DARPA Launches FastNIC Program to Fix Supercomputer Data Bottlenecks
DARPA aims to increase network interface speeds by a factor of 100 to keep up with modern processors and memory. The new FastNIC program targets the critical bottleneck that currently limits data sharing in supercomputers and data centers.
DARPA identifies the network interface card as the primary bottleneck in modern supercomputing systems. While processors and memory operate at impressive speeds of around 10^14 bits per second, and networking hardware like switches and fiber match this capability, the network interface typically functions only in the gigabit range.
This massive speed gap severely limits a processor's ability to ingest data and share information between computing units. In data centers and supercomputers made up of thousands of servers and GPUs, this bottleneck causes significant delays as incredibly fast components sit idle waiting for the slower network interface to catch up.
Through the FastNIC program, DARPA aims to reinvent the network stack and improve throughput by a factor of 100. Solving this challenge gives U.S. supercomputers a major competitive edge over international rivals, particularly China, in the ongoing race for high-performance computing dominance.