What controls Turbo Core in Xeons?

[randir]

I don't experience frequency drop, even though my E5-2696 V3 R2 has 18 cores

Can you run mprime (prime95) torture test with small ffts and then check frequency? I find it the most CPU demanding benchmark, far more demanding than linpack/renderers.

Is this motherboard any good to unlocked Xeons?

Supermicros are unlockable, though they require a manual BIOS editing. The only problem with them is that the BIOS lacks any overclocking features - BCLK, voltages, memory clocks, etc. So, you'd have to use external software for that - XTU/throttlestop do the job for windows (no BLCK still), but you're out of luck on linux unless you want to dig into msr registers progamming. They also have a very strange fan management, but that's solvable.

 

[sciff]

randir, yes it is. No other test or application can cause 340 W TDP. And believe me, HWInfo doesn't lie, I've checked it with socket power monitoring device.

Spoiler: Torturing mu CPU with Prime95's Small FFTs

Torture_Test_Anim.gif

 

[randir]

randir, yes it is. No other test or application can cause 340 W TDP. And believe me, HWInfo doesn't lie, I've checked it with socket power monitoring device.

And if you place the following content into the local.txt file in the prime95 folder, would you be able to hit more that x34 multiplier?

CpuSupportsAVX2=0
CpuSupportsAVX=0

PS: impressive chip, I've persuaded myself only for double 2683's)

 

[sciff]

randir, that's what I get:

Spoiler: Prime95's Torture test small FFTs no AVX/AVX2

Click to enlarge

Guys!
One thing puzzles me a lot. As you've seen before, this CPU can consume up to monstrous 340 W in Prime95 Torture test with Small FFTs. However, in Corona Benchmark and Corona Renderer it consumes only 240 W at x33 all-core multiplier (in other tests and applications, that don't use AVX, it consumes even less).

So why the hell it doesn't go to x38 in Corona and other less power-hungry applications, when 240 W is not even close to its maximum? The motherboard, obviously, is capable of providing the necessary power!

You can call me greedy, but I'm not. It's just that I see that this CPU/mobo setup has even greater potential and this keeps bothering me.

I doubt that this CPU would go up to those 340 W if all-core multiplier was x38. Power consumption would probably be somewhere between 240 and 340 W. It means that the temperatures could still be bearable for my particular cooler, Noctua NH-D15.

How to achieve this? Is there some way to get x38 all-core multiplier by EFI driver modification?

 

[sciff]

10-15 mins, if you use balanced power plan - it can drop to x12 (if you select minimal CPU state at 99 % it will stay at x20 without any load and up to 2,9 for all cores or 3.5 for few cores under load). Однако, здравствуйте 🙂

I would suggest using minimal CPU state of 5% like I do. Then, just like you said, the CPU will drop to x12 multiplier in idle mode. You won't loose any performance under full load by doing this.

 

[randir]

How to achieve this? Is there some way to get x38 all-core multiplier by EFI driver modification?

Can you try the following two EFI drivers? There's two modifications of the original V3_2 EFI with some additional msr hackery. I've compiled both v39 and uCode-less versions (I run v39 version on my system). They include 50mV vcore undervolting which is stable for me but may be unstable on your system - if you get any weird crashes, I can recompile without it.

!!! NOTE FOR ANYONE DOWNLOADING THOSE !!!

Use at your own risk! Ensure that your CPU has adequate cooling and monitor temperature if you decide to stress-test it! This breaks all in-system power consumption monitors - do not rely on them!

https://ufile.io/skgjj
https://ufile.io/1aiek

 

[sciff]

randir, thank you, both EFIs give plus x1 improvement!

This breaks all in-system power consumption monitors - do not rely on them!

Results:

0x39 microcode: x34 (instead of x33) in non-AVX applications and x29 (instead of x28) in AVX applications
By the way, this version causes HWInfo's core multiplier bars to be slow in response. No microcode version doesn't have this issue. Weird, huh?

