In a world where 128GB RAM costs 3 million won these days, I purchased Gigabyte's Spark AI TOP ATOM.
Since it won't be my main system, I bought 1TB.
What I mean is, I already have a DGX Spark.
I finished the configuration by connecting it with NVIDIA Cx7, but I realized there was something I had misunderstood.
I thought that if I connected these two, they would work together like multiple GPUs on a single server, but that wasn't the case. I learned that you need to treat the two devices as separate nodes and communicate based on Cx7, then run models through distributed frameworks like vLLM or Ray.
I thought if I ran a model on one side, the devices would automatically be recognized as multiple units... but the method was a bit different from what I had imagined... ㅠ
Although the structure turned out a bit different from my original plan, it's still usable enough, so I'm satisfied with this.
