What controls Turbo Core in Xeons?

[MOF]

RC5 binaries are now up:
https://github.com/freecableguy/v3x4/releases/tag/v0.1b-rc5

Awesome!!.

"v3x4_90_50_1pt650pc" everything works very well. (Powercut doesnt work after soft resets but full resets ok!) I got almost 3000cb on Cinebench.

"v3x4_90_50_1pt950pc" and "v3x4_90_70_1pt900pc_cnq" (Powercut never worked. Soft, hard, poweroff... i try everything. Powercut didnt worked.)

I would like to see 90-70, 100-50, 110,70 versions.

 

[MOF]

 

[Dufus]

Nice voltage for 3.4GHz. Are those GFLOPS at 3.4GHz? You should run one thread per core and at that problem size and 3.4GHz should be getting over twice the GFLOPS you have.

I don't use powercut myself but you may find the VCCIN is limited on some chips so when setting it should check the mailbox return code is okay (zero) and not in error and also check before hand it is not locked already if not already done that way.

 

[MOF]

Are those GFLOPS at 3.4GHz? You should run one thread per core and at that problem size and 3.4GHz should be getting over twice the GFLOPS you have.

You right but 8gb (4+4) ram doesnt allow bigger problem size. Also dual channel memory bandwith make bottlenecking.

 

[wiseguyuk]

You probably need one of these:
https://www.amazon.com/gp/product/B01I1EU9LG/ref=oh_aui_search_detailpage?ie=UTF8&psc=1

Ultimately, this is how I flashed the BIOS for my system... remove chip, program, replace.

Thank you , I have ordered it now.

 

[Micrornd]

Some amazing stuff going on in this thread.
I salute all those working on it.

I am by no means a programmer, but I did have a thought -
Since we have the ability to flash a bios from within Windows.
And we have the ability to change bios settings from within Windows.
And we can preload apps, preload scripts and run scripts prior to Windows loading (completely) -

Shouldn't it be possible for a well experienced Windows application programmer to basically automate what is happening in this thread and with a reboot (or 2), accomplish what you are doing?
Possibly incorporating voltage offset selection and whatever other parameters might be wanted.
It would definitely bring the ability to the masses (like myself and I am sure others) that are unskilled or unsure.

I realize there would need to be a separate version (??) for each MB, but it does seem like a money making opportunity for someone.
I, myself (and I am sure others), would be willing to "buy" an app like that, and since it would drive up the price of Xeon v3s, I'm guessing Those sitting on a pile of them (as on Ebay) might be willing to finance the building of the program.

Just a thought

 

[kjboughton]

Some amazing stuff going on in this thread.
I salute all those working on it.

I am by no means a programmer, but I did have a thought -
Since we have the ability to flash a bios from within Windows.
And we have the ability to change bios settings from within Windows.
And we can preload apps, preload scripts and run scripts prior to Windows loading (completely) -

Shouldn't it be possible for a well experienced Windows application programmer to basically automate what is happening in this thread and with a reboot (or 2), accomplish what you are doing?
Possibly incorporating voltage offset selection and whatever other parameters might be wanted.
It would definitely bring the ability to the masses (like myself and I am sure others) that are unskilled or unsure.

I realize there would need to be a separate version (??) for each MB, but it does seem like a money making opportunity for someone.
I, myself (and I am sure others), would be willing to "buy" an app like that, and since it would drive up the price of Xeon v3s, I'm guessing Those sitting on a pile of them (as on Ebay) might be willing to finance the building of the program.

Just a thought

Shhhhhhhh
Shut up

 

[Saren]

I tested with 0x1F, 0x27 and without ucode. I got best result without ucode.
Looks like stable.

My motherboard doesnt allow tdp limit increasing. 145w TDP limit allways there. Only with I can get better result with Randir's v3x2_50_vcc.efi and v3x2_50_39_vcc.efi drivers because those drivers blocking a tdp reading and shows 50w all time.

-0.090 volt is a very very good value, thats ok for me. Probably more than -90 will give a BSOD.

If you can make tdp reading mismatching like Randirs did, that will be perfect.

Same here, with my supermicro board, the EFI providing the best performance are Randir's vcc .efi drivers. No other efi can beat that.
If there are more vcc drivers that would be very great. Thanks.

 

[MOF]

Same here, with my supermicro board, the EFI providing the best performance are Randir's vcc .efi drivers. No other efi can beat that.
If there are more vcc drivers that would be very great. Thanks.

I got better result with kjboughton's latest v3x4_90_50_1pt650pc.efi driver

v3x2_50_vcc.efi = 3300mhz all core with ucode (without ucode goes 3400mhz but benchmark scores still same as ucoded 3300mhz )
v3x4_90_50_1pt650pc.efi = 3400mhz with ucode (without ucode makes solid allcore 3500mhz but benchmark result still same as ucoded 3400mhz)

 

[kjboughton]

RC7 binaries are now up for download:
https://github.com/freecableguy/v3x4/releases/tag/v3x4-0.10b-i306f2-rc7

Builds have no changes to VCCIN... I suspect most of you are using a motherboard which allows this to be set there... much better than doing with the driver as an additional reboot is required for this change to take effect.

