Linux or Windows 10/11 for a Threadripper PRO

Hi @wendell I have been crunching through these forums to understand builds and finally made one myself, here is the config:

  1. Threadripper Pro - 5995WX 64 Core/128Threads
  2. D4532R DDR4-3200 64GB ECC/Reg Server Memory - 512GB
  3. 1x6000 ADA + 1x4000 ADA (4000 Only for Display purpose)
  4. ASUS PRO WS WRX80E-SAGE SE WIFI II Extended-ATX

So here comes my question. I am a ML researcher and use the machine for model training and everyday development.

  1. Can Windows 10/11 handle this config efficiently ?
  2. I am a big fan and user of POP_OS and I want this machine to rock that. Do you think POP_OS or Ubuntu is a good choice for this beefy machine.
  3. Also could you please suggest a good RDP solution for POP_OS/Ubuntu OS to access it as is with complete GUI. I am referring something similar to Microsoft RDP. But before reaching out to you I tried the below:
    1. Logme in - only gives terminal access
    2. Chrome desktop - it’s a pain, it sticks, it’s draggy
    3. Real VNC - brought a license too but it still it stucks
    4. Tried installing matte server and tried using thin client from my laptop didn’t go well.

Could you and all experts here send me some suggestions please. Thanks in advance.

I would heavily lean towards Linux, when I built my ML rig I installed Ubuntu and I have not had any issues. No OS is perfect but I think Linux is far superior for development.

As far as remote desktop I don’t know exactly what your workflow is but have you looked at remote development through ssh in your ide? I know vscode has the feature and I am pretty sure pycharm does as well.

Thanks @thecoderx for the suggestion. VSCode ssh is a nice option but my use case is that I also need an eclipse running, we have a setup around it. That’s the reason I was looking for a Remote Desktop software as well. And my preference is also Linux too.

1 Like

Hi! Hmm… If I are on the same LAN or nearby. X-windows working fine for me on openSuSE “ssh -X [email protected]” And then open up apps with GUI in my desktop. Regards-

It sounds like you are going to be doing a lot of remote development and performance is really the key sticking point. I might consider a VDI setup with an Intel A770 or an older Nvidia card for your encoding. It may seem like overkill as you probably would only be running one VM but it would give you a desktop like experience on your thin client and allow you to easily share your machine with other researchers or easily run different operating systems. I am by no means an expert on the topic but I would at least look into it more. @wendell definitely knows a ton about this.

Thanks guys for all the suggestions. @thecoderx I guess I might have to switch to windows because I have tried all the solutions that google gave me including free and licensed and found every solution has same issues. When I access the Remote desktop for first time it comes up and all other subsequent times like if I try to minimize a window the connection is stuck like this. Attaching inage for reference.

@vrc I think you are having a hardware issue that might be being caused by a bad/outdated driver. I know when I first built my rig with a rtx 4090 and Ubuntu LTS getting the Nvidia drivers installed and recognized with cuda was much more of a headache then I was anticipating. I was used to Windows where these things seem to play together a little nicer.

Hmm Really and that never came to my mind that it could be a driver issue. So here are the below steps I would like to do and please correct if anything can change or not needed:

  1. Reinstall POP_OS again latest with Nvidia drivers. The distro has 2 versions, 1 generic & 1 specifically made for Nvidia hardware.
  2. Update MOBO Bios or firmware.
  3. Then try these softwares on top of it.

Also there is 2 weird issues I haven’t mentioned, some times when the system is rebooted sometimes the login screen goes to the left (or simple terms like unscaled version). So I have to reboot again to get to normal layout. Other once the system goes to sleep, it hibernates infinitely, so even tough we try to wake it up , the it doesn’t, I have to manually force shutdown the system.

Or is going for Plain Ubuntu 22 LTS would be better ?

Not sure about this, I don’t have enough experience to speak to the difference between the two. But I will say that getting NVIDIA to play with Linux in general can be difficult. Have you manually installed the latest recommended driver for your GPU? From the other issues you are describing I would be leaning towards a driver issue even more as a potential issue.

Have you tried running:

lsmod | grep nvidia

If the output contains nvidia you will know the driver is loaded. If the driver is loaded run the following:

nvidia-smi

It will tell you what version driver you have installed. Make sure everything matches what you expect.

I would look at the versions to see what is installed vs what is available. The processor you are on is pretty mature and any bios updates at this point are “probably” minimal. Usually the major bios updates come when the processor/platform is new. I am not sure this is the lowest hanging fruit and always has some potential to cause you other issues.

@VRC have you tried xrdp? I have a similar-ish situation to you were I would really really prefer to use linux and I need a GUI on the machine, but the machine is not in the same room as me.

I can have an xrdp session open for multiple weeks without dropouts or session wipes (which use to be an occurrence years ago using xrdp).

There’s also the nuclear option of just using a good ikvm.

I have a feeling this is more about Ryzen/Threadripper and Linux not playing together nicely. I seem to remember Wendell mentioning something about the 5000wx series not liking hibernation in one of his videos, but I can’t find anything on this forum about it. But a quick search on Google will show you all kinds of issues with hibernation both on Linux and Windows with Ryzen. I disabled hibernation on Ubuntu LTS because it was giving me weird behavior on my Ryzen 3900x.

Yes Nvidia driver is loaded and the nvidia-smi works too.

1 Like

No I havent, thanks for mentioning, that will be the next try before reinstallation.

Okay looks like I will have to research more into these compatibility issue.

It sounds like maybe your video encoding isn’t working properly and that is why the remote desktop solutions won’t send the signal over the network appropriately because it is not being compressed. I think the easiest way to diagnose this is to install ffmpeg and then check if the NVIDIA nvenc encoders are available.

ffmpeg -hide_banner -encoders | grep h264_nvenc

If they aren’t available this probably points to a bigger problem with the gpu not talking to the other hardware or OS for some reason. If you haven’t tried it yet I would roll back to a one GPU install at some point soon and see what changes.

@thecoderx fugured out the issue, it is the OS itself. POPos is okay as on device OS and ssh kind of but it’s not playing well if you want to render the GUI to remote devices.

I moved to Ubuntu and it’s freaking smooth and vivid. Just using free version of VNC for now and it’s amazing. I was doing all this testing on my other PC with 5950x with 128GB ram. Now I am actually building the Threadripper beast with 64 core and 512gb ram, I am excited about it.

However there is a great disappointment wine doesn’t work really well on Ubuntu too for some softwares that are absolutely needed

1 Like

@VRC glad you figured it out! I wonder why POPos doesn’t like that setup.

Not sure how wine runs on Fedora but it seems like a lot of the guys doing ML like it too as it is a little more up to date. May be worth a try.