El jueves 12 de Febrero, José Duato impartió la charla «Advanced interconnect solutions for high-performance AI accelerators», dentro del ciclo Las Charlas del Máster, con un gran éxito de afluencia.
Resumen de la Charla
The massive deployment of increasingly larger LLM models has triggered the design of advanced accelerators to provide the required computational power. Those accelerators implement HBM memory to provide the required memory bandwidth, which drastically limits memory capacity per accelerator. As a consequence, LLM models have to be spread among multiple accelerators, leading to previously unexisting communication needs. This talk introduces the different kinds of interconnects that exist in state-of-the-art datacenters and their main features. It also addresses power consumption and the need for advanced solutions to make datacenters more sustainable.
