An architect is designing an infrastructure solution for an AI workload that involves processing massive datasets for training. The goal is to minimize data transfer latency between the storage system and the GPUs in the compute nodes. The architect wants to enable the NVIDIA GPUs to fetch data directly from the NVMe storage array, bypassing the server's CPU and main memory.
Review the proposed architectural components:
```
- Compute Nodes: HPE ProLiant DL380a Gen11 with NVIDIA H100 GPUs
- Storage: HPE GreenLake for File Storage
- Interconnect: Ethernet with RoCE support
```
Which technology must be enabled and properly configured across these components to achieve this direct GPU-to-storage data path? (Choose 2.)