Cisco CAI-INF2-MI210 | 2-Way Infinity Fabric Bridge for MI210 GPU, 300 GB/s P2P

Q: What is the purpose of the CAI-INF2-MI210 bridge?
A: The CAI-INF2-MI210 enables high-speed, low-latency peer-to-peer communication between two AMD Instinct MI210 GPUs using Infinity Fabric Link technology. This direct GPU interconnect delivers up to 300 GB/s bandwidth per link, significantly improving performance for multi-GPU HPC simulations, AI training, and large-scale machine learning workloads.

Q: Which GPUs are compatible with this bridge?
A: This bridge is designed exclusively for AMD Instinct MI210 PCIe Gen4 accelerators (Cisco part number CAI-GPU-MI210). It cannot be used with other GPU models. For 4-way MI210 configurations, the CAI-INF4-MI210 bridge is required instead.

Q: How many GPUs can be connected with this bridge?
A: The CAI-INF2-MI210 connects exactly two MI210 GPUs in a 2-way configuration. Each MI210 has three Infinity Fabric links available. For larger deployments requiring 4-way or 8-way GPU configurations, multiple bridges or the 4-way variant (CAI-INF4-MI210) are needed.

Q: What performance advantage does Infinity Fabric provide?
A: Infinity Fabric Link offers up to 300 GB/s of peer-to-peer bandwidth between connected MI210 GPUs, far exceeding standard PCIe Gen4 x16 bandwidth of 32 GB/s. This enables efficient data sharing and synchronization across GPUs, reducing communication bottlenecks in distributed training and multi-GPU simulations.

Q: Can this bridge be installed in any server?
A: The bridge requires a server platform with adjacent PCIe Gen4 slots that support Infinity Fabric Link connectivity. The Cisco UCS C845A M8 AI server is a validated platform. Installation requires the GPUs to be mounted in neighboring slots with proper mechanical clearance for the bridge connector.