Broadcom PEX9749 PCIe 4.0 x16 Switch Chip – Enterprise Switch Card

The Broadcom PEX9749 is a PCIe 4.0 x16 switch chip delivering 48 lanes of switching capacity designed for enterprise AI and HPC infrastructure builds requiring GPU interconnect and NVMe-oF disaggregated storage topologies. The silicon supports non-blocking, low-latency packet switching across multiple high-speed PCIe 4.0 connections, enabling efficient pooling of compute and storage resources in hyperscale data center environments. With full PCIe 4.0 compliance and optimized latency profiles, the PEX9749 addresses the critical bottleneck of inter-device communication in GPU-centric and storage-intensive workloads.

The chip architecture integrates advanced switching fabric with support for hot-plug operations, dynamic reconfiguration, and transparent PCIe bridging. It accommodates complex multi-GPU training clusters and NVMe-oF fabric deployments where aggregate bandwidth and deterministic performance are non-negotiable. The PEX9749 launched in Q1 2024, targeting the rapid expansion of AI infrastructure builds requiring purpose-built interconnect silicon beyond standard motherboard topologies.

Broadcom's implementation includes enhanced error detection, power-efficient switching logic, and support for virtualization pass-through models. The chip integrates thermal management features suitable for dense rack-level deployments and operates within enterprise-grade environmental specifications for mission-critical infrastructure.

Ideal for

• Multi-GPU training clusters: Interconnecting 8–16 high-end GPUs with near-native PCIe 4.0 bandwidth for large model training and inference
• NVMe-oF disaggregated storage: Building high-speed storage fabric connecting multiple NVMe-oF appliances to compute nodes with sub-microsecond latency
• Hyperscale data center GPU pools: Enabling flexible GPU resource sharing across multiple compute nodes within a single PCIe switching domain
• AI inference acceleration pods: Aggregating multiple inference-optimized GPUs and storage for low-latency serving architectures
• High-performance computing (HPC) interconnect: Supporting tightly-coupled compute and memory hierarchies in scientific simulation and modeling workloads
• Real-time analytics platforms: Connecting FPGA accelerators, GPUs, and fast NVMe storage for sub-second data processing pipelines

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.