Skip to content
Sign in
Rung 09 Networking & InterconnectSwitch rung
Rung 09 / How chips and sites talk

Networking & Interconnect

How chips and sites talk to each other.

900 GB/s

Fourth-generation NVLink bandwidth per H100 GPU — what one accelerator can say to its neighbours

NVIDIA H100 Tensor Core GPU datasheet · 2023-03-01
5
Sources
1
Concepts
1
Entities
A brushed steel fibre coupler with copper collars, on paper.

In scope: optical interconnect and co-packaged optics, scale-up versus scale-out fabrics, NVLink, InfiniBand and Ethernet economics, bandwidth per dollar, topology, cross-datacenter links. Out: memory bandwidth inside a package (hardware), photonic compute (out of the knowledge base entirely).

A cable bundle of copper conductor and glass fibre terminating in one cartridge.
Optical interconnectCPOFabricsTopology
Download Report
Report only1Show all
Sources compiled for this topic
TypeSourcePublished
REPORTScaling AI Factories with Co-Packaged Optics for Better Power Efficiency
Ashkan Seyedi · NVIDIA (Technical Blog)

Vendor-stated absolute figures the GTC release lacked: pluggables 'often 30W per interface' vs CPO 'as low as 9W' (about 3.3x, not the 3.5x claimed); electrical loss up to 22 dB (200 Gb/s channels) vs ~4 dB, an 18 dB gap = ~63x, which is where the 63x/64x signal-integrity figure comes from. Availability restated: Quantum-X early 2026, Spectrum-X 2H 2026.

2025-08-18