Intel Xeon W9-3495X Sets World Record, Dethrones AMD Threadripper

[AleksandarK]

When Intel announced the appearance of the 4th generation Xeon-W processors, the company announced that the clock multiplier was left unlocked, available for overclockers to try and push these chips even harder. However, it was only a matter of time before we saw the top-end Xeon-W SKU take a chance at beating the world record in Cinebench R23. The Intel Xeon W9-3495X SKU is officially the world record score holder with 132,484 points in Cinebench R23. The overclocker OGS from Greece managed to push all 56 cores and 112 threads of the CPU to 5.4 GHz clock frequency using liquid nitrogen (LN2) cooling setup. Using ASUS Pro WS W790E-SAGE SE motherboard and G-SKILL Zeta R5 RAM kit, the OC record was set on March 8th.

The previous record holder of this position was AMD with its Threadripper Pro 5995WX with 64 cores and 128 threads clocked at 5.4 GHz. Not only did Xeon W9-3495X set the Cinebench R23 record, but the SKU also placed the newest record for Cinebench R20, Y-Cruncher, 3DMark CPU test, and Geekbench 3 as well.



View at TechPowerUp Main Site | Source

 

[P4-630]

Just , nothing more, it's useless for everything else....

 

[Daven]

Ryzen Threadripper Pro 7995WX: “Hold my beer”

 

[AnarchoPrimitiv]

Ryzen Threadripper Pro 7995WX: “Hold my beer”

Yeah, I know....the Zen 3 architecture is what, 2-1/2 years old?

 

[kondamin]

Yeah, I know....the Zen 3 architecture is what, 2-1/2 years old?

And maybe this time amd won’t keep them back 2 years after the rest of the line up was released.

 

[Bruno Vieira]

Ryzen Threadripper Pro 7995WX: “Hold my beer”

If the allow OCing

 

[Darmok N Jalad]

So at default clocks, it’s a 350W CPU with a 420W max. I wonder what it was consuming at these settings? Anyone able to do that math? 600W+ maybe?

 

[Bruno Vieira]

So at default clocks, it’s a 350W CPU with a 420W max. I wonder what it was consuming at these settings? Anyone able to do that math? 600W+ maybe?

1Kw min, at stock the clocks are around 3.5Ghz at 100% usage

 

[TumbleGeorge]

So at default clocks, it’s a 350W CPU with a 420W max. I wonder what it was consuming at these settings? Anyone able to do that math? 600W+ maybe?

6000W I just interpolated

 

[Dave65]

Did Intel hide the chiller this time?

 

[TumbleGeorge]

Did Intel hide the chiller this time?

Something is very hot to the right of the debug display with the digits 67

 

[Chrispy_]

"Dethrones Threadripper"

Pffft - only because nobody has yet pushed beyond a 96-core Genoa EPYC. Threadripper is old and tired and 96C/192T EPYC Genoa doesn't exist in a platform that's open to overclocking. It would murder Sapphire Rapids in 2P configuration (already has) but the limitation is CB23's 256-core limit. A 2P Genoa server is just wasting 128 cores because Cinebench needs updating to handle faster, wider platforms....

So at default clocks, it’s a 350W CPU with a 420W max. I wonder what it was consuming at these settings? Anyone able to do that math? 600W+ maybe?

If the TDP of 420W is 56 cores at 3.3GHz (as reported), that's around 950mv on Intel7 for Golden Cove P-cores.
LN2 voltage wall for 12th/13th gen is around 1.5V
(1500/950) for the ΔP, (5.4/3.3) for the ΔF, and 420W*(ΔP^2)*ΔF=1713W

YMMV based on silicon lottery. ofc. Primo-grade W9-3495X silicon is probably more efficient than most, but the 13900KS we're comparing against for Golden Cove is already cherry-picked, top-tier silicon so I don't think my guess is going to be out by more than 20% or so. 1400W+ for sure. I wonder what PSU they were using for this run...