No microcode: x34 both in non-AVX and AVX applications

Spoiler: v3x2_50_39_vcc

Spoiler: v3x2_50_vcc

No stability issues experienced, however I tested Prime95's Torture test only briefly, for obvious reasons 🙂

But why does it say Error at startup?

 

[sciff]

randir, is there some way to get x35 uncore multiplier in 3v2.efi, instead of x30? A friend asked if that's possible 🙂

 

[randir]

But why does it say Error at startup?

That's the core msr trickery failing( I can think of two possibilities:

- you've set a vccin override in bios that sets up a lock afterwards - try setting it to auto if you have such option
- you've done only a soft reset - try to shutdown and then load this driver to start from a clean cpu state

 

[randir]

What are you trying to say? That power consumption shown will not take into account additional +50mV?

No, with "error:vccin" you have everything as usual (and those 50mV are correctly accounted for). You'll see what happens if you get it working)

is there some way to get x35 uncore multiplier in 3v2.efi, instead of x30? A friend asked if that's possible

You can write to msr 0x620, say, 0x2323. On any one core - it propagates to all other. This sets uncore frequency limits to 35-35. It's successfully written & read back even on my system (with 30 mult max), but I'm not sure that this has any real world effect - lowering definitely works, but high values are capped as it seems.

This msr register is not locked and can be written to at any time - so you can give it a try, but I won't be too trilled about it.

 

[Alex S]

Can you run mprime (prime95) torture test with small ffts and then check frequency? I find it the most CPU demanding benchmark, far more demanding than linpack/renderers.


Supermicros are unlockable, though they require a manual BIOS editing. The only problem with them is that the BIOS lacks any overclocking features - BCLK, voltages, memory clocks, etc. So, you'd have to use external software for that - XTU/throttlestop do the job for windows (no BLCK still), but you're out of luck on linux unless you want to dig into msr registers progamming. They also have a very strange fan management, but that's solvable.

Seems i'll just use ASUS Z10PE-16 without WS. These have conflict between videocard and 1st CPU DIMMs, but videocard can be rised to 2,5-3 cm with simple riser. No fancy design, no sound, no BCLK, but these seems same, as their expensive WS cousins.

Is that new EFI driver will work in single socket mobo? I'll check that too.

 

[Alex S]

Wow. x30 in Cinebench (and slightly better scores) and LuxMark. u39 VMWare + new v3x2_vcc. x27 in Corona (but it's ok, not 25-26 as before). Same in Y-Cruncher. Author is genius!

 

[sciff]

That's the core msr trickery failing( I can think of two possibilities:

- you've set a vccin override in bios that sets up a lock afterwards - try setting it to auto if you have such option
- you've done only a soft reset - try to shutdown and then load this driver to start from a clean cpu state

I don't have such an option in BIOS. Everything else involving voltage is set at Auto.

I tried to do complete shutdown with no luck afterwards.

Enabling PLL Overvoltage didn't help either 🙁

Seems i'll just use ASUS Z10PE-16 without WS. These have conflict between videocard and 1st CPU DIMMs, but videocard can be rised to 2,5-3 cm with simple riser. No fancy design, no sound, no BCLK, but these seems same, as their expensive WS cousins.

Have you considered ASRock EP2C612 WS? Good mobo, and cheaper than both of ASUS WS's

Is that new EFI driver will work in single socket mobo? I'll check that too.

It does 🙂

 

[Alex S]

Have you considered ASRock EP2C612 WS? Good mobo, and cheaper than both of ASUS WS's

Yes, but this MB have only 4 DIMM socket per CPU. Right now i have only one CPU and currently using combo - 4x8GB + 4x4GB, totally 48GB, and with EP2C612 WS i'll only have 32GB. And there is no such thing as "enough memory", as you know 🙂

 

[Dingmel]

Hey guys. Sorry to digress a little. Has anyone had luck getting this to load each time automatically during bootup with bcfg? in my case, it registers fine, and shows up in the drivers list, but only autoloads probably 1/10 of the time. very weird.

