Tuesday, August 20, 2019

Custom eGPU enclosure for GTX1060 + HP Microserver Gen8

The goal

Use GTX 1060 6G GPU with HP Microserver Gen8 for deep learning.

Usability

Mostly for deep learning, but also mining or accelerated password checking is possible.

Parts

  • GTX 1060 6G, blower style from ASUS
  • Mean Well RSP-200-12 Industrial 12V, 200W power supply unit
  • PCI-E Riser card over USB 3.0 cable
  • Additional good quality USB 3.0 cable, 1.8m
  • 10A panel meter (could be also 15A)
  • Aluminium "L" profile 3x5cm
  • Power cables, 0.75mm (minimum)
  • 2 x PCI-E 6 pin power cables with 12V as 2 x 0.5mm (minimum)

Video



Description


Recently I was playing with deep learning with Keras, but I've stuck with computing power of my CPU. I don't have any other PC than HP Microserver Gen8 and really don't have place for buying something bigger. I've really wanted to dive into ML so ... I bought GTX1060 and started to figure out how to connect it to my little server.

Despite that card has PCIE 3.0 bus, and HP gen8 have PCIE 2.0, minus the fact of using only one PCI lane, it works really great for my purpose. The downside of my setup is a bottleneck on transferring data between PC <-> GPU, but since computing power of my GPU is the same it doesn't make a difference for me at this point.

If deep learning would be my next hobby, I could always buy this additional tower PC and move GPU there, but since I'm starting with the topic, I think my setup is sufficient.

To power up the GPU, I've bought Mean Well industrial PSU, which in my opinion is quite ok. I've choose it over normal PC PSU, because it's smaller and I need only 12V output.


The goal was to fit under wall mounted monitor.



Original USB 3.0 cable used for PCI-E lane no.1 was a little bit short  ~ 50-60cm, so I've bought good quality USB 3.0 cable and soldered it into riser pcb part that needs to be placed into PCI-E port on the motherboard.

New cable is 1.8m long and there are no problems with GPU communication at this point.


Additionally I've made this more secure by connecting dummy slot with riser PCB.


Now it fits perfectly and I'm not afraid that it accidentally fall of the slot.


Yep, the hole is not in the center, but I've add additional rubber tube to hold the cable better.


Before start to build my enclosure, I've tested if the card is working without any problems with riser and 1.8m cable.


I wasn't really sure if GTX will survive this kind of test, I would be really pissed of if I blew up brand new card, but at the end everything worked properly.


I've hooked it up into my custom "lab" psu, in idle GTX is consuming about 0.7A, it's really not that much, considering that it can eat > 10A at full speed.


The power line is under the riser card, it's glued to aluminium profile and additionally it had some capton tape on the top.


Riser card sits on 4 screws. I've added also additional plexiglass shield to separate GPU from power cables.




10A panel meter will tell me how much power is drained by the card. 10A on 12V = 120W, so same as GTX1060 has in it's specs, but from my tests looks like it can jump over 10A. On related movie on YouTube, I'm testing deep learning, gpu_burn and john the ripper power drain, check it out!


Power supply unit is a Mean Well RSP-200-12 with 89% efficiency, power factor correction (PF > 0.95), short circuit protection, overload, over voltage and over temperature.


It's connected to GPU through 10Amp panel meter. I'd plan to put here additional fuse, PSU gives > 16A.


It really looks nice!


It can also be used in vertical position, I think it should be better for the cooling system, since hot air likes to go up.


It's pretty stable.



And here you can see the measured 12V on the oscilloscope. For me it looks ok, Vpp is about 0.5V, it was captured during gpu-burn stress test, on which 10A panel meter was acting crazy - jumping from 6A to above max range.

Average Vcc is really nice - 12V, it looks like RSP-200-12 PSU was a good choice.

2 comments: