PPD/watt & Intel turbo boost

[biodoc]

I've finally got around to test the impact of intel turbo boost on PPD and PPD/watt on my dial socket Ivy bridge machine.

Hardware:

Dual socket motherboard
2x E5-2690 V2 (Ivy Bridge)
TDP for each processor =130W
Total of 40 logical cores
Turbo boost disabled: All cores at 3 GHz
Turbo boost enabled: All cores at 3.3 GHz
GTX980 (idle)

Application:

TN-Grid AVX on all 40 threads (full load)

Power draw from the wall:

Turbo boost on: 390 watts
Turbo boost off: 310 watts

PPD (calculated from 20 completed tasks):

Turbo boost on: 30136
Turbo boost off: 28455

PPD/watt:

Turbo boost on: 77.27
Turbo boost off: 91.74

So with turbo boost on I see a 10% increase in all core clock speed, 25.8% increase in power draw, 5.9% increase in PPD and 15.8% decrease in PPD/watt.

I guess the logical thing for me to do is disable turbo boost in the bios except maybe in races.

I was a bit shell-shocked to see this machine drawing nearly 400 watts from the wall. I suspect a non-avx app would draw significantly less power though.

 

[StefanR5R]

My Broadwell-EPs generally pull more power than their TDP too, because the BIOSes + my BIOS settings cause them to run at all-core turbo all of the time when fully utilized, no matter the power draw. I have them running in a desktop mainboard from Asus, and in "frequency optimized" server/workstation mainboards from Supermicro.

So far I liked the fact that the CPU clock is predictable --- and therefore, that performance is predictable (mostly, there is still variation from AVX2 vs. non-AVX2 turbo, as well as from periods when only few of the cores are used) --- more than I disliked the decrease of power efficiency. Since these are comparably low-clocked 14 nm CPUs, they are still quite efficient even with turbo.

However, with electricity rates only tending to increase every year, I should probably follow your example, revisit what BIOS options I have available, and optimize PPD/Watt for projects which I preferably run.

 

[biodoc]

Anandtech published an article explaining how intel measures TDP for it's processors. My take from that article is that TDP is roughly the power draw with turbo boost disabled. My dual socket system draws 93 watts at idle and 310 watts under full load with turbo boost disabled. The TDP for the processors is 130 watts x 2 for a total of 260 watts. Considering the power draw of the MB and GTX 980 at idle and difference of 50 watts (310 watts acrual - 260 watts total TDP) is not unreasonable.

 

[crashtech]

That also may put a different spin on CPU selection as well. Is there any consensus about what clockspeed on, for instance, Haswell cores is required to properly feed a bigger GPU like a 1080ti?

 

[TennesseeTony]

For Folding@Home, 1080Ti/10-series just keeps putting out more PPD with higher clockspeed in my experience.....3.4GHz is my personal max Haswell CPU speed... (hard to believe I am so outdated already.....)

 

[emoga]

@biodoc Are you still using turbo boost or have you turned it off and never looked back?

I was thinking of summerizing my machines and this seems like the most logical thing I could do.

 

[biodoc]

@biodoc Are you still using turbo boost or have you turned it off and never looked back?

I was thinking of summerizing my machines and this seems like the most logical thing I could do.

I turned it on for the pentathlon but switched it off again after the race. I'll probably keep it off now "forever".

 

[Assimilator1]

Interesting thread, I was wondering recently how much power savings I would make if I sold my i7 4930k @4.1 GHz & bought a coffee lake i7 8700k (which also has 6C/12T), until I saw it's price that is!
But anyway, a THG article had the 8700k drawing ~160w running P95 with AVX fully on, whilst AT's bench has 2 hugely different figures between the 2 CPU benches!
CPU 2019 has it drawing 151w (running who knows what?), whilst CPU (no year) shows it drawing 122w, with the i7 4960X (more similar to my o/c 4930k) drawing 118w! (probably from the Skylake x review figures). What's going on there? Surely the coffee lake cpu is going to draw less power?, even with it's higher clock?

Also, anyone know why the big power difference between the 2 benches? Reading the 2 AT articles about it, the figures were different there too, with the original 1 showing just 86w! Wth?? And no explanation in the later tests as to why! The other matching the CPU 2019 bench.
Although the 8700K is going to be roughly 25% faster than my rig's CPU (via IPC & clock gains), it seems I wouldn't be saving anything on electricity, possibly spending more!

I'm going to dig out the power draw figures for my own rig, to throw into the conversation, once I've found them!
Anand's original review of the 4960X shows a total system power draw of 210w for it running x264 HD 5.

 

[crashtech]

I don't have anything specific to add, but re CFL power draw, I wouldn't discount the effect of the higher clock, the power vs speed graph is not linear at all. Plus a lot of motherboards are effectively OCing the 8700K by default, applying all-core max Turbo and extra voltage to boot.

 

[Assimilator1]

Ah ok, so it could be possible to lower the voltage with the same clocks to lower power usage! Actually, now that I think about it, I think an AT review said something along those lines.

Btw, dug up my paperwork for my main rig, although it turns out I didn't power test it with DC on the final settings, I got close enough.
So at 4.04 GHz with the same vcore as now (1.13v setting), it was pulling 224w running 'SETI v8 avx' on 12 threads, OCCT ~247w.
At 4.1 GHz with OCCT it was pulling ~250w.

I've just realised that it's actually running at 4.125 GHz atm, I don't remember nudging it up a little , vcore is the same IIRC. Hmm, although CPU-Z is showing 1.3v atm, .....
I do remember that I hit temperature limits before I hit my voltage limits, I think their might be room to drop vcore at the same speed, I was going to try that a few years ago

[update] So it turns out that somewhere along the line the Asus bios decided to whack up the vcore from 1.13 v to 1.3v, & CPU VCCSA from 1.1 to 1.3v too! Grrr, I wonder how long it's been like that & how much that's cost me in electricity!!?
So I've dropped the voltages back down now, and (unsurprisingly) it runs MUCH cooler, only tested in DHEP but CPU temps dropped by 10-13C! And power has dropped from 268w to 241w (close to the original test with OCCT).
Funny thing is I also don't remember (& have no record of) setting the 'FSB' to 125 MHz....... looks like I will need to do some stability testing.