Packet Buffer Reference

Broadcom Tomahawk 3 Architecture

The Broadcom Tomahawk 3 (BCM56980) is the third generation of Broadcom's hyperscale-optimized Tomahawk data center switching ASICs. Focused on bringing 400GbE to volume production, it maintains the family's hallmark shallow-buffered, low-latency, and high-throughput design philosophy.

Overview

Tomahawk 3 doubled the capacity of the previous generation, achieving a total switching capacity of 12.8 Tbps. This allowed for the industry's early mass transitions to dense 400G routing and switching.

Field Value
ASIC BCM56980 (Tomahawk 3)
Capacity 12.8 Tbps full-duplex per NPU
On-Chip Buffer (OCB) 64 MB
Off-Chip Buffer None
Buffer Model Shared Memory (SmartBuffer)
Process Node 16 nm
Interface Support 400G, 200G, 100G, 50G, 25G, 10G

Data quality note: The 64 MB on-chip buffer is an architectural constant for the baseline BCM56980. Effective buffer usability per port or per flow heavily depends on the vendor's Network Operating System (NOS) thresholds and shared pool configurations.

Buffer Architecture

Shared Pool Scaling for 400G

Tomahawk 3 continues the monolithic on-chip buffer model without any external HBM capabilities.

  1. On-Chip Buffer (OCB): Features a 64 MB completely integrated shared-buffer. This bump from the 42 MB in Tomahawk 2 was necessary to handle the vastly increased pipeline bandwidth (256 individual 56G-PAM4 SerDes lanes) and the burst dynamics of 400GbE ports.
  2. RoCEv2 and RDMA Focus: Despite being "shallow buffered," the Tomahawk 3's memory management was extensively tuned to support lossless Ethernet use cases (like RDMA over Converged Ethernet or RoCEv2) for machine learning clusters and NVMe-oF storage fabrics, relying aggressively on Priority Flow Control (PFC) and Explicit Congestion Notification (ECN).

Trade-offs

By capping the buffer at 64 MB, Tomahawk 3 maximizes thermal efficiency and density (delivering 12.8 Tbps in a single 1U form factor) but completely sacrifices the ability to absorb profound network bursts. Compared to a deep-buffer router like Jericho 2 (which has 8 GB of HBM and a similar release timeframe), Tomahawk 3 is strictly an aggregation, leaf, or spine switch that mandates sophisticated host networking to prevent traffic drops.

Comparison to Previous Generation

Feature Tomahawk 2 (BCM56970) Tomahawk 3 (BCM56980)
Capacity 6.4 Tbps FD 12.8 Tbps FD
OCB 42 MB 64 MB
SerDes 25G NRZ 50G PAM4
High-Density Ports 64 x 100G 32 x 400G (or 128 x 100G)
Process 16 nm 16 nm

Known Deployments

Tomahawk 3 was the anchor for the industry's first wave of 400G fixed-configuration switches:

See Also