High temperatures with i7-4770 and LINPACK, is CM Hyper 212 EVO enough? (no OC)

[enra]

Hi,

I'm having some temperature issues when running linpack-heavy numerical optimization code on my i7-4770 (thus, no-OCd) with stock heatsink and AS5 compound. Usual temperatures when idle are ~28-35°C and at full load with regular FPU/integer computations around ~70-85°C (depending on how many threads in parallel, HT). So far, so good.. but with MKL/linpack etc. (presumably AVX2 instructions) it hits temperatures up to 95-100°C almost instantly. The heatsink is mounted correctly, no issues with that, and there are no general issues with airflow etc. I found a previous thread on Linpack and Haswell, but with an overclocked i7-4770K (http://forums.anandtech.com/showthread.php?t=2333848).

I'm considering getting a CoolerMaster Hyper 212 EVO to replace the stock heatsink/cooler. I don't really want to invest much more, given that it's not OC'd etc. Does anyone have experience with that and can tell me if that cooler is capable of keeping the temperature at a more acceptable level?

Any other suggestions are welcome as well!

Thanks,
enra

 

[fastamdman]

The stock heat sink is absolute garbage, no matter what paste you are using. I would highly recommend an aftermarket heat sink. The hyper 212 evo+ is good to roughly 4.5ghz with decent temps depending on your system.

 

[tviceman]

Hyper 212 will probably get you down to upper 70's or low 80's with the linpack run, all other things being equal.

You may also find that you can undervolt your chip and shave off 5-7 degrees.

 

[Phynaz]

If you're not over clocking the stock heat sink is perfectly fine.

 

[enra]

Thank you all for your replies!

I think I'll get the CM Hyper 212, but first I'm going to experiment with fixed Vcore settings (currently set to "auto" in bios, my MB is a GA-H87-D3H). It seems that Vcore is adaptively increased when AVX2 is in use, which results in these extreme temperatures (I'm seeing spikes in my lm-sensors vcore output). From reading different threads it seems to be notoriously hard to run a fully stable OCed system with even much better heatsinks when testing load using AVX2 instructions for longer time spans (e.g., linpack 11). Unfortunately, that's pretty much the kind of workload I'm expecting with the code I'm running, but I'm running it on a non-OCed system. Evidently the stock cooler is not enough even without OC with these "auto" adaptive vcore increases.

 

[BonzaiDuck]

Thank you all for your replies!

I think I'll get the CM Hyper 212, but first I'm going to experiment with fixed Vcore settings (currently set to "auto" in bios, my MB is a GA-H87-D3H). It seems that Vcore is adaptively increased when AVX2 is in use, which results in these extreme temperatures (I'm seeing spikes in my lm-sensors vcore output). From reading different threads it seems to be notoriously hard to run a fully stable OCed system with even much better heatsinks when testing load using AVX2 instructions for longer time spans (e.g., linpack 11). Unfortunately, that's pretty much the kind of workload I'm expecting with the code I'm running, but I'm running it on a non-OCed system. Evidently the stock cooler is not enough even without OC with these "auto" adaptive vcore increases.

Unless something has changed, it seemed over a period of years the multi-core chips had a margin of error for any given sensor of +/- 5C or 6C.

Do you get your numbers from the highest core value?

For the price, the 212+/EVO is supposedly good for modest overclocking, even if that's not what you want to do. You obviously know what you're doing with the stress-tests. I just made sure the software revisions include the AVX instructions. Didn't think much more about it .. .

 

[Skurge]

I had a hyper 212EVO on my 4770K at stock and running an AVX2 workload would put it neat 90C. Although it is summer over here and I don't had airconditioning. So ambiants were pretty high.

In a none AVX2 workload, temps were much lower. in the 70s

 

[VirtualLarry]

If you're not over clocking the stock heat sink is perfectly fine.

Obviously not...

 

[Magic Carpet]

Same here. Easily gets to 100c plus but with a passive heatsink. Luckily, there aren't many AVX workloads around, except for benches.

 

[ShintaiDK]

Also make sure you mobo doesnt autooverclock, when you wish to compare stock temps. Specially turbo OC seems to be a defaukt for many mobo makers.

 

[coffeejunkee]

Haswell cpu's automatically increase vcore when running AVX loads.

When you say "when running linpack-heavy numerical optimization code" do you mean you actually run AVX optimized software or just some stresstest like Linpack?

If it's the latter, simplest solution is to not run Linpack. If it's the former, a 3rd party cooler will help.

Edit: read your post 5. Seems you're serious about AVX. I'd probably go for something beefier than Hyper 212. I recommend Thermalright Macho or True Spirit 140 from Amazon. Both around $50 but need to check case width for TS 140.

 

[Kenmitch]

Hyper 212 will probably get you down to upper 70's or low 80's with the linpack run, all other things being equal.

You may also find that you can undervolt your chip and shave off 5-7 degrees.

My 4670k I recently sold would pump WAY MORE vcore running stock speeds than was required for stability. I was running it 24/7 at 4ghz with a fixed vcore of 1.06v's.

I suggest playing around with your chip too see what she'll do. Even just bumping it up to the stock 1 core turbo speed under 4 core load would be a decent speed boost....Most likely you'll be able to shave off some vcore if you fiddle around with it.

 

[Phynaz]

Obviously not...

I suppose if the purpose of his computer is to run Linpack all day then it's not. For normal workloads it's fine.

 

[JoeRambo]

Unfortunately, that's pretty much the kind of workload I'm expecting with the code I'm running, but I'm running it on a non-OCed system.

I would not worry about it before testing it. Real world code is not equal to Intel's optimized Linpack library. Your code, even if it is using FMA, will have branches, cache misses and so on and will probably not multithread good enough to keep 4 cores loaded 100% of time.

