Dual E5-2687W clocking with turbo modes?

[niko084]

I'm looking at building a system with 2 of these chips (E5-2687W), and I would really love it if someone could verify if it's possible to lock them all cores active at their max turbo value being 8 cores active locked at 3.8ghz without messing with bclk.

This system is being built for doing some heavy floating point operations, stability will be key so I want to stay away from blck adjustments, however if I could get them locked in at 3.8ghz all cores active 24/7 I would be a very happy person.

****
Sorry wrong cpu in title.
****

 

[Shou Miko]

what benefit will u get from having them not using EIST (Speedstep) and so on?

it will use more power, and it will be driving at max speed all the time which will just be pointless.

 

[niko084]

what benefit will u get from having them not using EIST (Speedstep) and so on?

it will use more power, and it will be driving at max speed all the time which will just be pointless.

No that's more so the entire point. The operations this system will be handling will be running for days/weeks 24/7 on all cores.

A little more in depth these are calculations for dispersion of gases given various atmospheric conditions.

 

[OneMoar]

depends on the board some-boards have controls that let you define what loads trigger what multipliers
Dual socket LGA 2011 boards have pretty much gone the way of the DODO
really tho you are barking up the wrong tree with Xeons
they just aren't worth the money and effort
not when you can pickup a i7-3970X and clock it to the moon

 

[niko084]

depends on the board some-boards have controls that let you define what loads trigger what multipliers
Dual socket LGA 2011 boards have pretty much gone the way of the DODO
really tho you are barking up the wrong tree with Xeons
they just aren't worth the money and effort
not when you can pickup a i7 3770k or a 4770k and clock it to the moon 4.5Ghz+

With this work cores is everything, already using a 4.2ghz 980x and a 4.8ghz 3960x, along with an older i7-930 stock clocks.

This is not a desktop application, the software is best suited for a cluster for distributed computing, but cost overhead is too high, it will utilize as many cores as I tell it to and it scales incredibly. Playing with testing, the 980x clocked with HT off barely kept up with the i7-930, with HT on, it's >75% faster.

The future of this will be dual cpu systems only as the jobs are too large and complex, currently a few of these projects are looking like 3-4 MONTHS of 24/7 use on the 3960x. Goal is to get that number down without building a cluster ($$$$$$).

BTW: It looks after a few hours now I've found my answer, sounds like that chip has a locked maximum multiplier for turbo mode with all cores active, so looks like 3.5ghz is the cap for 8 cores, which while not as great as I would like, it's better than 3.1ghz.

 

[Micro]

BTW: It looks after a few hours now I've found my answer, sounds like that chip has a locked maximum multiplier for turbo mode with all cores active, so looks like 3.5ghz is the cap for 8 cores, which while not as great as I would like, it's better than 3.1ghz.

Are these retail 2687w's or ES ?
I believe 3.4ghz is max for 6-8 cores on retail and 3.2ghz for 6-8 cores on ES's.
I do not know if they can be locked at any given TurboBoost multiplier, but most boards allow locking out TurboBoost, leaving you at 3.1ghz.
But TurboBoost IS triggered by Power State so you will end up with a low multiplier if your software scales as well as you think.
Of course the same thing would happen with any CPU using TurboBoost unlocked.

But with 16 cores and 16 HT running full out, heat is going to be a problem with the VRMs.
Cooling the CPUs is standard stuff.
You're looking at each 2687w or ES using 110-125w (real world) so the VRMs are going to be throwing a lot of heat, you may want to step up the sinks on them or add an auxiliary cooling fan to them.

What board are you looking at?
And if the software really likes cores, did you consider the 4600's?
ES 4600s can be had at a decent price.
Boards for quads are high, but cheaper than 2 dual setups.

BTW -

Dual socket LGA 2011 boards have pretty much gone the way of the DODO
really tho you are barking up the wrong tree with Xeons
they just aren't worth the money and effort
not when you can pickup a i7-3970X and clock it to the moon

You're looking at this wrong, he needs maximum threads, not max speed.
He's not gaming.
As to dual 2011 boards being Dodos, well, Intel makes at least 7, Asus 4, Gigabyte 4, and Supermicro at least 15 that I know of.
Intel and SM also make quad 2011 boards.

 

[niko084]

I think for the time being a quad system is going to cost a little much, however might be the route if they don't get on top of CUDA support so I can build a nice hybrid.

I'm looking at using the Asus Z9PE-D8 WS, partially due to the fact the system will be running Win7.

Cooling I don't think will be an issue, both cpu's would get an AIO liquid cooling system, with decent case fans and placed inside an enclosed air filtered closet that has a nice A/C unit blowing into it, currently ambient temps stay right around 60F and it seems to keep up pretty easily on low still.

 

[McSteel]

Any chance a Xeon Phi could be useful here? Until CUDA support kicks in for some Teslas...

 

[OneMoar]

or even a AMD Optron granted its not a fast as a Intel chip but you could have 24 threads with a dual socket setup

 

[McSteel]

As far as I understood, the OP already has some Xeons (E5-2687W), that he wants to get the most out of. A Xeon Phi is an easy drop-in for the already existing system. It's a significant expense, but could also mean a significant performance boost, potentially higher than building another system and distributing workload to it.

