Description

I have not built a custom PC since 1999 (Pentium III days!). Recently I decided I would begin to learn about machine learning and neural networks, and realized that the off-the-shelf Macs I had at home are inadequate for the task.

Compiling the part list was quite a process. For the longest time, I was settled on getting the Asus X399 Strix-E Gaming motherboard, which is an E-ATX board. That constrained me in terms of case selection quite a bit, coupled with my desire to use the Enermax Liqtech 360 AIO cooler/radiator, whose thickness also presented some issues when selecting a case.

Selecting the 1080 Ti-based GPU card really came down to performance per price, and I thought the EVGA 1080 Ti SC2 had the right balance and came at a great price.

Assembling in the be quiet! Dark Base Pro 900 was...an interesting experience. The case is incredibly modular, but that also meant that moving things around meant disassembling many sub-parts. Taking out the optical drive cage, for example, to make room for the Enermax radiator and fans was quite the task. But overall, I think it worked OK - I was glad to have a full-tower case that gave me room to maneuver all the cables, and had generous cable tie-down points.

Luckily it all worked when fired up the first time. Got Debian 9.2.1 Stretch installed, and through some digging around, even found the firmware for the Asus PCE-AC88 4x4 MIMO Wi-Fi card. Drivers for everything else worked great as well, including the NVIDIA drivers for the GPUs.

Installing Windows 10 was another affair. I was trying to create a bootable USB flash drive from inside Debian, and when I booted into it, the Windows installer indicated it required some missing drivers. I turned off CSM compatibility mode, used WoeUSB to create the Windows installer USB drive inside Debian, and then the Windows installer was able to work properly.

One final problem that I had was that while the system recognized the 8 sticks of DDR4 memory I had installed, both Windows and Linux indicated that only 48GB out of the 64GB were usable - i.e. that two sticks were essentially useless.

I suspected that the motherboard BIOS was the culprit, and I was right. Flashing the latest BIOS from ASRock's website (version 2.00) resulted in both Windows and Linux recognizing the 64GB of RAM.

Now the real fun begins. Still deciding between Keras and PyTorch for a machine learning framework, but hoping to converge on a decision soon.

Please let me know if you have questions.

Comments

  • 21 months ago
  • 5 points

That’s a very powerful system. Congrats on your build. I love the fact that your last item in your expensive list is the $10 mouse...

  • 21 months ago
  • 2 points

Wow what a beast! great job!

  • 21 months ago
  • 2 points

Wow! I would like to be at the point in my life where $6k to learn something new isn't such a stretch. This is a nice build good sir.

  • 21 months ago
  • 1 point

Damn, that's super nice! Great job! You liking that Threadripper? I love mine.

  • 21 months ago
  • 1 point

Gotta spend more time with it to be honest. The real test for me will be to see how I can multi-thread neural network training tasks.

  • 21 months ago
  • 1 point

Its so good

  • 21 months ago
  • 1 point

What better the most expensive mac money can buy that offers the best performance for an Apple, or this beast you have here?

  • 21 months ago
  • 4 points

this thing blows macs out of the water, macs are really overpriced, and this is pretty close to the best you can buy consumer computer wise, soooo.........

  • 20 months ago
  • 2 points

I agree with acs590. I am living the exact same situation he is, I have the 2017 IMac with the top specs available not the new pro and it was nice but with some pain when it came to video editing and model training times in tensorflow. I built a machine almost identical to acs590’s (even in a be quiet case albeit smaller) and it’s awesome.

  • 21 months ago
  • 1 point

very nice

  • 21 months ago
  • 1 point

Great PC, you can do anything with it.

  • 21 months ago
  • 1 point

I think I am going to copy this build. Why did you stop at 64GB? You might run out of RAM and have to deal with batch coding.

  • 21 months ago
  • 1 point

I figured that it should be plenty for now, and I can upgrade to 128GB later if needed.

  • 21 months ago
  • 1 point

"used WoeUSB to create the Windows installer USB drive inside Debian"

Once PCI Passthrough on Threadripper becomes more mature, I guess you could add another lower power card like a 1060 6GB as the host OS card and run Linux off of the 1060 and the SLI 1080 Tis off of Windows at the same time.

  • 21 months ago
  • 1 point

Damn nice. I'd be curious to see what you can OC out of this bad boy.

  • 21 months ago
  • 1 point

Yeah I don't have any experience with overclocking, but got the Enermax with the hope of learning how to OC at some point.

  • 21 months ago
  • 1 point

A lot of boards today come with their own flavor of OC software. I run an ASUS Z270 board and it has some real handy features for you to OC without having to have really any experience with it. It'll gradually increase the system and if it gets to an unstable point, it'll go back down to the setting before when it was stable. It's not bad. The ASRock should have some sort of software in the BIOS that allows you to easily (or rather more easily) OC the system. With the kind of cooling that you have and the hardware you're running, I'd be surprised if you didn't hit a steady 4GHz with what you've got at least.

  • 21 months ago
  • 1 point

What's the plan for the machine learning?

  • 21 months ago
  • 1 point

Will probably start with traditional image classification and style transfer neural networks for learning purposes, then I have some aspirations for video classification for sign language interpretation!

  • 21 months ago
  • 1 point

Check out the channel Sentdex. He covers a lot of the basics. Getting sign language data sources is probably going to be rough.

  • 18 months ago
  • 1 point

Can you mount the fans on the 360 on either side of the radiator? And if so, can i replace the enermax fans with corsair rgb fans?

  • 16 months ago
  • 1 point

How are the Hitachi holding up? Loud?

  • 2 months ago
  • 1 point

Can this fit 4 gpus?

[comment deleted by staff]
  • 21 months ago
  • 1 point

Not yet. My primary intention is really to use it on ML, but I really should install some games just to see what they're like.