 

[Alex S]

Hey guys. Sorry to digress a little. Has anyone had luck getting this to load each time automatically during bootup with bcfg? in my case, it registers fine, and shows up in the drivers list, but only autoloads probably 1/10 of the time. very weird.

First, check if your system is not changing drive order. For example, i have a RAID0 array of four SSDs and after every BIOS visit i must add load string to EFI again, or driver will not load after next reboot. (Mine EFI boot partition is fs2).

 

[Dingmel]

First, check if your system is not changing drive order. For example, i have a RAID0 array of four SSDs and after every BIOS visit i must add load string to EFI again, or driver will not load after next reboot. (Mine EFI boot partition is fs2).

Thanks for that suggestion.
I've been suspecting that my board may have been doing that, as my system has multiple drives attached, including a 4 drive RAID0 included.
I've tried copying the driver EFI to several drives, and adding multiple BCFG entries for some kind of redundancy just in case there is a problem loading, but I still can't get it to run consistently. Any suggestions?

I'm on a Rampage v

 

[Alex S]

Thanks for that suggestion.
I've been suspecting that my board may have been doing that, as my system has multiple drives attached, including a 4 drive RAID0 included.
I've tried copying the driver EFI to several drives, and adding multiple BCFG entries for some kind of redundancy just in case there is a problem loading, but I still can't get it to run consistently. Any suggestions?

I'm on a Rampage v

Try to add them to EFI boot partition and avoid BIOS resets and even settings. Working here, so might work for you too.

 

[Dingmel]

Try to add them to EFI boot partition and avoid BIOS resets and even settings. Working here, so might work for you too.

Yup, have them on the EFI boot partition, which is usually FS1:, and added them to FS2: which would be the master drive where the OS resides. Still no luck. I'm perplexed 😕
On a sidenote, rebooting from within windows, sometimes gets the driver to load on its own, versus coldbooting, or rebooting from outside windows. No idea what's causing it to randomly load/unload

 

[randir]

I don't have such an option in BIOS. Everything else involving voltage is set at Auto.

I tried to do complete shutdown with no luck afterwards.

Enabling PLL Overvoltage didn't help either 🙁

As an alternative, you can try to enable ThrottleStop's PowerCut feature on a clean system (without any efi's loaded). If it also doesn't work than it's something to do with the BIOS.

 

[GTZ]

Thanks for that suggestion.
I've been suspecting that my board may have been doing that, as my system has multiple drives attached, including a 4 drive RAID0 included.
I've tried copying the driver EFI to several drives, and adding multiple BCFG entries for some kind of redundancy just in case there is a problem loading, but I still can't get it to run consistently. Any suggestions?

I'm on a Rampage v

Have you tried disabling fast boot and secure boot? Those two setting did the trick for me for the driver to auto load.

 

[Dingmel]

Have you tried disabling fast boot and secure boot? Those two setting did the trick for me for the driver to auto load.

already have. the issue isn't that it isn't loading, but rather it is loading intermittently.

 

[Frakkls]

Hello! I saw you achieved 3.4 Ghz on all 18 cores with the e5-2696 v3.
I have the same processor and my Asrock x99 Taichi board arrives soon with a noctua d15-like cooler.
Can you give some information about how I can achieve the same and if there are some disadvantages to it please?
I'm new to all this stuff. I read some pages on this forum but I'm not sure if I can figure it myself.
Thank you!

 

[itou31]

Hi sciff, randir,
Thanks for the new efi drivers, I have the same issue vccin error as sciff. We have the same board and CPU, but mine go only up to x32 with the new efi, but the ratio still fluctuates even after 10min, not as before (and twice power consumption while idle (about 80W in HWinfo)).

 

[Alex S]

Had anyone noticed any instabilities with -50 mV offset driver?