What controls Turbo Core in Xeons?

Page 90 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Gellért

Member
May 29, 2017
30
6
41
Cata40, since Noctua's backplate is mentioned, the guy is probably talking about socket 1151 not 2011-3 where the backplate is initially installed on the motherboard and provides sufficient hardness to prevent the motherboard from being significantly bent under the weight of CPU cooler heatsink.


I use a modified BIOS with removed µCU (microcode update) and removed ME (Intel Management Engine - the only way to flash such BIOS is to use a SPI Flash device like this one - works flawlessly). For the turbo unlock you just need the µCU to be removed.

You also need to tune the BIOS a little and install an EFI driver. I recommend using one of kjboughton's RC9 drivers (however, I compiled one for myself to get different offsets for my asymmetrical CPU duo - minus 60/50/50 mV for OEM and minus 70/50/50 mV for ES)







They will work together, but in AVX2 software you might observe different behaviour (lower turbo for OEM CPU) if you use 1F microcode which Kris (kjboughton) recommended. I get x27 for the SR1XK and x34 for the QGN7 for some reason. But if you use no microcode at all, you get maximum turbo for both of them in AVX2. With µCU versions 27 and 39 (and 3A for that matter) I did not observe any difference between them, they both got lower AVX2 turbo (x29) which is perfectly normal.


Sorry for asking but there is way to much to read: how can you make the v3x4......efi to run automatically?
For the microcode part it is enough to delete the mcupdate file in sys32 or I should use a modified bios?
There are way too much .efis. How about the v3x4-0.10b-i306f2-rc9_90_50_50_pc1.9v.efi one with 2x QGN7 2696v3 and the asus d8 ws board?
 
Last edited:

KW8cSgF53G

Junior Member
Oct 4, 2017
9
3
51
They will work together, but in AVX2 software you might observe different behaviour (lower turbo for OEM CPU) if you use 1F microcode which Kris (kjboughton) recommended. I get x27 for the SR1XK and x34 for the QGN7 for some reason. But if you use no microcode at all, you get maximum turbo for both of them in AVX2. With µCU versions 27 and 39 (and 3A for that matter) I did not observe any difference between them, they both got lower AVX2 turbo (x29) which is perfectly normal.
I may have found the answer...
It may dut to the "TDP% downclocking limitation(will explain below...)",
"the other (extra)" "fuse code" that built in a Retail version of e5-v3s chip...
It's not exist in a QS2(Eng. Sample) version of chip!

As I mentioned before, I got two different stepping cpus,
SR1XH(Retail version of E5-2683 v3), and QGN5(QS2-Eng. Sample of E5-2683 v3).
Then through keeping tracking the AIDA64+RTSS reading when an "AVX(2) Acceleration Enabled" video application is loading,

I found that when the wattage of SR1XH reaches 90.00W, equal to 90.00W/120.0W="75%TDP",
some cores(in High Performance Power Plan, almost half cores) frequency forcely goes down to default rated speed of 2.5GHz!
Compare to QGN5, same setup of that program as before, however result in lower CPU Utilization and multiplier ratio than QGN5(100.0MHz*27)!
Under the same load, QGN5 could reach higher CPU usage(average about +7.5%) and TDP% with complete the same voltage as SR1XH,
which gives no other restriction on all cores, all the way up to max turbo frequency until hitting around rated cpu TDP:120.0W.

And if I locked the VCCIN to 1.825v and with Power Cut enabled on SR1XH, then "TDP% downclocking limitation" behaved just like QGN5, GONE...
(although that way would lost the ability to measuring real power consumption of the cpu...)

Final, I heard some words from somewhere a few months ago (An video tutorial about unlocking e5-v3s all cores multiplier):
Intel has fixed "the BUG" in later version of the e5-v3s chip...Well, I guess that's all there is about...
(even if you did found a way to unlock the all cores multiplier to rated single cores max turbo frequency,
but still, you can't fully utilize it on medium to heavy AVX2 workload.)


