Software Engineer, Collective Communication

The role

This role sits on the Workload Networking team, which owns the communication layer that large training runs depend on. The engineer works alongside ML researchers to fold new collective operations directly into the training stack. The position is based in San Francisco with a hybrid schedule.

What you would do

• Design and build efficient collective operations in C++ and CUDA together with researchers
• Make sure the biggest training runs fully exploit each network transport in the clusters
• Run network simulations to guide the design of future supercomputer fabrics

What they asked for

• Background in low-level, performance-critical software
• Comfortable writing performance-sensitive CPU and/or GPU code
• Experience working with collective communication is a plus

Nice to have

• Prior experience writing distributed algorithms over RDMA
• Familiarity with network simulation methods