| Backend | Comm. Functions | Optimized for | Float32 | Float16 |
|---|---|---|---|---|
| MPI | All | CPU, GPU | Yes | No |
| GLOO | All (on CPU), broadcast & all-reduce (on GPU) | CPU | Yes | Yes |
| NCCL | broadcast, all reduce, reduce and all gather (on GPU) | GPU only | Yes | Yes |
分布式的通信有两大类:
目前机器学习中主要由两种分布式架构:
Broadcast

Scatter
