Inter-node vs Intra-node

I was reading the Multi node training section, > So here we are continuing with the idea of 2B param model and we will now use 32 GPUs across 4 nodes to speed up the training even more. > While each group of 8 GPUs is still connected with super-fast NVLink technology, the inter-node connections are usually in an order of magnitude slower. Is NVLink used for inter-node connection as well? From my preliminary research on the topic, intra-node use NVLink and for inter-node either InfiniBand or RoCE is used. Here are few references that I used - https://lambdalabs-files.s3-us-west-2.amazonaws.com/lambdalabs.com_presents_distributed-training-a-gentle-introduction.pdf - https://apxml.com/courses/how-to-build-a-large-language-model/chapter-18-hardware-considerations-llm-training/interconnect-technologies-nvlink-infiniband