What controls Turbo Core in Xeons?

Page 52 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

plonialmoni

Junior Member
Aug 4, 2017
5
0
36
Also, one more thing... Could VRMs get too hot with this? Will a ASRock gaming professional i7 be good enough in that department?

Sent from my VS986 using Tapatalk
 

lucien_br

Member
Feb 13, 2017
37
7
41
Hi, I guess I'm a latecomer to this thread, I have an ASRock x99 motherboard. Is all I need to do is flash one of the modded BIOSes for it to get an E5 2690 v3 at max turbo on all cores, or are there more steps I'm missing? I have adequate cooling (240mm water) by the way.
 

lucien_br

Member
Feb 13, 2017
37
7
41
Hi, I guess I'm a latecomer to this thread, I have an ASRock x99 motherboard. Is all I need to do is flash one of the modded BIOSes for it to get an E5 2690 v3 at max turbo on all cores, or are there more steps I'm missing? I have adequate cooling (240mm water) by the way.




No links to downloadable BIOS allowed. Link needs to from the manufacturer.


esquared
Anandtech Forum Director
 
Last edited by a moderator:
  • Like
Reactions: plonialmoni

plonialmoni

Junior Member
Aug 4, 2017
5
0
36
I should be getting it all together in a couple of weeks. Sorry for posting earlier with my phone's tapatalk signature, I didn't know the default signature was a rules violation here. Hopefully I've disabled it now. If it shows up, a moderator should let me know - it would be not intentional at this point.
 
Last edited by a moderator:

timk1980

Member
May 11, 2017
25
1
41
Okay, partially solved my problems with Z10PE-D16WS + 2x E5-2696v3...

Had to kill the windows microcode file; running with no microcode loaded at all, individual cores can go up to 38x, though they tend to hover in the low 30-33x range for large threaded workloads. Still seeing the massive throttling happening when using all 18 cores, though: CPU-Z score and Intel Burn Test performance both reflect that clearly. (Oddly enough, Cinebench works better with 18 than 16 cores).

Best I can do with undervolting is using one of Payne's -20 EFI drivers. Any of the -50 causes crashes (if it loads at all).

I switched in the BIOS to use only 16 cores on each CPU, and CPU-z earns me a little over 13k for multi, and 440 single, which both seem pretty good.

Using throttlestop and HWinfo, I can see that I'm getting what amounts to tdp throttling (never see the power go above about 165-170W max).

So, a few questions for anyone that may know:

Any BIOS settings that can allow a higher power allowance? (I'm getting upgraded cooling to deal with this, as well)

For those that have gotten consistent 33-34x on all 18 cores under heavy load, and/or have been able to keep even AVX workloads from downclocking, what settings did you tweak, etc?

Anyone else see the oddness in performance (from throttling presumably?) with the full 18 cores?

Thanks to everyone!
 

timk1980

Member
May 11, 2017
25
1
41
If you have load line calibration in your bios it should help stabilize your cpu
Indeed it does! Doesn't help running 18 cores, but when limiting to 16 cores, CPU-Z multi went from approximately 13k to just shy of 15k.

Also, increasing "Power Thermal Control" from 130 to 140 (max seems to be 151, haven't tried going that far yet) brings the time for 10 linpack runs in IBT from 130s -> 121s, a noticeable improvement.

Those are the only two settings I've confirmed to make a difference. Still kind of annoyed that I can't get all 18 cores to run reliably well with the hack, but I'm happy with the 16c performance no doubt!
 

Dufus

Senior member
Sep 20, 2010
675
119
101
@MOF so many different variables for throttling one just has to do what one can. The biggest problem I personally see is with AVX2, this seems to operate via a VID table, that is it seems to try and keep core VID below a certain value and what ratio one gets is tied to that.

Here's an example running just one thread with affinity set, just a simply loop, no AVX2. Turbo in yellow.

CPU runs at it's max 30x ratio


Now see what happens when we run just one AVX2 opcode once every 50 milliseconds on one thread of each core.

Power control gets ready for AVX2 by slamming down the ratio to 25x and from what I am led to believe leaves it there for 1 millisecond after the last AVX2 instruction before returning to the top ratio, 30x in this case. The irony in this case is that it doesn't even run the AVX2 as 256bit but needs a warm up period. Some info on Agner Fog's blog, for instance http://www.agner.org/optimize/blog/read.php?i=387. Limit reasons shows EDP logged, with the event being logged in yellow.

Now if we run the same opcode once every millisecond

Ratio is now dropping to 25x and we can see EDP continually active (red) and only using 37W of power, no current or power throttling..
 
Last edited:
  • Like
Reactions: kjboughton

MooNiSLe

Junior Member
Apr 23, 2017
4
1
81
@MOF so many different variables for throttling one just has to do what one can. The biggest problem I personally see is with AVX2, this seems to operate via a VID table, that is it seems to try and keep core VID below a certain value and what ratio one gets is tied to that.

Here's an example running just one thread with affinity set, just a simply loop, no AVX2. Turbo in yellow.

CPU runs at it's max 30x ratio


Now see what happens when we run just one AVX2 opcode once every 50 milliseconds on one thread of each core.

Power control gets ready for AVX2 by slamming down the ratio to 25x and from what I am led to believe leaves it there for 1 millisecond after the last AVX2 instruction before returning to the top ratio, 30x in this case. The irony in this case is that it doesn't even run the AVX2 as 256bit but needs a warm up period. Some info on Agner Fog's blog, for instance http://www.agner.org/optimize/blog/read.php?i=387. Limit reasons shows EDP logged, with the event being logged in yellow.