(PS: If you locking the VCCIN and use Power Cut, with sufficient quantity of power phases like 2*8-pin EPS12V for ONE cpu,
the QS2 one will result in the same way as the Retail one, unlock TDP limitation but also got very hot in AVX(2) workload.
However, your MB is Dual-Socketed and with only pool quantity of 5 power phases, one 8-pin EPS12V for each cpu,
then even if you disable the F-IVR Telemetry (Power Cut Enabled) and locked the VCCIN, it still would't
give you much impact on extra performance gain...like my X10DAL-i.)

(sorry for my bad english.)
 
Last edited:
  • Like
Reactions: sciff

Gellért

Member
May 29, 2017
30
6
41
Can someone help me out pls?
I'm totally stuck with this turbo thing.
Got an Asus Z10PE D8 WS and 2x 2696v3 QGN7.
Before the ASUS board I got the x99 Taichi. I could delete the microcode (but not in UBU) and flash the bios easily. Then I was able to run the uefi shell to have the .efi driver easily as well.
Right now I can't delete the microcode in UBU (following the guide on page 9 here), can't flash the bios, and I wasn't been able to run the UEFI shell from an USB stick.
Could someone guide me or guide me to the proper informations I need? I read through almost all posts.
 

sciff

Member
Mar 6, 2017
136
52
71
Can someone help me out pls?
I'm totally stuck with this turbo thing.
Got an Asus Z10PE D8 WS and 2x 2696v3 QGN7.
Before the ASUS board I got the x99 Taichi. I could delete the microcode (but not in UBU) and flash the bios easily. Then I was able to run the uefi shell to have the .efi driver easily as well.
Right now I can't delete the microcode in UBU (following the guide on page 9 here), can't flash the bios, and I wasn't been able to run the UEFI shell from an USB stick.
Could someone guide me or guide me to the proper informations I need? I read through almost all posts.
I can walk you through the process from A to Z. Just start a private conversation
 
  • Like
Reactions: Gellért

proworker

Junior Member
Jul 29, 2017
12
0
6
Last edited:

Tansel Turgut

Junior Member
Feb 17, 2018
2
0
1
Hi, I have 2 x E5 2699 v3 ES processors (I think it is very similar to E5 2696v3)
QFQM, stepping 2, R2

I bought a Supermicro X10DAL-i-0 motherboard.
Base speed is only 2.3 with 36 cores.

can you guys help me get better speeds?

thanks!

tturgut@aol.com
 

Tansel Turgut

Junior Member
Feb 17, 2018
2
0
1
I may have found the answer...
It may dut to the "TDP% downclocking limitation(will explain below...)",
"the other (extra)" "fuse code" that built in a Retail version of e5-v3s chip...
It's not exist in a QS2(Eng. Sample) version of chip!

As I mentioned before, I got two different stepping cpus,
SR1XH(Retail version of E5-2683 v3), and QGN5(QS2-Eng. Sample of E5-2683 v3).
Then through keeping tracking the AIDA64+RTSS reading when an "AVX(2) Acceleration Enabled" video application is loading,

I found that when the wattage of SR1XH reaches 90.00W, equal to 90.00W/120.0W="75%TDP",
some cores(in High Performance Power Plan, almost half cores) frequency forcely goes down to default rated speed of 2.5GHz!
Compare to QGN5, same setup of that program as before, however result in lower CPU Utilization and multiplier ratio than QGN5(100.0MHz*27)!
Under the same load, QGN5 could reach higher CPU usage(average about +7.5%) and TDP% with complete the same voltage as SR1XH,
which gives no other restriction on all cores, all the way up to max turbo frequency until hitting around rated cpu TDP:120.0W.

And if I locked the VCCIN to 1.825v and with Power Cut enabled on SR1XH, then "TDP% downclocking limitation" behaved just like QGN5, GONE...
(although that way would lost the ability to measuring real power consumption of the cpu...)

