Cisco and Nvidia develop 12.8Tbps optical modules for AI data centres
Cisco and Nvidia are collaborating on an ultra-dense optical pluggable called USFP, designed to carry far more data through a single port in AI data centres. If adopted, it could reduce the number of fibre bundles needed to connect GPUs and networking equipment as clusters grow.
Cisco’s specification allows up to 32 lanes of 400 Gbps, or 12.8 Tbps per module, compared with about 1.6 Tbps for today’s most capable pluggable optics. The modules may use up to 100 watts and a two-stage insertion mechanism to make their many electrical contacts reliably. A fault could disrupt more equipment than with a conventional pluggable, though less than a failure in co-packaged optics; USFP is intended as an open standard.
- Cisco and Nvidia are developing a 12.8 Tbps optical pluggable.
- USFP could consolidate cabling for up to eight GPUs.
- High power use and reliability are key design challenges.
New here? Start with this
AI data centres contain large groups of processors called GPUs, which handle the calculations used to train and run artificial intelligence systems. Networks connect the GPUs to one another and to other equipment, and those connections need to carry large amounts of data quickly.
Cisco makes networking equipment, while Nvidia makes GPUs and related systems used in AI data centres. Optical modules convert electrical signals into light so data can travel through fibre-optic cables; higher-capacity connections could mean fewer cables as these centres expand.
The proposed USFP design is a type of removable optical module intended to carry more data through each network port. Its capacity and reliability matter because a fault in a high-capacity connection could affect multiple pieces of equipment, while an open standard could allow wider adoption across suppliers.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
USFP could make expanding AI data centres easier to connect by carrying much more data through each port and reducing the number of fibre bundles. An open standard could also give operators more choice and encourage a wider supplier ecosystem, while keeping the optics replaceable rather than integrating them directly with processors.
The case against
The higher capacity comes with practical trade-offs: modules drawing up to 100 watts may add power and cooling demands, and a fault could affect more equipment than a conventional pluggable. Operators may therefore favour established, lower-capacity links until the new design’s reliability, maintenance needs and compatibility are demonstrated at scale.
Read the full article at the source →
Originally published by The Register as “Cisco, Nvidia team up to build fatter, faster pluggable”.