A data center architect is designing infrastructure for training very large language models across many GPUs and needs a high-speed interconnect that enables direct multi-GPU communication within the scale-up domain. Which NVIDIA technology is specifically described as providing this capability?