TL;DR
- Tomahawk 5 (BCM78900) is Broadcom's fifth-generation StrataXGS switch ASIC, delivering 51.2 Tb/s switching capacity in a single chip.
- Common port configurations: 64 × 800GbE, 128 × 400GbE, or 256 × 200GbE — used by Arista, Cisco, Edgecore, and most white-box vendors.
- Shallow-buffer architecture optimised for low latency; relies on PFC/ECN tuning for lossless RoCEv2 behaviour.
- Direct competitor to NVIDIA Spectrum-4 for AI fabrics; common in hyperscaler-built networks where vendor diversity matters.
Overview#
The Tomahawk 5 is Broadcom's flagship merchant-silicon switch ASIC for data centre Ethernet. Announced in 2022 and shipping in volume from 2023-2024, it provides 51.2 Tb/s of switching capacity using 512 lanes of 100 Gb/s PAM4 SerDes. That capacity can be packaged as 64 × 800GbE, 128 × 400GbE, 256 × 200GbE, or various mixed configurations.
Tomahawk 5 powers most of the non-NVIDIA 800G Ethernet switch market. Arista's 7060X6, Cisco's Nexus 9332D-H2R, Edgecore AS9817-64D, and Dell Z9864F-ON all use the BCM78900 family. It is the default merchant-silicon choice for hyperscalers building bespoke AI fabrics.
Specifications#
| Property | Value |
|---|---|
| Family | BCM78900 (StrataXGS) |
| Aggregate capacity | 51.2 Tb/s |
| SerDes | 512 lanes × 100 Gb/s PAM4 |
| Typical port mix | 64×800G or 128×400G |
| Buffer architecture | Shallow, on-chip |
| Packet processing | Programmable via NPL |
| Telemetry | BroadView, Inband Network Telemetry |
| First shipments | 2023 |
Shallow Buffers and AI Fabrics#
Tomahawk-class silicon prioritises low latency and high throughput over deep buffering. For AI fabrics this is a deliberate trade: incast bursts must be absorbed by tight ECN/PFC control rather than by large queues. Properly tuned, this works well — a Tomahawk 5 fabric running RoCEv2 with disciplined congestion control delivers competitive AllReduce performance.
The alternative — deep-buffer silicon like Broadcom's Jericho 3-AI or DNX-family chips — trades latency for tolerance of less-disciplined traffic patterns. Many production AI fabrics combine the two: Tomahawk leaves for low intra-rack latency, Jericho-class spines or super-spines for absorbing transit bursts.
Operational Notes#
- PFC and ECN tuning is non-optional for RoCEv2 fabrics — defaults from the switch vendor are starting points only.
- Dynamic Load Balancing (DLB) replaces static ECMP and is essential to avoid flow polarisation under elephant flows.
- Telemetry: Inband Network Telemetry (INT) lets per-packet path and queue depth be exported for hotspot diagnosis.
- NOS choice (SONiC, Arista EOS, Cisco NX-OS, Dell Enterprise SONiC) varies operational tooling more than performance.
References
- Broadcom Tomahawk 5 Product Page · Broadcom
- Tomahawk 5 51.2 Tbps Switch Brief · Broadcom
- Open Compute Project — AI Ethernet Fabrics · OCP