Bug#1022889: ITP: rccl -- ROCm Communication Collectives Library
Package: wnpp
Severity: wishlist
Owner: Cordell Bloor <cgmb@slerp.xyz>
X-Debbugs-Cc: debian-devel@lists.debian.org, cgmb@slerp.xyz, debian-ai@lists.debian.org
* Package name : rccl
Version : 5.3.0
Upstream Author : Advanced Micro Devices, Inc.
* URL : https://github.com/ROCmSoftwarePlatform/rccl
* License : BSD 3-Clause
Programming Lang: C, C++, HIP
Description : ROCm Communication Collectives Library
RCCL (pronounced "Rickle") is a library of collective communication routines
built on the AMD ROCm platform and optimized for discrete AMD GPUs. It
implements routines such as all-reduce, all-gather, reduce, broadcast,
reduce-scatter, gather, scatter and all-to-all. The library is optimized for
high bandwidth over channels including PCIe, xGMI, InfiniBand Verbs or TCP/IP
sockets and supports an arbitrary number of GPUs.
This library provides multi-GPU communication functionality that helps users
depending on NVIDIA's NCCL library to port their code to run on AMD GPUs. RCCL
is a dependency of both PyTorch and Tensorflow.
This package is part of AMD's ROCm stack and will be maintained under the
Debian AI team umbrella.
Reply to: