A data center team is designing an AI training cluster with servers containing 8 GPUs each. Within a single server, GPUs must exchange gradients at maximum bandwidth with the lowest possible latency, bypassing CPU and PCIe bottlenecks. Across servers, GPUs connect through a lossless Ethernet fabric using RoCEv2. Which technology provides the intra-server (scale-up) high-speed GPU-to-GPU interconnect that the design should specify?