I wanted to set up an ML workstation and dive deeper into CNN and Transformers. My question is, is it necessary to run them via NVLink or are the PCIe Lanes fast enough? I found multiple different sources on that and im just confused. Configuration so far: https://pcpartpicker.com/list/kLmfrv
From my understanding, it’s not possible to directly combine their VRAM, thus training larger models requires data/model parallelism. Therefore using NVLink provides higher bandwidth between both GPUs, right? I’m uncertain whether I should simply opt for 1 x4090 or 2x3090’s.
I would genuinely appreciate any recommendations and helpful insights you could provide.