The trend towards denser, high-power datacenters is largely driven by the increasing size of AI models, highlighting a crucial need for efficient cooling and interconnect solutions.
Researchers note that AI models now grow in parameter count by a staggering 32-fold every three years, necessitating innovations both in cooling methods and interconnect bandwidth.
Nvidia's Jensen Huang observed that using copper connections for 72 GPUs requires immense power, and while optics could theoretically improve distance, they currently lack efficiency.
Ayar Labs emphasizes the potential of integrating optical interconnects within compute architectures to combat bandwidth bottlenecks and ultimately lower the operational density of datacenters.
Collection
[
|
...
]