I have had some interest from a couple of forum members with regards to an upgrade project I have been working on which involves installing a Tesla M2090 GPU card in my HP Z620. I thought it may help to provide a little bit of background to give some context to this post;
Originally purchased my HP Z620 in March 2016 and it arrived with the following spec;
E5-2660 x1, 32GB (8GB x4) PC3-12800R ECC REG RAM, Quadro K4000 and 1TB 3.5” HDD
Since then I have added/upgraded the following;
2nd CPU riser and 2x Xeon E5-2670’s, replaced the Quadro K4000 with a K4200, added a Tesla M2090 GPU card, upgraded the RAM to 64GB (8x 8GB) PC3-14900R ECC REG, added a 2.5” + 3.5” Swappable ICY Dock (connected via an additional 2-Port SATA III PCIe card), and added a 256GB SM951 as a boot drive (via an ASUS PCIe card), and filled the drive bays with a 480GB SanDisk Extreme II SSD, 4TB and 2TB HDD’s.
Basically looking to max out the machines performance but within a sensible budget. I mainly use the Z620 at home for SolidWorks modelling and music production but my son started at University last year and is studying for a degree in 3D modelling and computer animation. Consequently, I was looking at ways of improving rendering speeds but shuddered at the price of second hand high end Quadro cards.
I've been working on a mod project to add a relatively cheap Tesla M2090 GPU card to my current HP Z620 workstation and thought it may be of interest to some forum members. Currently there are several Tesla M2090 on eBay selling for around the £70 – £80 mark. The main issue with the Tesla M2090 card is that it was never designed to be installed in a standard PC enclosure since it relies purely on forced air cooling to stop the GPU from overheating and shutting down. It’s a 225W card which generates a lot of heat. My plan was to use liquid cooling for maximum cooling efficiency and hence maximum performance.
Before I get to the build details, I thought some benchmarks may be of interest to help anyone decide if the hassle outweighs the benefits. First up is a render of a sample scene in SolidWorks Visualize;
Image size 1920 x 1080 pixels, 2500 passes
https://www.dropbox.com/s/fvbxuqwnymlbape/1969Camaro%2010.jpg?dl=0
The results;
Octanebench is another great benchmarking tool for measuring GPU performance.
Benchmark scores;
From L-R - K4200 only (score=39.94), M2090 only (score=52.20), K4200 + M2090 (score=91.85)
For comparison, here’s a link to the results page; https://render.otoy.com/octanebench/results.php
So in summary, faster than a pair of Quadro K4200’s (score: 79), and the Quadro K6000 (score: 87). Checking on eBay tonight, the cheapest Quadro K6000 listed is £1250 + shipping!!!
So now the build project. This being my first ‘proper’ PC mod, (and my first hands-on encounter with liquid cooling), there was a certain degree of discovery involved during the project, (or in other words, I made it up as I went along but with a rough plan). I suppose in hindsight, there are a couple of minor tweaks I would make if I were to start again since there was a bit of confusion and miss-match regarding the tubing barb sizes. I should also caution that this mod did require some minor modification to the back of the Z620 chassis. Also, HP does not provide any support for the M2090 when installed in a standard desktop workstation and any modding is likely to invalidate your HP warranty.
First step was to remove the existing passive cooler on the Tesla M2090 and fit the custom Tesla active water cooling block and the 10mm barbs.
The next step was to mount the PWM fan temperature controller PCB board to the back of the Tesla card. Due to the short length of the thermocouple sensor, it was not possible to mount the PCB anywhere else. The controller board is secured in place using 4 layers of 3M car number plate tape attached to the back of the PCB. Another piece of tape was used to secure the thermocouple directly behind the GPU, (i.e. dead centre of old cooler mounting holes).
Now the brave part. Using a ‘Dremel’ like multi-tool, I cut a small opening in the rear of the chassis sufficiently large enough to accommodate the cooler tubing and wiring.
The 4-pin Molex connector tails and PWM extension cables were then attached to the controller board and the card fitted into the HP Z620.
I mounted the radiator, fan and pump/reservoir onto a metal plate I fabricated at work and started with the plumbing. My online search indicated the best arrangement is for the pump to feed directly into the GPU card cooler, the GPU card outlet is then connected to the bottom radiator connection, the top radiator connection then feeds back to the reservoir which sits above the pump. Keeping the reservoir located above the Tesla card, pump and radiator eliminates the risk of air pockets in the loop.
Next was connecting the Tesla cooler to the external radiator set-up.
Then finally connecting the fan controller board to a spare 4-pn Molex connector located behind the DVD drive;
To make the build look a bit less DIY, I also made a simple cover to tidy the appearance of the wiring and tubing. Below are images of the final build.
Anyway, I hope this post is of interest to some and perhaps add a little insight should you be considering adding liquid cooling.
By fitting the temperature controller board the Z620 is virtually silent when not rendering or performing GPU intensive work since the fan will only spin at 40% speed. As the M2090 starts to heat up the fan speed is controlled proportionally to the M2090 temperature.
Finally, here’s the complete parts list;
If you have any questions or are considering a similar type of project and would like some further details then please feel free to pm me.