My new 4770k rig - temperatures

[Aikouka]

Shortly after starting this thread the pump on my Corsair H100i went.

However it did show me something with regard to the Haswell that I would probably not have realised otherwise:



So all in all the first few days of the new 4770k system have been interesting to say the least.

What that tells me is that whatever you were using for temperature reporting was probably lying to you. I have an issue where my pumps won't turn on if I have them plugged into the PWM header (they are PWM-based pumps). I checked my temps in the BIOS, and I could easily watch them gradually rise (maybe 1C every 1-3 seconds) from a starting ~40C up to about 55C where I'd normally shut the machine down.

If you're running it without any cooling, that CPU is most likely throttling itself to keep it from burning up.

 

[Nec_V20]

What that tells me is that whatever you were using for temperature reporting was probably lying to you. I have an issue where my pumps won't turn on if I have them plugged into the PWM header (they are PWM-based pumps). I checked my temps in the BIOS, and I could easily watch them gradually rise (maybe 1C every 1-3 seconds) from a starting ~40C up to about 55C where I'd normally shut the machine down.

If you're running it without any cooling, that CPU is most likely throttling itself to keep it from burning up.

What happens with the Haswell processors is not only can they throttle down, but also they can vary the voltage. So when doing just normal Windows things which don't really require a lot of CPU time, such as Internet Browsing, the voltage is at a really low level and the cores will only ramp up the CPU multiplier as needed per core and then ramp back down again.

As far as I know this ability of the CPU to vary its own voltages as well as the clock multiplier is unique to Haswell

 

[calc2]

I wonder how many of these 4770K oc'd systems really are stable after having run linpack 11 on my system. If I increase the problem size/dimensions on my system, eg 60000/60000, it will quickly throttle even at stock settings with a Noctua NH-U12P SE2. I've seen examples of people running at high overclocks with very low GFlops numbers, I get in the range of 190-230 GFlops depending on speed. Even after delidding, and lapping the bottom of the ihs it still exceeds 90C at even low overclock speeds. Part of the reason is that running with larger problem sizes makes it clear higher voltages are needed or it will fail, eg 1.200v at 3.5GHz, when set to not adjust for voltage for AVX2.


- I lost my old account I had for around a decade, so now I get to start over with this 'first' post.

 

[supermag24]

Well, temp issues aside, congrats on the sweet i7 build. Shouldn't have to upgrade your cpu for years.

 

[Kenmitch]

I wonder how many of these 4770K oc'd systems really are stable after having run linpack 11 on my system. If I increase the problem size/dimensions on my system, eg 60000/60000, it will quickly throttle even at stock settings with a Noctua NH-U12P SE2. I've seen examples of people running at high overclocks with very low GFlops numbers, I get in the range of 190-230 GFlops depending on speed. Even after delidding, and lapping the bottom of the ihs it still exceeds 90C at even low overclock speeds. Part of the reason is that running with larger problem sizes makes it clear higher voltages are needed or it will fail, eg 1.200v at 3.5GHz, when set to not adjust for voltage for AVX2.


- I lost my old account I had for around a decade, so now I get to start over with this 'first' post.

Haswell overclocks, temps seem to be scattered all over the place compared to past generations. The new double lotto system sucks!

Not all of them need massive voltage to be stable. Guess it's just a matter of clockspeed goals.

Here's a experiment I did with 1v with my rig in sig.

Core temps produced 63,66,67,62 by Linpack 11 run below.

Results look in the norm for the clockspeed.

 

[BallaTheFeared]

I use Prime95 26.6 (for the first 3 hours, than I move on to the real test, games) because it was the last version with 0 support for AVX code paths.

 

[JimmiG]

I wonder how many of these 4770K oc'd systems really are stable after having run linpack 11 on my system. If I increase the problem size/dimensions on my system, eg 60000/60000, it will quickly throttle even at stock settings with a Noctua NH-U12P SE2. I've seen examples of people running at high overclocks with very low GFlops numbers, I get in the range of 190-230 GFlops depending on speed. Even after delidding, and lapping the bottom of the ihs it still exceeds 90C at even low overclock speeds. Part of the reason is that running with larger problem sizes makes it clear higher voltages are needed or it will fail, eg 1.200v at 3.5GHz, when set to not adjust for voltage for AVX2.


- I lost my old account I had for around a decade, so now I get to start over with this 'first' post.

Yeah, that "AVX boost" Intel built in to Haswell really *is* needed.

This is why I've changed to adaptive voltage now (+0.08V). In normal, non-AVX loads it tops out at 1.245V and temps are in mid 60's. With AVX loads it hits 1.323V, which is a bit higher than it needs to be, causing temps to get really high with the Aida64 FPU test.

The other option would have been ~1.3 - 1.31V fixed voltage. I figure it's better for the CPU to stay at or below 1.245 (EIST activated) 99.9% of the time, than to dial in a constant 1.3xV just for the AVX test.

Honestly, I no longer care about Linpack 11. If it throttles even at stock clocks *with* an aftermarket cooler, then why bother making your overclocked chip pass the test? The Aida64 FPU-only test is already ridiculously intensive compared to normal use. My criteria is to not throttle with that test.

 

[BallaTheFeared]

My thought is that we don't need to test for max temps, because of the throttle on Haswell. Which is really the only thing these tests are good for, there is nothing real world about the stress they place on a processor.

Also if AVX2/FMA3 take shape in the form that is present in Linpacks with everyday software we're not even going to need to overclock, the performance gap Haswell opens up over Sandy/Ivy in these tests is outstanding even comparing stock Haswell vs OC SB/IVY.

That said the only thing I've run so far that supports AVX/AVX2/FMA3 is Handbrake, and I haven't needed more voltage than I need to game for that application.

 

[JimmiG]

I guess Handbrake is limited to some extent by RAM and drive performance, so even though it uses AVX instructions, it's not as intensive as the more artificial Linpack test. It's similar to how Prime95 with larger FFT sizes runs 20C cooler than smaller sizes that fit in L1/L2 cache.

 

[calc2]

With further testing it seems AVX2 needs more voltage but perhaps a uniform amount of voltage until very high overclocks, ie:

3500 - 1.200v (1.175v failed) - (forgot to write down temp)
4000 - 1.200v - 97C
4100 - 1.210v - 98C

I changed the tests to be 4 runs of each 5k-60k in 5k increments. So finding an old pre AVX2 version of linpack and burning in with it with offset voltage may be a good way to go.

It took me a while to track down a later linpack 10.3 version so I linked them below in case anyone else was interested in them.

http://registrationcenter.intel.com/irc_nas/2592/w_lpk_p_10.3.10.017.zip
http://registrationcenter.intel.com/irc_nas/2592/l_lpk_p_10.3.10.016.tgz

It looks like you may need the settings from the 11.0 version for KMP_AFFINITY, at least under linux 3.11 otherwise it appears to use the wrong cores by default. I doubled my GFlops output by modifying the script on 10.3 to use the settings below from 11.0.

export KMP_AFFINITY=nowarnings,compact,1,0,granularity=fine