Final, I heard some words from somewhere a few months ago (An video tutorial about unlocking e5-v3s all cores multiplier):
Intel has fixed "the BUG" in later version of the e5-v3s chip...Well, I guess that's all there is about...
(even if you did found a way to unlock the all cores multiplier to rated single cores max turbo frequency,
but still, you can't fully utilize it on medium to heavy AVX2 workload.)


(PS: If you locking the VCCIN and use Power Cut, with sufficient quantity of power phases like 2*8-pin EPS12V for ONE cpu,
the QS2 one will result in the same way as the Retail one, unlock TDP limitation but also got very hot in AVX(2) workload.
However, your MB is Dual-Socketed and with only pool quantity of 5 power phases, one 8-pin EPS12V for each cpu,
then even if you disable the F-IVR Telemetry (Power Cut Enabled) and locked the VCCIN, it still would't
give you much impact on extra performance gain...like my X10DAL-i.)

(sorry for my bad english.)
Hi, I have 2 x E5-2699 v3 ES processors (I think it is very similar to E5-2696v3)
QFQM, stepping 2, R2 (Base speed is only 2.3 with 36 cores).

I have a Supermicro X10DAL-i-0 motherboard.

can you help me get better speeds?

thanks!

tturgut@aol.com
 

kniaugaudiskis

Junior Member
Jun 4, 2009
18
3
81
Hey guys, I was finally able to crack the 3000 point mark in Cinebench R15 with my single E5 2696V3! It's definitely not a 24/7 OC due to the BLCK being 105MHz which is a little too high for full stability.

 
Last edited:

rottwag

Member
Apr 2, 2017
77
11
41
Hey guys, I was finally able to crack the 3000 point mark in Cinebench R15 with my single E5 2696V3! It's definitely not a 24/7 OC due to the BLCK being 105MHz which is a little too high for full stability.


Very cool! I think i could get such a result as well... I stopped at 29xx when my 2696 showed that it was not fully stable...

But aiming for the highest CB scire (one run is enough) maybe even more is possible? What was your offset? 100mV and multi of 36x ?
 

kniaugaudiskis

Junior Member
Jun 4, 2009
18
3
81
Very cool! I think i could get such a result as well... I stopped at 29xx when my 2696 showed that it was not fully stable...

But aiming for the highest CB scire (one run is enough) maybe even more is possible? What was your offset? 100mV and multi of 36x ?
Well not exactly, my core offset is -60mV and cache offset is set to -50mV. This yields a multi of 33x on all cores and combined with a 5MHz OC on the BLCK I have 3.46GHz all-core speed and that is fast enough to me to score 3001 points in Cinebench R15 :) As for my 24/7 settings, I'm running offset voltages mentioned above and 102.1MHz BLCK which gives me 3.37GHz all-core and 3.88GHz single-core speeds. If I leave the BLCK at 103MHz (3.4GHz all-core speed), I get a BSOD at idle once a week or so, even though my c-states are off.
 
Last edited:

rottwag

Member
Apr 2, 2017
77
11
41
Well not exactly, my core offset is -60mV and cache offset is set to -50mV. This yields a multi of 33x on all cores and combined with a 5MHz OC on the BLCK I have 3.46GHz all-core speed and that is fast enough to me to score 3001 points in Cinebench R15 :) As for my 24/7 settings, I'm running offset voltages mentioned above and 102.1MHz BLCK which gives me 3.37GHz all-core and 3.88GHz single-core speeds. If I leave the BLCK at 103MHz (3.4GHz all-core speed), I get a BSOD at idle once a week or so, even though my c-states are off.
Whaaat??

Actually i run 34x 101,9 for 24/7 which is 3,46 GHz and i get approx 2875 Points with offset -60 -50

At the beginning i just tried to find the limits and increased the offset until i was nit able to have 3 successful runs in CB...

