. If he did, I would've bought one from him too

The problem is not legality but rather how much it is representative of the final consumer product.

 

[igor_kavinski]

The problem is not legality but rather how much it is representative of the final consumer product.

Yeah, I totally understand that. Curious to see how much the final retail sample will differ from the results posted so far.

 

[PJVol]

Your ES is fed with the right cores voltages but the VRMs voltage is kept at a fixed value whatever the cores effective voltage.

I'm curious what do you draw this conclusion from?

 

[igor_kavinski]

@Josh128

Enjoy your prize from our excellent ES user!