Full Water Cooling Round Up 11/11

Introduction

This is a round up of all my recent testing on one page.  It includes the Corsair H100, 1-2xPMP-450s's at 12 and 24V, 1-2xRX360s and XSPC Rasa vs Raystorm Performance.  These are then split out into individual reviews with more analysis:
- The Corsair H100 review is here
- The 1-2xPMP-450s's is in construction
- The Rasa vs Raystorm review is here

Test setup:

All tests were completed on an i7-990x running at 4.4GHz (23multx193bclk) with 1.412Vcore and 1.35Vqpi, memory was run at 1600-9-9-9-27-1N.  Hyperthreading on, LLC enabled, all throttles, speedstep etc were turned off.  Loading was tested using stresscpu v2 for linux 64 bit under redhat 5.7 kernel 2.6.18-274.  Stress cpu is based on the gromacs core (from folding at home) and is fairly stable vs time wrt to temps.  Temps were logged on every cpu core every minute using kmod-coretemp and lm-sensors lib.  The loop was given 10 mins of time to stabilize before logging for 10 mins.  The recorded temperature was an average of all 6 core temperatures for 10 minutes i.e. 60 readings.  All thermal interface material (TIM) was shin-etsu x23-7723D unless otherwise stated.  This was selected as it is very insensitive to poor application and still has decent performance.  To be really fair I should have remounted each cpu block 5 times and taken the best result, shin-etsu seemed the best option given that I didn't have time to do this.

Load Temps


Approximate Overclocks

Some of you are more interested in maximum overclock than necessarily reducing temps.  By making some estimates we can therefore come up with an approximate change in overclock by doing some math.

Some assumptions:
- CPU power = 200W
- voltage required for overclocking is linear with frequency (which it isn't) and was based on 2 points - 1.4V for 4.4GHz and 1.5V for 4.6GHz
- That the cpu power load is entirely dynamic switching power that follows the standard cmos power equation 1/2 * cap * freq * volts^2
- Max "safe" temperature is 50C above ambient

Using this gave me a formula that from experience looked to be off by a factor of two, applying a correction factor gives us the following plot:

Obviously this all assumes that any require vcore is safe and that there are no voltage "walls".


Cost Efficiency

As expected the more you spend the less proportinally you get:

Full Data