Broadcom Tomahawk 5

Learning Objectives

• Understand why high-throughput Ethernet switching is critical to AI training clusters
• Identify Tomahawk 5's specs and how it compares to NVIDIA InfiniBand alternatives
• Evaluate when to choose Ethernet-based AI fabrics vs. InfiniBand

What Is Broadcom Tomahawk 5?

Tomahawk 5 is the fifth-generation flagship Ethernet switch ASIC from Broadcom (BCM78900) — the silicon at the heart of the Ethernet switches connecting GPUs to GPUs in AI training clusters. It delivers 51.2 Tbps of total switching throughput on a single chip, doubling the prior generation, and is the first Broadcom switch ASIC to support 800 Gigabit Ethernet (800GbE).

For AI training, network throughput is often the bottleneck — large model training requires constant gradient and weight synchronization across hundreds or thousands of GPUs. Tomahawk 5's 51.2 Tbps throughput at low (~1 microsecond) latency, with shared-buffer architecture optimized for RoCEv2 (RDMA over Converged Ethernet) and other modern AI protocols, makes Ethernet a credible alternative to NVIDIA InfiniBand for the back-end AI fabric.

Pricing & Access

Tomahawk 5 is silicon sold to switch OEMs, not directly to data-center operators. End customers buy finished switches that incorporate the silicon.

A typical large AI training cluster spends as much on networking as on the GPUs themselves — Tomahawk 5 economics directly impact total cost of training.

Core Capabilities

51.2 Tbps Throughput

Industry-leading single-chip throughput. Configurable as 64x800GbE, 128x400GbE, or 256x200GbE — flexible deployment across leaf, spine, or aggregation roles. Built on TSMC 5nm process, with power consumption only ~10% higher than the 7nm Tomahawk 4 (around 500W per chip) despite doubling throughput.

Sub-Microsecond Latency

Approximately 1 microsecond cut-through latency (64-byte packets) — competitive with InfiniBand for AI workloads. Critical for AllReduce, AllGather, and other collective operations that dominate large-model training.

RoCEv2 Optimization

Industry's most advanced shared-buffer architecture, providing the lowest tail latency for RDMA over Converged Ethernet (RoCEv2). RDMA lets GPUs access remote GPU memory without involving the CPU — the foundation of efficient distributed AI training. Tomahawk 5's buffer architecture means RDMA performance stays consistent under congestion, where lesser silicon stalls.

800GbE Support

First Broadcom switch ASIC to support 800 Gigabit Ethernet ports. 800GbE is the new standard for AI back-end fabrics — most NVIDIA H100/H200/B200 cluster designs assume 800GbE per GPU NIC. Tomahawk 5 deployments scale from 64-GPU pods to 10,000+-GPU clusters.

Telemetry and Congestion Control

Hardware-level telemetry feeds congestion-control algorithms. Visibility into queue depths, packet timing, and ECN (Explicit Congestion Notification) signals lets operators tune AI fabrics for sustained throughput rather than peak-only performance.

Strengths

• 51.2 Tbps single-chip throughput: Industry-leading — doubles prior generation
• Mature ecosystem: Multiple switch OEMs (Arista, FS, NADDOD, Cisco) ship products built on Tomahawk 5 — buyers have vendor choice
• 800GbE first-mover: Enables next-generation AI cluster deployments without waiting for the next switch silicon generation
• RoCEv2-optimized: Best-in-class shared-buffer architecture for RDMA workloads
• Power efficiency: Only ~10 percent power increase over Tomahawk 4 despite doubled throughput
• Ethernet ecosystem: Open standards, multi-vendor optics + transceivers, broad interoperability

Limitations & Considerations

• Switch silicon, not solutions: Buyers need switch OEMs (Arista, FS, NADDOD) to ship finished products — Broadcom doesn't sell direct
• Optics cost: 800GbE optical modules can cost $1,000-$5,000 per port; cable + optics costs frequently exceed switch silicon costs
• NVIDIA InfiniBand comparison: For tightly coupled training workloads at the largest scale, NVIDIA Quantum InfiniBand still offers some workload-specific advantages — Ethernet is closing the gap, not yet universally winning
• Software stack maturity: RoCEv2 software stacks have matured rapidly but still require expert tuning for production AI deployments
• Tomahawk Ultra coming: Late 2026 successor will offer higher throughput and tighter AI/HPC scale-up integration — long-deployment buyers should track the roadmap

Best Use Cases

When to choose alternatives:

• Tightly coupled HPC + AI training at the very largest scale → NVIDIA Quantum InfiniBand still offers workload-specific benefits
• Smaller deployments (under 100 GPUs) → lower-throughput Tomahawk 4 / 3 silicon may be cost-effective
• Specialized AI fabrics → some hyperscalers (Google TPU pods, Cerebras wafer-scale) use proprietary interconnects rather than commodity Ethernet

Key Takeaways

• Broadcom Tomahawk 5 (BCM78900) is the dominant 51.2 Tbps Ethernet switch silicon for AI data center back-end fabrics
• Supports 800GbE — the new standard for AI cluster networking — across configurable port modes (64x800GbE, 128x400GbE, 256x200GbE)
• RoCEv2-optimized shared-buffer architecture delivers RDMA performance approaching NVIDIA InfiniBand at meaningfully lower cost
• Sold as silicon to switch OEMs (Arista, FS, NADDOD, others); end customers buy finished switches typically priced $80,000-$200,000 each
• Tomahawk Ultra successor in late 2026 will offer higher throughput; long-deployment buyers should track the roadmap