Broadcom Jericho3-AI BCM88850 Routing ASIC

The Broadcom Jericho3-AI BCM88850 is a next-generation routing ASIC architected from the ground up to address the unique traffic patterns and congestion challenges inherent in large-scale AI and machine learning training clusters. Unlike general-purpose routing silicon, the BCM88850 incorporates deep-buffer capabilities, hardware-accelerated ECMP with adaptive load balancing, and telemetry primitives optimized for the all-to-all, collective-communication workloads characteristic of GPU-to-GPU data movement across InfiniBand and Ethernet fabrics. The device achieves 25.6 Tbps of aggregate forwarding capacity while supporting high-radix topologies such as fat-tree, dragonfly, and rail-optimized network designs commonly deployed in hyperscale AI infrastructure.

Ideal for

• Spine-layer routing in large-scale GPU cluster networks (1,024+ GPUs) running distributed deep learning training workloads requiring lossless, high-bandwidth collective operations
• AI fabric leaf-tier aggregation connecting NVLink or PCIe GPU servers via high-radix 400GbE or 800GbE uplinks within rail-optimized topologies
• Hyperscale data center core routing for LLM fine-tuning and inference clusters where consistent sub-microsecond latency variance is critical to model convergence speed
• Multi-tenant AI cloud infrastructure requiring per-flow telemetry, adaptive load balancing, and hardware-enforced QoS to isolate training jobs across shared fabric
• HPC center fabric modernization replacing legacy InfiniBand-only designs with high-performance Ethernet backbones capable of supporting both AI workloads and conventional HPC MPI traffic
• Next-generation line-card and fixed-switch platform design for networking vendors building AI-optimized switching systems targeting cloud and enterprise AI infrastructure buyers

Technical specifications

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