General > General Technical Chat
New PC build p0rn photos.
<< < (9/14) > >>
paulca:
Well, the fun I'm having with the cooling system and testing, tweaking, testing to see what the limitations are and just how hot it will run or how cool it will run, for each component and how they interact.

Using:
* Prime95 for synthetic CPU loading (of different flavours and levels of torture)
* Furmark @3440x1440 for synthetic GPU stress loading
* MSI Afterburner for detailed metrics and graphs (stupid name, great program, previously known as RivaTuner until it was bought by MSI)
* MSI Afterburner for GPU clocks, voltage, limiters tuning.
* Argus Monitor for fan/pump control, more detailed metrics and graphs
* 3DMark "Time Spy" benchmark for short term performance 'scores' and comparisons with other peoples scores, not for competition, but as actual ... well benchmarks.

First with all fans and pumps at max, maximum noise, all vents, doors and grills open to the dust produced a performance benchmark score of 15,960.  I soak tested the system with both prime95 and furmark and the GPU didn't get past 50*C, the CPU did get to 80*C (more on that later). 

Then with all the doors, vents and filters back in place and a "sensible" quiet fan profile which is almost silent when not under load, but turns into the tornado again when it sees GPU and/or CPU load.  That only dropped the benchmark score by 40 points.  Heat soaking it showed only a few degrees increase and virtually no effect on peak clocks, though I didn't test/graph average clocks across cores.  The "peak core clock" will be the "golden core" the rest will be less than that.

So, I tried to set up a "quiet gaming" profile.  Again, pumps, fans all low or off until load is detected, but this time not immediately spinning up to full performance mode, but slowly following the temperature not reaching 100% until 80*C on the GPU.  Similar profile for the CPU.  The benchmark was still 15,9xx and the GPU didn't get past 61*C even under full stress load 61*C with a fan speed of 25% and it was stable.  Wow.  Maybe I went OTT on the cooling LOL, a bit.

So, why the hell not?  I switched all the fans off and hit it in both the CPU and GPU with max load.  After 5 minutes, the CPU was at 90*C, but still holding peak core clocks of 4.7Ghz!  Damn psychopathic thing! However it was the GPU that was raising much more concern.  It was slowly creeping up through 70*C.  The GPU radiator had become too hot to hold my finger on.  When would it stop?  What temperature of water can the fittings, tubes and radiators handle (quick google, target under 40C, don't exceed 60C).  Abort!

This test was so interesting because it forced me to try and derive the water temperature from the core temperatures.  Alternating between 100% load and 0% load with no fan support produces a sloppy square wave on the thermals with an amplitude of, in my case about 30*C for the GPU and a huge 50*C on the CPU. Poking at the radiators with my mark one temp probe (my finger), it seemed fair that this amplitude more or less accounts for the thermal junction delta between the cores and the water (the main heat mass) in the loop and as it's circulating, the radiator.

So at 75*C which is where I chicken out, the water temp would only have been 45*C, too hot for me to stand in a shower that temp, so as it was uncomfortable under my finger after 10 seconds or so, I took 45*C as the finger in the air estimate.  Which is fine.

I'm tempted to see how far I can actually push that, can I get the card to hit the thermal limiter before the water cooling system blows a radiator fitting out.

But first.  I need a set of stick on thermal couples and a way to data log them on to the PC. (Sit down at the back Ardy folks!).  Don't want to push it higher without an actual measured temp on the radiator and/or block itself.

Back to the CPU.  50*C is too high, and higher still if the CPU gets up to 95*C the water temp is still under 40C.  I will have to remove the cooler, clean off the "came on it from the factory" thermal goo, add some "actually bought" stuff and see if that helps.  I fear it won't help much.  The issue is the heat spreader and multi die design.  Unlike the GPU which is a dead flat block of silicon to interface with, the CPU has a tin lid on it which is thermally bonded to the dies underneath and tries to spread the heat out and prevent hot spotting.  But it also introduces another thermal junction (or two) which do add up.  De-lidded CPUs (when they were single die) would usually lower their junction loss to 40*C or less.  I'm not going to start delidding it.  As long as I can use the fans to keep the radiator under 35*C a 50*C delta I can live with.
paulca:
Overclocking the GPU got me a little more benchmark score of 16,444 however it later turned out to be unstable.  +200 core and +800 RAM seemed to be stable, but to be honest, I don't think I'll bother to keep it overclocked.  It's hitting it's power limiter 99.9% of the time, so the boost clocks are limited by that anyway, if I clock it up, it will just clock itself back down again, the higher boost clocks will only effect light loads not hitting the limiter, but will most likely hit the voltage limiter and become unstable.

Much the same for the CPU, switch on it's auto AI overclocking gubbins and let it be.
paulca:
Now... I could pull the wool over it's eyes and flash a compatible BIOS from a "more expensive" card, like the OC version BIOS which goes to 110% power limiter.  Or the top of the range which goes to 125% power limit, I believe.  Trouble is, at some point something in there that wasn't specced for a 125% power limit card is going to melt or let magic smoke out and that's an £800 replacement as I'm not getting that one past RMI inspection.
Berni:
To get significantly more speed out of a graphics card you usually need to upgrade the cooling to keep up. Watercooling does wonders on those, but it can get rather complicated because you not only need to cool the GPU but also the RAM around it and the VRMs. Due to every card being slightly different means that you need a solution designed for a particular card or be creative with adapting a generic solution. Even more serious overclocking involves replacing the shunt resistor that measures the power draw, this can trick a card into pulling >500W but as you might think not all VRMs are designed with enough heard room to survive that. Also an exploding VRM has a good chance of frying the GPU since it is what is making its core voltage.

For actual day to day use i tend to downclock graphics cards. They run a lot quieter at about 75% of max TDP and are typically still fast enough to do the job. Also helps with the room not getting as hot in the summer.

For the CPU the easiest speed gain is adjusting the bios to allow all cores to constantly run at the max boost clock speed. That generally will never become unstable unless the motherboard can't keep up with power delivery or the cooling can't keep it cool. Each core is capable of reaching that boost speed anyway, it just takes a lot of power if all do it.
paulca:
Interesting points.  It seems things have changed somewhat in 2020+.  NVidia GPUs used to come with power and temp limiters set at 100%, but could be slammed up to 120% to "overclock" the card out of the box.  At 120% power limit it is able to clock high enough to overwhelm it's cooling.  So throwing more and more cooling at it gives you higher and higher clocks and more and more fps.

But today, the cards are hard limited to 100% power limit, unless you pay more for the card with upgraded VRMs et.al and a different BIOS to get 110% power limit... it's only when you pay through the nose for the top of the line variants do you get access to the full 120/125% overclock.  Only then is it likely to overload an air only cooler and could thermal throttle the boost clocks.

Which is another NVidia change from 10xx cards to 30xx cards.  They are absolute psycopaths with their own internal overclocking.  I've push it to 85*C before it even started to clock back the boost.  It carried on until it hit 95*C when the screen went black for a heart stopping second, it force slammed it's fans to 100% for a few seconds, recovered at 80*C again and carried on.

I just ran 3 3DMark bench marks.  With full fans.  With minimal fans and with no fans at all.  All of them gave scores within 1% of each other.  All of them limited by the power limiter.
Navigation
Message Index
Next page
Previous page
There was an error while thanking
Thanking...

Go to full version
Powered by SMFPacks Advanced Attachments Uploader Mod