Now if we run the same opcode once every millisecond

Ratio is now dropping to 25x and we can see EDP continually active (red) and only using 37W of power, no current or power throttling..
Great Job! Dufus could u provide this AVX2 test program here? Seems very handy to prove some hypothesis about the power limt, thx!
 

kjboughton

Senior member
Dec 19, 2007
330
117
116
http://www.anandtech.com/show/11698/intel-finalizes-skylakex-processor-specifications-18cores-44-ghz-165w-on-september-25th

Looking at the bottom of page 1; versus E5-2696 v3 (18C/36T), you get:
  • Higher single-core turbo of 4.2GHz vs. Xeon with hacked single-core turbo of ~4GHz (38 x 105 BCLK)
  • Higher multi-core turbo up to 12C but LOWER multi-core turbo from 13C and up (3.5 down to 3.4GHz vs. Xeon at sustained 3.57GHz (34 x 105 BCLK)
  • Higher TDP vs similiar core count Xeons available since September 2014
  • Single socket operation only versus 2P Xeon
  • AVX-512 and other IPC gains since Haswell
For the price of a single i9-7980XE you can find a pair of E5-2696v3's and a used/refurbished MB and still put money in the bank.

Disclaimer: Long HCC v3/v4 Xeons. This market is frozen.
 
Last edited:

timk1980

Member
May 11, 2017
25
1
41
Does anyone have a copy of the v3x2_vcc.efi by @randir that does not undervolt? The original link is long dead, but the -50mV version works great for me, alas not fully stable.
 

sciff

Member
Mar 6, 2017
136
52
71
Does anyone have a copy of the v3x2_vcc.efi by @randir that does not undervolt? The original link is long dead, but the -50mV version works great for me, alas not fully stable.
Here: https://yadi.sk/d/ZGu0BU3X3LXubC

Guys. Will a QS and a product version of E5-2696v3 work together in a dual-cpu configuration? Given that their specifications, frequencies, steppings etc. are identical?



 
Last edited:

evilr00t

Member
Nov 5, 2013
29
8
81
QGNx is C1 QS, which is also the retail stepping.

Generally speaking, mixing steppings in multi-cpu configurations is supported by Intel, as long as the stepping difference isn't greater than one. The UEFI is responsible for downgrading the features to the oldest stepping. The last time this actually mattered was with the Sandy Bridge 206d6/206d7 chips, where the 206d6 chips had VT-d disabled due to an erratum; Intel has more recently not done any CPU steppings other than the initial one, so mixed stepping support is no longer relevant.

You're supposed to use the newer CPU as the bootstrap processor, but that rule is enforced by the UEFI. I believe mixing QS C0 and QS C1 Haswell-EP cpus would work, but definitely not the B-step ES2s with the QS's.

TLDR: should work fine with retail. Have fun.
 
  • Like
Reactions: sciff

sciff

Member
Mar 6, 2017
136
52
71
You're supposed to use the newer CPU as the bootstrap processor, but that rule is enforced by the UEFI. I believe mixing QS C0 and QS C1 Haswell-EP cpus would work, but definitely not the B-step ES2s with the QS's.
So I should use the QGN7 as CPU0, that is to install it into the right hand side socket? Is that correct?
 

evilr00t

Member
Nov 5, 2013
29
8
81
So I should use the QGN7 as CPU0, that is to install it into the right hand side socket? Is that correct?
If I'm interpreting this correctly, the bootstrap processor is usually CPU0, so I'd put the newer (retail) chip there. It probably doesn't matter because both are QS-C1 chips (the retail version just has the ES flag unset).
 
  • Like
Reactions: sciff

CANONKONG

Member
Jul 11, 2017
98
62
46
 
Last edited:

Kalistoval

Junior Member
Aug 2, 2017
13
1
41
Awesome I'm excited, fixed lol
 

Cata40

Member
Mar 2, 2017
156
6
81
I do not know how you did, but this time, bios for asrock oc formula, it works charming, greeting
..............................................................................................
I'm coming back after 2 hours of continuous use with Handbrake and I mean it's slipping and going round this time, its perfect, like original bios
 
Last edited:

James Kim

Junior Member
Jun 3, 2017
17
3
41
Can you provide some instruction on how to use the tool? are you using mmtool or uefitool? Where should I insert a ffs file to? And, I should work on the BIOS file that has mcode removed, correct? Thanks!
 

CANONKONG

Member
Jul 11, 2017
98
62
46
Can you provide some instruction on how to use the tool? are you using mmtool or uefitool? Where should I insert a ffs file to? And, I should work on the BIOS file that has mcode removed, correct? Thanks!
mmtool and uefitool are both needing.Use hex edit and uefi tool to replace the original bios microcode with the mod microcode.And use mmtool to add the ffs to the DexCore and save,than use the UBU to recheck the bios.
 

CANONKONG

Member
Jul 11, 2017
98
62
46
I do not know how you did, but this time, bios for asrock oc formula, it works charming, greeting
..............................................................................................
I'm coming back after 2 hours of continuous use with Handbrake and I mean it's slipping and going round this time, its perfect, like original bios
I will wtite a teaching and upload,because I can't upload the pictures on Anandtech.
 

ASK THE COMMUNITY