At -100mv i got even a multi of 36x on all cores, but that did not run long....

I do not have any microcode in place and i tend to think that results are worse compared to my old setup where microcode 39 was loaded...

I simply compared the stock clockings and saw that with microcode i got almost 2475.. without microcode loaded i got only 2350...

Maybe i just start experimenting with different microcodes...

Which one do you use, if any??

Andy
 
Last edited:

kniaugaudiskis

Junior Member
Jun 4, 2009
18
3
81
Which one do you use, it any??
I'm running the 39 ucode. I believe you've scored only 2875 points at 3.46GHz due to not "optimizing" some Windows settings. Firstly, make sure you have your Windows power plan set to "High Performance" because it sets minimum processor state to 100% and in such a case your CPU doesn't have to waste any time jumping to its maximum speed when the benchmark has started and thus hurting the score. Secondly, go to task manager and set the Cinebench R15 process to "High" or "Realtime" priority. Your score will be higher now, let me know how it goes :)

By the way, I wish my CPU did 34x multi at -60mV vcore. What motherboard are you running?
 
Last edited:

rottwag

Member
Apr 2, 2017
77
11
41
I'm running the 39 ucode. I believe you've scored only 2875 points at 3.46GHz due to not "optimizing" some Windows settings. Firstly, make sure you have your Windows power plan set to "High Performance" because it sets minimum processor state to 100% and in such a case your CPU doesn't have to waste any time jumping to its maximum speed when the benchmark has started and thus hurting the score. Secondly, go to task manager and set the Cinebench R15 process to "High" or "Realtime" priority. Your score will be higher now, let me know how it goes :)

By the way, I wish my CPU did 34x multi at -60mV vcore. What motherboard are you running?
I tried different microcodes now (1F, 3A, 27 and 39). With all microcodes i only get a multi of 33x under full load.... BUT: With the 1F microcode, I managed to get 2904 in Cinebench - with the other 3 the results where quiet close to 2875 Points what I also get WITHOUT microcode but with multi of 34x ! ("High performance" mode and "Realtime" priority did not have any effect on the result on my side by the way)

Will continue making some tests - but if performce is the same (or even better) with 1F and with less power consumption i probably will use it. Does not matter if the multi is 33x or 34x when the performance is the same!

My mobo is an Asus X99-A
 
Last edited:

rottwag

Member
Apr 2, 2017
77
11
41
More thoughts:

#1 with microcode 1f, i get 33x101,9 = 2904.. if i would raise Blck to 105 i should also get approx 3000

#2 with my older 8-core i had 33x104,9 running stable with microcode 39 stable for almost 1 year.
Maybe i should try to raise bclk now having a microcode loaded.
Possibly, bclk can go higher with microcode than without?

Without microcode even 102,5 were not fully stavle with my 2696

What do you think?
 

kniaugaudiskis

Junior Member
Jun 4, 2009
18
3
81
Possibly, bclk can go higher with microcode than without?
Well, I haven't tested it myself but I'd say it all comes down to how many USB and SATA devices you have connected to your motherboard, because as you OC your BLCK, you're basically overclocking your USB and SATA controllers as well and these guys don't really appreciate it to say the least. For instance, when I had my BLCK at 105MHz, some of my USB devices stopped working and I had to reboot. Oh and my motherboard is the ASUS X99 Deluxe II.

With the 1F microcode, I managed to get 2904 in Cinebench - with the other 3 the results where quiet close to 2875 Points what I also get WITHOUT microcode but with multi of 34x !
I see, I'll try to load my PC without a ucode and see if I can hit 34x multi at -60mV, too.

