There is a newer version of NCCLThe NVIDIA Collective Communications Library (NCCL) implements multi-GPU and multi-node collective communication primitives that are performance optimized for NVIDIA GPUs.
Accessing NCCL 2.8.3-CUDA-11.1.1
To load the module for NCCL 2.8.3-CUDA-11.1.1 please use this command on the BEAR systems (BlueBEAR, BEARCloud VMs, and CaStLeS VMs):
module load NCCL/2.8.3-CUDA-11.1.1
BEAR Apps Version
The listed architectures consist of two part: OS-CPU.
- BlueBEAR: The OS used on BlueBEAR is represented by EL and there are several different processor (CPU) types available on BlueBEAR. More information about the processor types on BlueBEAR is available on the BlueBEAR Job Submission page.
- BEAR and CaStLeS Cloud VMs: These VMs can have one of two OSes. Those with access to a BEAR Cloud or CaStLeS VM should check that the listed architectures for an application include the OS of VM being used. The VMs, irrespective of OS, will use the haswell CPU type.
For more information visit the NCCL website.
This version of NCCL has a direct dependency on: CUDAcore/11.1.1
This version of NCCL is a direct dependent of: libgpuarray/0.7.6-fosscuda-2020b PyTorch/1.7.1-fosscuda-2020b TensorFlow/2.4.1-fosscuda-2020b
These versions of NCCL are available on the BEAR systems (BlueBEAR, BEARCloud VMs, and CaStLeS VMs). These will be retained in accordance with our Applications Support and Retention Policy.
Last modified on 23rd February 2021