AI scale depends on the network

Your AI investment depends on how efficiently your systems exchange data.

On October 12, Meta shared new networking designs at the Open Compute Project Global Summit:

• IcePack, its first liquid-cooled network rack.
• IceCube switches delivering 102.4 terabits per second of switching capacity, twice its previous 51.2T platforms.
• A fabric design distributing GPU traffic across eight independent network planes.
• An Open Switch Platform Architecture initiative, co-led with Microsoft and Alibaba. engineering.

Multiple network paths help keep GPUs working when a component fails, although available bandwidth falls.

The same day, Cisco and NVIDIA announced design plans for USFP, a new format for pluggable optical modules connecting AI systems.

Their stated design targets include:

• Four times the bandwidth density of today’s 1.6T pluggables.
• Approximately 1.3 times greater power efficiency.
• Field-replaceable modules using a familiar maintenance model.

The goal is moving more data within the same physical footprint while using less energy per bit.

These are announced design targets, with benefits requiring validation in deployed systems.

My takeawayinclude networking and facilities in the AI business case from the start.

When your team proposes more AI compute, do they also show how the network will support productive use?