Multi-rack GPU Clustering
Delivering high data throughput for larger AI clusters with multi-rack PCIe connectivity
Requirements and Challenges
- Larger clusters of GPUs with high performance connectivity are required to address the growing complexity and size of AI models
- AI infrastructure must scale GPU clusters across several racks as each rack can only accommodate a limited number of GPUs due to space, power and thermal management constraints
- Passive PCIe DACs can only support up to ~3 meters for PCIe 5.x, limiting the number of GPUs in a cluster that can be connected in a single rack
- Clustering with passive PCIe DACs and general-purpose AECs lack advanced cable and fleet management features that are essential for managing data center infrastructure
Aries Smart Cable Module (SCM) Benefits
- Purpose-built for demanding GPU clusters and delivers robust signal integrity and link stability over long distances
- Improves cable routing and rack air flow with low power active electrical cables (AEC) using thin copper cables to scale GPU clusters within and across racks
- Supports reliable PCIe 5.x reach extension up to 7 meters to enable larger GPU clusters in a multi-rack architecture
- Real-time lane, link and device health monitoring and management features with Astera Labs’ COSMOS suite that provides comprehensive diagnostics and advanced Link, Fleet and RAS management for cloud-scale deployments