Networking & Interconnect
How chips and sites talk to each other.
Fourth-generation NVLink bandwidth per H100 GPU — what one accelerator can say to its neighbours

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).
| Type | Source | Published |
|---|---|---|
| PAPER | Deep Neural Network Inference on an Integrated, Reconfigurable Photonic Tensor Processor Photonic tensor processor team · imec / silicon photonics ecosystem First production-ready photonic tensor processor for DNN inference — packaged in a 19-inch rack unit, PyTorch-integrated, MNIST + CIFAR-10 benchmarked. Runs pretrained networks without chip-specific retraining | 2026-01-15 |
| PAPER | Integrated Photonic Neural Network with On-Chip Backpropagation Training Farshid Ashtiani, Mohamad Hossein Idjadi, Kwangwoong Kim · Nokia Bell Labs First integrated photonic deep neural network with end-to-end on-chip gradient-descent backpropagation — all linear AND nonlinear computations on a single silicon photonic chip. 92.5% accuracy on 2D classification; automatically compensates for fabrication variations | 2026-03-20 |
| ANALYSIS | Broadcom Announces Tomahawk 6 - Davisson, the Industry's First 102.4-Tbps Ethernet Switch with Co-Packaged Optics Broadcom Inc. (press release) · Broadcom Per-bit optical power claim on a named part (BCM78919): 102.4 Tb/s optically enabled from 16 x 6.4 Tb/s engines at 200 Gb/s per link, 70% lower optical interconnect power — 'more than 3.5x lower than traditional pluggable solutions'. 512 XPUs scale-up, 100,000+ in two tiers. SAMPLING at announcement, not volume; 70% and 3.5x are the same claim stated twice. | 2025-10-08 |
| ANALYSIS | NVIDIA Announces Spectrum-X Photonics, Co-Packaged Optics Networking Switches to Scale AI Factories to Millions of GPUs NVIDIA (press release, GTC 2025) · NVIDIA Second vendor CPO power claim on comparable framing: 3.5x power efficiency, 4x fewer lasers, 63x signal integrity, 10x resiliency, 1.3x faster deployment vs pluggables; 100 Tb/s and 400 Tb/s Ethernet configs, 144x800G InfiniBand. ROADMAP, not deployment — dated 2025-03-18, Ethernet stated as coming in 2026. Lumentum named in NVIDIA's own CPO partner list (bears on LITE). | 2025-03-18 |