Elastic Fabric Adapter
Ultra-fast network card for HPC
A network interface for EC2 that lets HPC and ML training applications communicate between nodes with low latency using OS-bypass (via libfabric, MPI, NCCL), skipping the operating system kernel.
- Cost$ Low
- Difficulty●●●●●
- Costs money when idle?No
- Self-managed
- In examsSAA
Think of it asLike a private lane with no red lights connecting factories directly to each other.
When to use it
✓ Use it when
Multi-node workloads that exchange data constantly: MPI, distributed training of large models.
✗ Avoid it when
Ordinary web applications or microservices → EC2's default ENA is enough.
Common uses
- CFD and weather simulation with MPI
- Distributed ML model training on GPUs
- HPC clusters using ParallelCluster
What does it cost?
- Pricing model: No additional charge; available only on certain instance types, paid at EC2 prices.
