What Building-to-Building Optical Interconnect Will Enable Gigawatt Scale Training Clusters?

–

Moscone Center

The size and complexity of generative AI tasks are forcing compute clusters to grow beyond the size of a single building. GPUs performing training functions must tackle massive datasets and model sizes, requiring synchronous operation and data sharing across GPUs in the cluster. Networks to support these tasks require dedicated communication fabric between these GPUs. This panel will explore optics optimized for inter-building links. Our panel will evaluate these design constraints:

Early adoption: high capacity switches will ramp quickly and need optics to support
 
Large radix: ≤1.6T/port in 100T switch generation

Low power: can optics fit in a port optimized for 2 km optics?
   
Reach: ≥10 km

Latency: what will balance optical performance to training efficiency?
   
Fungibility: is there any alternative to pluggable optics if the ports need to mix long and short reach?
   
DWDM: Is it cheaper than parallel fibers?
  
Cost: enable network scaling
    
 

Organizers