More nuggets of wisdom:

1) ALWAYS, ALWAYS, ALWAYS boot from cold metal when *changing* drivers. This means shutdown and flip rocker switch on PSU or otherwise remove *all* power from the board for *at least* 30 seconds.

2) PowerCut does nothing IMHO. On, off... CPU clocks the same and performs the same amount of work. Only difference is you don't get to see what the CPU think's is going on wrt power.

3) If you compile to set fixed VCCIN the driver will warn you to reboot to take effect. In other words, if booting from cold state, once you get to Windows login, reboot for VCCIN to take effect. No need to do so on subsequent warm restarts.

4) Enabling C6 in particular allows the cores to go to *max* turbo. If you disable C6, at least on my 2-way board (ASUS Z10PED8-WS), core frequency will be limited to max *all-core* sustainable frequency. This isn't cool as gaming with much fewer cores does allow all cores to 1C max.

5) The more you can reduce VID for the Core domain (and the Uncore domain to a lesser extent) the higher your all-core maximum turbo boost. You will reach a point of instability where your OC fun will end.

6) Not withstanding point 5, you *may* find you need more VCCIN as you increase your negative core voltage offset. I recommend 1.900 ~ 1.950V max before going too crazy.

7) For me, Uncore negative voltage offsets > -50mV were simply *not* stable in the long term. Play with finding where's you're stable on your Cores before moving to the Uncore domain. If need be, we can trade cache speed for core speed for the final tweak later down the line.

8) Running no uCode is far from optimal. It's clear that the final PCU governor logic is not in place for initial silicon production. Cores *may* boost higher but all you're seeing is more power contention. On average there is no benefit to this approach. My recommendation remains uCode 27 (0x1F).

 

[Cata40]

D

Same here, with my supermicro board, the EFI providing the best performance are Randir's vcc .efi drivers. No other efi can beat that.
If there are more vcc drivers that would be very great. Thanks.

You're not right, all the efi are good. And those of Randir, yes and those of C-Payne and Kjboughton. I use everything and go well.
Everything can be even better, but let us be happy that these people help us!

 

[custom90gt]

RC7 binaries are now up for download:
https://github.com/freecableguy/v3x4/releases/tag/v3x4-0.10b-i306f2-rc7

Stupid question, will these work for single cpu setups?

 

[kjboughton]

Stupid question, will these work for single cpu setups?

yes, up to 2S

 

[MOF]

RC7 binaries are now up for download:
https://github.com/freecableguy/v3x4/releases/tag/v3x4-0.10b-i306f2-rc7

I try all, didnt work. Boot normally, but it freezes a few seconds later on the desktop.

 

[MOF]

2) PowerCut does nothing IMHO. On, off... CPU clocks the same and performs the same amount of work. Only difference is you don't get to see what the CPU think's is going on wrt power.

I think it changes from mobo to mobo. On my system powercut makes huge difference.

Left side v3x4_90_50_1pt650pc.efi (Powercut enable) - Right side Powercut disable
Same setting and same benchmark running on both.

 

[Cata40]

To me work v3x4-0.10b-i306f2-rc7 and v3x4-0.10b-i306f2-rc7_50x2. After that, all i get freeze
I don't know to do a powercut

 

[KW8cSgF53G]

8) Running no uCode is far from optimal. It's clear that the final PCU governor logic is not in place for initial silicon production. Cores *may* boost higher but all you're seeing is more power contention. On average there is no benefit to this approach. My recommendation remains uCode 27 (0x1F).

Update: missing considering a factor... See below for further information.

 

[Dufus]

You right but 8gb (4+4) ram doesnt allow bigger problem size. Also dual channel memory bandwith make bottlenecking.

Yes it's amazing how much that RAM can make a difference in things like Linpack and XTU bench. Still should see ~550 GFLOPS at 20014 (3072MB) and dual channel. Try setting affinity. APIC ID's for my processor are every odd logical CPU is a second thread so running something like 'Start "Linx" /AFFINITY 0x5555555555555555 Linx.exe'.

I see your having power limits but powercut is really a last resort IMHO. For instance with my Taichi and 2683 I cannot set VCCIN via the CPU above about 1.82V. The BIOS does however have an option to set VCCIN above that but this is configured externally, perhaps through the SIO, and when that happens anything set via the CPU is overridden. Your probably still current limited with powercut too so should make that reasonably high.

Shouldn't it be possible for a well experienced Windows application programmer to basically automate what is happening in this thread and with a reboot (or 2), accomplish what you are doing?
Possibly incorporating voltage offset selection and whatever other parameters might be wanted.
It would definitely bring the ability to the masses (like myself and I am sure others) that are unskilled or unsure.

