Full documentation for RCCL is available at https://rccl.readthedocs.io
- Compatibility with NCCL 2.12.10
- Packages for test and benchmark executables on all supported OSes using CPack.
- Adding custom signal handler - opt-in with RCCL_ENABLE_SIGNALHANDLER=1
- Additional details provided if Binary File Descriptor library (BFD) is pre-installed
- Removed experimental clique-based kernels
- Compatibility with NCCL 2.11.4
- Managed memory is not currently supported for clique-based kernels
- Compatibility with NCCL 2.10.3
- Managed memory is not currently supported for clique-based kernels
- Packaging split into a runtime package called rccl and a development package called rccl-devel. The development package depends on runtime. The runtime package suggests the development package for all supported OSes except CentOS 7 to aid in the transition. The suggests feature in packaging is introduced as a deprecated feature and will be removed in a future rocm release.
- Compatibility with NCCL 2.9.9
- Managed memory is not currently supported for clique-based kernels
- Ability to select the number of channels to use for clique-based all reduce (RCCL_CLIQUE_ALLREDUCE_NCHANNELS). This can be adjusted to tune for performance when computation kernels are being executed in parallel.
- Additional tuning for clique-based kernel AllReduce performance (still requires opt in with RCCL_ENABLE_CLIQUE=1)
- Modification of default values for number of channels / byte limits for clique-based all reduce based on device architecture
- Replaced RCCL_FORCE_ENABLE_CLIQUE to RCCL_CLIQUE_IGNORE_TOPO
- Clique-based kernels can now be enabled on topologies where all active GPUs are XGMI-connected
- Topologies not normally supported by clique-based kernels require RCCL_CLIQUE_IGNORE_TOPO=1
- Install script '-r' flag invoked alone no longer incorrectly deletes any existing builds.
- Managed memory is not currently supported for clique-based kernels
- Compatibility with NCCL 2.8.4
- Additional tuning for clique-based kernels
- Enabling GPU direct RDMA read from GPU
- Fixing potential memory leak issue when re-creating multiple communicators within same process
- Improved topology detection
- None
- Experimental support for clique-based kernels (opt in with RCCL_ENABLE_CLIQUE=1)
- Clique-based kernels may offer better performance for smaller input sizes
- Clique-based kernels are currently only enabled for AllReduce under a certain byte limit (controlled via RCCL_CLIQUE_ALLREDUCE_BYTE_LIMIT)
- Performance improvements for Rome-based systems
- Clique-based kernels are currently experimental and have not been fully tested on all topologies. By default, clique-based kernels are disabled if the detected topology is not supported (override with RCCL_FORCE_ENABLE_CLIQUE)
- Clique-based kernels may hang if there are differences between environment variables set across ranks.
- Clique-based kernels may fail if the input / output device pointers are not the base device pointers returned by hipMalloc.
- Adding support for alltoallv RCCL kernel
- Modifications to topology based on XGMI links
- None
- Support for static library builds
- None
- Updated to RCCL API version of 2.7.6
- Added gather, scatter and all-to-all collectives
- Updated to RCCL API version of 2.6.4
- Compatibility with NCCL 2.6
- Network interface improvements with API v3
- Fixing issues and built time improvements for hip-clang
- Network topology detection
- Improved CPU type detection
- Infiniband adaptive routing support
- Switched to hip-clang as default compiler
- Deprecated hcc build