Summary
- DS6005 uses Broadcom Tomahawk 6 silicon and a stated 102.4Tbps switching capacity.
- The 1.5OU Open Rack V3 design integrates co-packaged optics and liquid cooling.
- Celestica has not provided independently measured energy savings or a named production installation.
Celestica has unveiled the DS6005, a fully liquid-cooled data centre switch using co-packaged optics and Broadcom Tomahawk 6 silicon. The company specifies up to 102.4Tbps of switching capacity in a 1.5OU device intended for Open Compute Project Open Rack V3 environments. It proposes 64 ports operating at 1.6Tb Ethernet and places optical engines close to the switching application-specific integrated circuit, rather than relying exclusively on conventional front-panel pluggable modules.
Electrical links between a switch chip and removable optical transceivers consume power and become more difficult to maintain as bandwidth rises. Moving optical conversion closer to the switching silicon can reduce some of those high-speed electrical paths. The potential gains include signal integrity and lower energy consumed per transferred bit, but the effect depends on the complete system, including the laser supply, cooling equipment, cabling and operating conditions.
The design uses liquid cooling to remove heat from the tightly packed switch. This is a significant mechanical change from many conventional network devices, which have been cooled by airflow. It also creates new service requirements: fluid connectors, leak control, coolant quality and coordinated maintenance procedures become part of network equipment operation. A liquid-cooled switch may fit well in an already liquid-cooled AI rack, but retrofitting it into an air-cooled network hall would require additional facility equipment.
Co-packaged optics also changes the maintenance model. If optical components sit close to a critical switch chip, serviceability becomes more complicated than replacing a failed front-panel module. Celestica refers to field-replaceable external laser modules and software features intended to support the optical architecture. The distinction between replaceable laser sources and integrated optical engines will be important when calculating repair time, spares inventory and the potential consequences of a component fault.
The announcement describes bandwidth capacity and product architecture rather than a completed commercial deployment. Celestica is demonstrating the switch at the Open Compute Project event, but no customer site, ship date or independently witnessed performance comparison has been established from the source. Claims that the technology can reduce energy consumption or latency are technically plausible in certain configurations; the scale of the reduction remains unproven without comparative measurements.
Higher port speeds also affect the wider data centre network. Fibre distribution, patching density, fault localisation and management tooling must be able to support the system. AI workloads can be sensitive to congestion and packet loss, so available switching capacity alone is not enough to guarantee training or inference performance. Network topology and software controls determine how well the hardware resources are used.
Thermal management for network hardware is a new maintenance concern where the rest of the room was designed around air cooling. A switch may reject less heat than nearby compute racks yet still require a dedicated fluid path and environmental controls. Mixing air- and liquid-cooled devices can be sensible, but it adds distinct maintenance and monitoring requirements. The data centre must establish whether network equipment can be isolated or replaced without draining shared loops or interrupting other critical assets.
Energy efficiency statements should disclose what is included. Power used by an optical engine, external laser, switch chip and facility pumps may fall into different accounting boundaries. A comparison based only on optics may overstate the benefit if additional cooling or power conversion is required, while a complete system measurement gives operators a more realistic picture. The company's press release presents an architecture and a rationale, not a independently audited facility power study.
Beyond the main switch, operators will also be evaluating fibre termination and interconnection. Large AI clusters can involve thousands of cables and require disciplined records of which paths connect particular racks. Increasing per-port bandwidth can reduce some physical port counts, but it also raises the effect of losing one link. Topology redundancy, spare equipment availability and validated management software will determine whether very dense switching improves operational resilience.
The next evidence will come from product availability, service procedures and tests reporting energy use under comparable traffic patterns. For facilities already planning high-density racks, the DS6005 illustrates how liquid cooling is extending from processors into the networking fabric. Its effect on operating costs and resilience remains a project-level question.
The design will also require a clear division between thermal management of the switch ASIC and the optical interfaces that carry traffic out of the chassis. A nominal port density provides one engineering reference point, but acceptance of a production design requires sustained traffic testing, cooling performance at specified inlet conditions, and demonstrated maintenance procedures for failed components. Those operational details have not been established by the launch specifications alone.
