NCCL Support in EFA ( Nitro V3)

Question

NCCL Operations uses RDMA WRITE/WRITE IMM for few of the collectives ( all_reduce, all2all) but looking the support page
Nitro V3 dont support RDMA RDMA Write , that means NCCL dont run on Nitro V3 ?

MikeLim · Answer

As per [Get started with EFA and NCCL for ML workloads on Amazon EC2](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/efa-start-nccl.html), EFA and NCCL is supported on 
- p4d.24xlarge, p5.48xlarge and newer P series instance types
- Amazon Linux 2023, Amazon Linux 2, Ubuntu 24.04, and Ubuntu 22.04 base AMIs 
- NCCL 2.4.2 and later

As per [documentation](https://docs.aws.amazon.com/ec2/latest/instancetypes/ec2-nitro-instances.html#nitro-version-network-features), Nitro v3 `Supports RDMA read with EFA for p4d(e).24xlarge instances`.  p5en.48xlarge runs on Nitro v5 and support RMDA write

NCCL Support in EFA ( Nitro V3)

Aggiungi la tua risposta

Contenuto pertinente