 

[niko084]

or even a AMD Optron granted its not a fast as a Intel chip but you could have 24 threads with a dual socket setup

The AMD chips perform simply terribly, not sure why, haven't tried personally but the software company and other users confirm.

Xeon Phi, I forgot about that, I wonder how it might get along, doing to do some more research.

***
Do NOT already have chips, this is a new build that is being pondered.

 

[McSteel]

Ah, my bad, it seems my English comprehension is still asleep, even though I'm up now

Anyway, the Phi should offer better "computational density", if such a term even exists, and the software should scale nicely since it uses x86-compatible cores. The price should be $1000-1500, so it should also be a good bang for the buck, considering it offers ~1 TFLOPS of FP64 performance...

 

[OneMoar]

what is the name of this software ??

 

[niko084]

I took a look and I'm just a little uncertain of how the Phi works, I mean is it only compatible with certain software? Certain OS?
I know nothing about these things beyond the release articles I read, which granted made them look pretty promising as a new technology.

Going to check on a few things...

***
I'm not exactly certain of the software, I believe it's Ansys Fluent but I'm waiting for an email back to confirm.
If so the new version supports CUDA as well.

There is a secondary software that is used as well.

 

[Micro]

I think for the time being a quad system is going to cost a little much, however might be the route if they don't get on top of CUDA support so I can build a nice hybrid.

I'm looking at using the Asus Z9PE-D8 WS, partially due to the fact the system will be running Win7.

Cooling I don't think will be an issue, both cpu's would get an AIO liquid cooling system, with decent case fans and placed inside an enclosed air filtered closet that has a nice A/C unit blowing into it, currently ambient temps stay right around 60F and it seems to keep up pretty easily on low still.

The CPUs are easy to cool, just watch the VRM temps.
The server arm of Asus has/had during the design phase of that board, optional cooling fans for the VRMs.
The 2pin plugs for the fans on that board made it into production (right next to the VRM heatsinks), but I don't know if the fans were ever production.
But normally, there is not a pair of fan headers on a production board labeled "VR_FAN" if someone didn't see a need for them

 

[Micro]

Xeon Phi, I forgot about that, I wonder how it might get along, doing to do some more research.

http://software.intel.com/en-us/articles/is-the-intel-xeon-phi-coprocessor-right-for-me

 

[OneMoar]

The CPUs are easy to cool, just watch the VRM temps.
The server arm of Asus has/had during the design phase of that board, optional cooling fans for the VRMs.
The 2pin plugs for the fans on that board made it into production (right next to the VRM heatsinks), but I don't know if the fans were ever production.
But normally, there is not a pair of fan headers on a production board labeled "VR_FAN" if someone didn't see a need for them

what are you on about with the VRM' I have never seen the VRM's on a enterprise level board ever be a problem they get hot but they are MADE to get hot Temps in excess of 100c are considered Operating temp

 

[niko084]

The CPUs are easy to cool, just watch the VRM temps.
The server arm of Asus has/had during the design phase of that board, optional cooling fans for the VRMs.
The 2pin plugs for the fans on that board made it into production (right next to the VRM heatsinks), but I don't know if the fans were ever production.
But normally, there is not a pair of fan headers on a production board labeled "VR_FAN" if someone didn't see a need for them

I wouldn't imagine given good case cooling they would have any issues, however it's something I will check if this is built. I'm certain something can be easily be designed or upgraded to keep them cool, thanks for the heads up.


***
Software is Ansys Fluent, which in the newest version does support CUDA, however for only particular functions... However, custom modules are being run... Going to do some checking into how far they have got it and maybe do some testing.

Thanks for the link on the Phi, going to read over now.

 

[niko084]

what are you on about with the VRM' I have never seen the VRM's on a enterprise level board ever be a problem they get hot but they are MADE to get hot Temps in excess of 100c are considered Operating temp

I think the concern here is that when the cpus will be running at 100% for weeks -> months at a time and must remain perfectly stable and aiding in ensuring lifespan.

already burned out 2x 980x

 

[Micro]

what are you on about with the VRM' I have never seen the VRM's on a enterprise level board ever be a problem they get hot but they are MADE to get hot Temps in excess of 100c are considered Operating temp

Because I work with this stuff for a living every day

His A/C cooled closet will help much, but without adequate airflow, you end up with localized hotspots that over time degrade and eventually kill components.

The VRMs are the weak link on the board.
He said he intends to run the board 24/7/365 at full load.
Asus does not design a workstation (which the Z9PE-D8 WS is) or server class boards to handle that much load without additional cooling as provided by a rackmount case (with the accompanying screaming fans) and even then they are not designed for continuous full load without additional cooling.

On top of that he intends to use AIO water coolers for the CPUs.

Dual (and quad) LGA 2011 boards are designed to have the CPU air coolers blow air over the VRM heatsinks to help cool them. That's the reason the VRM sinks are aligned to get airflow from the CPU HS/fan rather than from case fans..

At a continuous full load 2687w's will use 115-125w and long term VRM temps can climb above 120c, that's why Asus originally looked into provisions for VRM fans.
VRM operating temps are designed to be in the 75c-80c range, 100c is definitely hotter than they should be.
They will handle 100c, but it shortens their life.
You don't need my word for it, ask any OEM