 

[TumbleGeorge]

the TDP of 420W is 56 cores at 3.3GHz (as reported), that's around 950mv on Intel7 for Golden Cove P-cores.
LN2 voltage wall for 12th/13th gen is around 1.5V
(1500/950) for the ΔP, (5.4/3.3) for the ΔF, and 420W*(ΔP^2)*ΔF=1713W

The consumption function when overclocking is not a straight line but an exponential. If I'm not mistaken.

 

[Chrispy_]

Correct. That's why there's a 'squared' in there. P=I^2*R, and P-IV, so the voltage delta squared is the cause of exponential power consumption, all other factors being equal.
It's only a rough calculation but for the last 20 years of CPUs and GPUs it's nearly always been in the right ballpark.

 

[TumbleGeorge]

And why not raised to the third³ degree? Why does the formula use the second degree? I don't know much about math and I'm curious why.
Edit:
Oh i wrote before you finished your comment. Now it's clearer. From observing the matches.

 

[Chrispy_]

Basic physics formulae:

P=IV and I=V/R. so substituting I from Ohms Law into Watt's Law gives you two V's on the top side of the equation P=V*V/R.
That's where your squared comes from for power.

 

[eldon_magi]

In Greece? So, where did he get the W790 SAGE from? There are no known price quotes or sellers for it in europe, or is there?

 

[Ferrum Master]

Basic physics formulae:

P=IV and I=V/R. so substituting I from Ohms Law into Watt's Law gives you two V's on the top side of the equation P=V*V/R.
That's where your squared comes from for power.

Most people around here including reviewers get confused seeing word "power" measured in Watts... they have their own term, weird term Wattage.

I am not sure you can finish school without knowing Ohms law even when taking Music oriented Class.

 

[Crackong]

And it sucks around 1500W + from the wall ?
wow just wow

 

[TumbleGeorge]

Don't the parasitic currents in the circuits increase even more as the voltage increase is applied? I think that the mere application of the law of OM is not sufficient to validate the whole number of consumption in this case. I'm also curious, since the CPU isn't the only consumer in the system under test, and the VRMs and power supply aren't 100% efficient, how much electricity is really being pulled from the wall?

 

[Ferrum Master]

Don't the parasitic currents in the circuits increase even more as the voltage increase is applied?

Not voltage is a fault here, but dielectric properties. Term parasitic should be used when we simulate a real component with inductor capacitor properties, considering the size of the CPU, there is no space for that, it is mostly leakage here. If you cram so many things besides and the operate at high frequencies, the outcome is obvious, at least for Intel tech nodes. It just acts as it is designed at the stock frequency bracket. Anything higher isn't calculated and even should not be. Years back it wasn't a issue, it started around Ivy Bridge era where you hit a frequency wall then the CPU becomes a leaking pig with Broadwell being definition of it.

We are hitting limit of the silicon process, some parts even refuse to shrink more. The outcome is obvious. This sport... LN ePeen contest is not a indicator for something. The tuning of the silicon efficiency is not directly related to the maximum performance under LN. It is a design plan.

 

[Jism]

If the TDP of 420W is 56 cores at 3.3GHz (as reported), that's around 950mv on Intel7 for Golden Cove P-cores.
LN2 voltage wall for 12th/13th gen is around 1.5V
(1500/950) for the ΔP, (5.4/3.3) for the ΔF, and 420W*(ΔP^2)*ΔF=1713W

it's not so linear.... Once the temperature goes down your actual required voltage becomes less. The chip consumes a tad less current as well at the proces because there's less leakage.

I would guess around 900 to 1200W....

 

[Wye]

PR stunt.
The customer base of Xeons care more about stability than overclocking.

 

[medi01]

The problem for such PR is that the handful of big fish that command the cloud market, are not susceptible to it.

 

[BertMacklenFBI]

Just , nothing more, it's useless for everything else....

Actually it's an amazing HEDT CPU for Intel-considering that it's using Golden Cove P-Cores. And 112 PCIE lanes.......

This would actually be a perfuect MacPro (don't get me started lol) CPU, but I'm excited to see the Trashcan incident repeat intself...lol