I got my first DGX (a GN100) and yesterday I got my second one (Spark). I’ve been non stop trying to configure a cluster to run together. they work separately but i need them to work together. its been 8 hours straight and i’ve typed export VLLM_HOST_IP=10.10.10.1 export NCCL_SOCKET_IFNAME=enp1s0f1np1 export GLOO_SOCKET_IFNAME=enp1s0f1np1 maybe 20x already.
I know I’ll get it done but in the meantime while I wait, I wrapped my boxes! Too many close calls of putting soda and coffee on top of it so wanted to make sure they dont scratch. I’m gonna get 2 more but wanted to do 2 first before 4.
Glad this forum exists, the moment this works, I"m going to read as many posts here as I can.
Agreed! I pointed it to the docker after doing the playbooks. They’re (the playbooks) are actually great to learn the basics, use Claude GPT or Gemini if you get stuck.
Once you get the two nodes solidly clustered (and do update both nodes’ firmware etc!), point the model/agent to the community docker, and read it as well to understand the basic configuration.
Once you get the docker configured, there are threads here that will help you with recipes. At that point I hope you spend some time playing with the capabilities
Great to see another Acer GN100 owner here! I think they’re a great value so far, compared to the other GB10s. And like others have said, to setup a cluster you can cheat and use Claude Code to do it for you. I pointed Claude to forum posts here and said read all the well researched and tested findings and tips to get to the best docker configuration.
Thanks gang! I got it working (with the help of Gemini Cricket of course) and that was fun but I’m still somewhat skeptical about the performance but I"m gonna play around more with it tomorrow. Here’s a video I made on the parallelism but for now I’m really not all that impressed. I mean I can do this on a macbok pro with no fuss and the model I was playing with was only 35B tomorrow I’ll play with 70. I essentially was listen to gemini copy and paste what it told me to do then paste the response then it tells me to do something etc. I know I can give AI (like claude and openAI) permission to do autonomous things but I’m in learning mode so I appreciate the copying understanding it then pasting then figuring out what and why.
I love my GN100. I got for a great price $3710 brand new and the spark I got for $3999 and with the cable added in i’m looking at launch prices with the pair so I’m ok with it. I drove out to microcenter bought one, went back the next day to buy 2 more and they were all gone. I’m figuring maybe in the fall they’ll come down in price more. I dont really need the 4TB, I’d be ok with another DGX with just 1TB of storage.
Oh and the wrap, its only on the top and sides and its the one that goes on cars (its a gold carbon fiber weave) and it can handle the sun if If these DGX starts to get hot enough to hurt the wrap I’ve got bigger problems.