I realize there would need to be a separate version (??) for each MB, but it does seem like a money making opportunity for someone.

Depends on the board, just setting variables might not be enough. Should have been the manufacturers making the possibility to start with but platform is too old now. Might have helped sales back then. Did start a BIOS mod myself but with the increase in CPU prices that has come to a stop and remains in stasis. I personally can no longer justify buying a 2696 and some other chips to continue for now. Perhaps some time later on when prices come down if I have not already sold everything by then.

Wouldn't count on the money making opportunity either.

Thanks, I agree, and i found that with the same EFI driver on idle status, when i measured the "temperature",
whitout microcode higher than with it for at least 5 degree Celsius (to CPU Cores and small variable to Packages)!

If using DTS please check your Tjmax with and without ucode. Possibly different as mentioned before. Usually better to use PECI temperture IMHO.

 

[KW8cSgF53G]

If using DTS please check your Tjmax with and without ucode. Possibly different as mentioned before. Usually better to use PECI temperture IMHO.

Yes, you got it: 95 and 90...

 

[kjboughton]

If using DTS please check your Tjmax with and without ucode. Possibly different as mentioned before. Usually better to use PECI temperture IMHO.

E5-2696v3, no ucode = 83C, with ucode = 78C

 

[Saren]

D

You're not right, all the efi are good. And those of Randir, yes and those of C-Payne and Kjboughton. I use everything and go well.
Everything can be even better, but let us be happy that these people help us!

I understand what you are trying to convey. However it's true that the new EFI's performance is not optimal in supermicro boards.

I am using X10DAL-I and a pair of 2696 v3. Since any kind of undervolting is unstable to my cpus,
with v3x2_vcc.efi and 0x3A microcode, I can get 3.1&3.0 with normal full load, and 2.6&2.6 with avx2 full load.
With other EFIs including the new ones, I can only get 2.6&3.0 with normal full load, and 2.4&2.6 with avx2 full load.
I think the TDP readings limited the CPU performances.

 

[Cata40]

I understand what you are trying to convey. However it's true that the new EFI's performance is not optimal in supermicro boards.

I am using X10DAL-I and a pair of 2696 v3. Since any kind of undervolting is unstable to my cpus,
with v3x2_vcc.efi and 0x3A microcode, I can get 3.1&3.0 with normal full load, and 2.6&2.6 with avx2 full load.
With other EFIs including the new ones, I can only get 2.6&3.0 with normal full load, and 2.4&2.6 with avx2 full load.
I think the TDP readings limited the CPU performances.

Like I said many times, I'm not a specialist, but I think. Can not make an efi, which leaves freedom for bios setting?
I mean, I mean those who do not have Asus boards, but others like Asrock, Msi, Gigabyte where we can not set this up
All those offset voltages should be left free to put them on but to unblock my motherboard.
Because it differs
At the Asrock OC formula along with my processor, good volts without freezing are -70 (-65) and -50
Is that something like that?

 

[Saren]

Like I said many times, I'm not a specialist, but I think. Can not make an efi, which leaves freedom for bios setting?
I mean, I mean those who do not have Asus boards, but others like Asrock, Msi, Gigabyte where we can not set this up
All those offset voltages should be left free to put them on but to unblock my motherboard.
Because it differs
At the Asrock OC formula along with my processor, good volts without freezing are -70 (-65) and -50
Is that something like that?

No. Supermicro boards do not have the option to ignore TDP limit. Using turbo hack makes the CPU reach the TDP limit, and motherboard will slow down CPU for that.
Randir's EFI is very special. It breaks TDP reading, that's why I can use his EFIs to achieve best performance.
@kjboughton did a very good work. However it lacks the special function to break the TDP reading. So, for supermicro board users, @kjboughton's EFIs are not optimal as Randir's.
I hope @kjboughton can add that function as an extra EFI, that would be perfect.

If you uses supermicro server board, you will understand my situation. Regular X99 boards are fine, but supermicro boards not.

 

[KW8cSgF53G]

No. Supermicro boards do not have the option to ignore TDP limit. Using turbo hack makes the CPU reach the TDP limit, and motherboard will slow down CPU for that.
Randir's EFI is very special. It breaks TDP reading, that's why I can use his EFIs to achieve best performance.
@kjboughton did a very good work. However it lacks the special function to break the TDP reading. So, for supermicro board users, @kjboughton's EFIs are not optimal as Randir's.
I hope @kjboughton can add that function as an extra EFI, that would be perfect.

If you uses supermicro server board, you will understand my situation. Regular X99 boards are fine, but supermicro boards not.

I'm in the same boat with you. Same motherboard, NO settings for multiple frequency, voltage, external clock, memory timing...

 

[Cata40]

I have asrock x99 oc formula, but at voltaic settings, I can not the only thing they can do is C-Payne, where they can set off the bios System agent offset, the rest are hand-bound (that's what works in windows) and they can set fsb!
So there is not much difference!