At OFC2024, IPEC held the Global Optoelectronic Technology Summit. Jim Theodoras, a technical expert from HGGenuine, was invited to attend the IPEC Forum and deliver a speech.

The theme of the forum is "Low-Latency High-Speed Optical Interconnection Technologies for AI Compute Era". For AI deep learning models, massive high-performance computing capabilities require transmission networks, which are the key to computing.

At the IPEC forum, participants discussed the development trends and technological innovation directions of linear drive and CPO, AI basic models, new computing power, and next-generation low-latency high-speed optical connection infrastructure.

Currently, AI clusters have 3 network inside them. 

1. Ethernet for large data transfers.

2. Infiniband for connecting the CPUs and low latency.

3. GPU network like NVlink that guarantees coherency between all the GPUs. Optical links have penetrated all but the GPU network. Nvidia’s latest Blackwell architecture used electrical interconnect for GPU interconnect, 5000 cables with over 2 miles of copper. Nvidia chose copper cables for power savings.

   
   

In this presentation, Theodoras presented the argument that optical technology that exists today can increase the coherence size of the cluster for a single rack to an entire row, thus offsetting the additional energy required for optical communications.