So test real code first and worry after. Linpack and related highly uber tuned and optimized power viruses have little to do with real world loads.


P.S. Even running Intel's optimized math libraries we are far away from Linpack loads.

 

[enra]

Apologies, I should have made myself clear wrt. what I'm doing, or what this computer is being used for.

I never actually tested it with Linpack benchmarks or any other benchmark software; I'm seeing these temperatures running statistical simulations using my own (vectorized) MATLAB code, which (presumably) heavily utilizes Intel Math Kernel Library (MKL) routines (I erroneously referred to Linpack in my original post; what I meant was linear algebra routines contained in the BLAS and LAPACK libraries).

I previously ran other software which doesn't use these optimized routines, but "normal" integer/floating point math instructions, which gave me the ~70-85°C when using all cores under full load.

@BonzaiDuck, I agree, it may be that the sensors don't report accurate values. I'm getting my readings from the coretemp linux kernel driver via lm-sensors. During these spikes all 4 cores report between 95°C and 100°C :-/. According to the coretemp docs:

Temperature is measured in degrees Celsius and measurement resolution is 1 degree C. Valid temperatures are from 0 to TjMax degrees C, because the actual value of temperature register is in fact a delta from TjMax.

(TjMax apparently is 100°C for desktop haswells)

@JoeRambo, you are right, of course, I doubt my code fully saturates all cores at all times. What I'm seeing is mostly short spikes for at most a few seconds. However, I'm expecting this code to run for days on end, given the size of my dataset, which makes me quite worried. I stopped running it after seeing these spikes for the first couple times, though I know the CPU is supposed to throttle down.

@Skurge thanks, that sounds quite promising!

@coffeejunkie thanks for your suggestions! I'll look into Thermalright Macho or True Spirit 140. Though I'm in Australia, which limits my options..

@Kenmich yes, lm-sensors reports vcore at 1.21 even before I get to the code that drives temperatures close to 95-100°C (!).

@ShintaiDK thanks. The Turbo Boost options were set to auto in BIOS, resulting in multipliers of 39/39/38/37 for 1/2/3/4 core loads.

 

[coffeejunkee]

@Kenmich yes, lm-sensors reports vcore at 1.21 even before I get to the code that drives temperatures close to 95-100°C (!).

I'm not really familiar with Linux but are you sure lm-sensors is showing the vcore and not the vid?

If so, 1.21 vcore is quite high for Haswell so I suppose even though you don't run Linpack your workload appears to trigger the voltage bump as well.

Normal loads should be around 1.05 or so, assuming sustained load (vcore will fluctuate a bit).

But Kenmitch is right, there's usually a fair amount of undervolting you can do. If you use fixed vcore it might even be enough for the stock cooler to handle (it will still have to work hard though). I'd also do some extended stability testing with many different programs.

 

[VirtualLarry]

But Kenmitch is right, there's usually a fair amount of undervolting you can do. If you use fixed vcore it might even be enough for the stock cooler to handle (it will still have to work hard though). I'd also do some extended stability testing with many different programs.

Just like I wouldn't suggest overclocking on production code, neither would I suggest undervolting. Leave the CPU voltages at stock, and improve the cooling, is what I would do.

 

[enra]

Thank you all, again!

I followed VirtualLarry's advice and did not undervolt. I also did not set vcore to a fixed value. Instead I got a used Thermalright HR-02 heatsink with a 12cm fan (turned out to be the best price/performance and cheaper than a new CM Hyper 212 Evo; everything in Australia is expensive..).

I'm keeping a log of vcore and temperatures when running my code. Vcore is automatically adjusted to up max 1.28V; at ~1.10-1.22V sustained for more than a minute on average. Temperatures vary between ~60°C (1.10V) and ~80-87°C (1.22V).

Overall I'm optimistic about this solution. If I encounter even higher temperatures I'm going to get a better (14cm) fan and another case fan.

Edit: I should add that this is with the stock turbo boost multiplier of 37 on each core for a 4-core load, i.e., 3.7GHz per core.

 

[Bubbleawsome]

Sweet jeebus, does no one here know about fan control? My i7 870 runs up to 1.3v and on the stock cooler would run up to 99c, I add in fan control (tornado 4k rpm 80mm fan) and max load is 77c. Now I have an aftermarket cooler, max load ~80c, max fan 62c. And still quiet on a 120mm fan. (Don't know RPM.)

 

[enra]

@Bubbleawsome: Of course, but I'm not sure how that would help me with my issues.

After running my code for an extensive amount of time, I saw Vcore being sustained at ~1.21V over the whole time (i.e., never goes much below that; no switching into higher c-states). This code is going to run for days. I'm not sure how fan control would help, except for running the fan with sustained high speed, or maybe shortly alternating max_speed<->lower_speed.

 

[Bubbleawsome]

I thought the issue here was with temps. Later it goes to voltage, which can be partially solved with temps. Now it looks like it might be code?

 

[coffeejunkee]

Nice deal, excellent heatsink. Much better temperatures now, still pretty high but it is to be expected with such a high vcore.

Now the question is, why does Haswell need so much vcore for AVX loads (or does it even need it)? Predecessor Ivy Bridge has AVX support too but didn't increase vcore at all. Can you run some other software like video-encoding, to see what your 'real' vcore is?

Normally I would agree with Larry about not undervolting/overclocking a production machine but in this case I would make an exception. Most Intel cpu's have generous margins, I run at a massive 0.1V undervolt at the moment, everything working fine. You could start with something like 0.025V to see if it helps out. But like I said, check for stability thoroughly.