("High performance" mode and "Realtime" priority did not have any effekt on the result on my side by the way
Are you temperatures in check? Does your CPU stay at the maximum speed throughout the test? :)
 
Last edited:

rottwag

Member
Apr 2, 2017
77
11
41
Well, I haven't tested it myself but I'd say it all comes down to how many USB and SATA devices you have connected to your motherboard, because as you OC your BLCK, you're basically overclocking your USB and SATA controllers as well and these guys don't really appreciate it to say the least. For instance, when I had my BLCK at 105MHz, some of my USB devices stopped working and I had to reboot. Oh and my motherboard is the ASUS X99 Deluxe II.
yeah, you are right - but I worked with excactly the same setup but differet CPU with BCLK 104,9 for 1 year - so either it depends on the CPU how much BCLK is stable or on the microcode... because the rest of my system is still the same.... on other thing that would also be possible is, that by increasing the BLCK the max freq of the CPU raised - due to that i could get instable as the Voltage is too low... so actually the low voltage is limiting the overclocking then... by saying that - i should really try if 33x 104,9 works stable with my -60mV offset as the max freq on all cores is excactly the same as with 34x101,9! ....i keep you updated...

I see, I'll try to load my PC without a ucode and see if I can hit 34x multi at -60mV, too.
I am asking myself now, when it runs @33x - could I increase the offset as it needs less voltage as the freq is lower?
But on the other hand - when it switches to 34x, it does not help as I need -60mV to be stable with 34x....

Are you temperatures in check? Does your CPU stay at the maximum speed throughout the test? :)
My tempereatures are fine - I have a Noctua DH-15 Cooler and overall 4 bequiet SilentWings 3 Fans... my max Temp running prime is 75C... with all other apps I get less.
After 2 hours of gaming i got 62C on the highest core! .... so cooling is def. sufficient... :)
 

Irgen

Junior Member
Feb 22, 2018
6
1
41
I can walk you through the process from A to Z. Just start a private conversation
Hey sciff, i'm just wondering isn't you the same person sciff from 3dcenter forum?

I have very similar setup with z10pe-d8 and 2696, currently its cinebench score is 4400 (Vray 32sec)
So just asking how well your system perforiming in cinebench? Or maybe you did tests in Vray bench?
I would like to unlock turboboost, but the procedure seems kinda complicated to me..,
but if it may go up to, say, 5500 or so, could be the game is worth the candle.
 

Cyras

Junior Member
Feb 24, 2018
1
0
1
german-xflight.de
What I ask myself for a long time is whether you can adjust / replace the microcode of the Xeon E5 V3 so that he thinks he is a 6950X? Had someone tried that?
 

Cata40

Member
Mar 2, 2017
156
6
81
As far as I know, since I am a member, she has not thought of it, at least declaratively here on the site.
Yeah, why did you try?
 

rottwag

Member
Apr 2, 2017
77
11
41
yeah, you are right - but I worked with excactly the same setup but differet CPU with BCLK 104,9 for 1 year - so either it depends on the CPU how much BCLK is stable or on the microcode... because the rest of my system is still the same.... on other thing that would also be possible is, that by increasing the BLCK the max freq of the CPU raised - due to that i could get instable as the Voltage is too low... so actually the low voltage is limiting the overclocking then... by saying that - i should really try if 33x 104,9 works stable with my -60mV offset as the max freq on all cores is excactly the same as with 34x101,9! ....i keep you updated...
Tested with microcode 1F on load 33x 102,9 and Computer crashed.... so no difference to withou microcode.
I also compared power consumption with/withoug microcode but did not recognize any difference in power consumtion.

In the CPU-Z Benchmark, results are worse with microcode loaded (max Multi 33) compared to without microcode (max Multi 34).

I think I have found my final Setup now: no microcode - all core 34x101,9 @ -60mVcore Offset :)
 

MOF

Member
Jul 31, 2017
118
33
101
In the CPU-Z Benchmark, results are worse with microcode loaded (max Multi 33) compared to without microcode (max Multi 34).
Without microcode synthetic benchmark results can be inaccurate. You will get higher or lower scores without microcode but this is not mean you can get better or worst result in the real world application.

Using system without microcode is not recommended
 

ASK THE COMMUNITY