This page provides release notes for each version of NCCL/gIB on x86 64-bit
systems. Each release section describes the included binaries, supported
machine series, and new features and enhancements.
NCCL/gIB v1.1.2
Release date: May 14, 2026
Prebuilt binaries
The following sections list the prebuilt release and diagnostic binaries.
Release
Install these binaries for production workloads. For installation instructions,
see Run NCCL on Compute Engine instances.
Docker installer
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib:v1.1.2 |
| SHA256 |
b71627dcf88862cd0e64434dfd0224441960db58f2022f94b5f56123659c207e |
| Base |
Alpine 3.21.0 |
| Built in |
Ubuntu 22.04 + CUDA 13.0 |
.deb package
| Repository |
https://packages.cloud.google.com/apt gpudirect-gib-apt |
| Packages |
nccl-gib (1.1.2), nccl-gib-plugins (1.1.2) |
| Base |
Ubuntu 22.04 |
| Built in |
Ubuntu 22.04 + CUDA 13.0 |
.rpm package
| Repository |
https://packages.cloud.google.com/yum/repos/gpudirect-gib-rpm |
| Packages |
nccl-gib (1.1.2-1), nccl-gib-plugins (1.1.2-1) |
| Base |
Rocky Linux 9 |
| Built in |
Ubuntu 22.04 + CUDA 13.0 |
Diagnostic
Install these binaries to test or troubleshoot network performance.
Docker container
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib-diagnostic:v1.1.2 |
| SHA256 |
61c917022538e59d24f597038a8ad7087a026d60715256ae39e7f06ef20ebde0 |
| Base |
Ubuntu 22.04 + CUDA 13.0 |
Version compatibility
| Machine series |
Bundled NCCL version |
Supported NCCL versions |
Minimum GPU driver version |
Minimum CUDA runtime version |
| A3 Ultra |
2.30.4 |
>= 2.23 |
550.90.07 |
12.6 |
| A4 |
2.30.4 |
>= 2.25 |
580.126.09 |
12.8 |
Component breakdown
| Component |
Version |
| rdma-core (not included in .deb and .rpm packages) |
2307mlnx47 |
| nccl |
2.30.4 |
| nccl-plugin-gib |
1.1.2 |
Version logging
To see the version logs, set the NCCL_DEBUG=INFO and NCCL_DEBUG_SUBSYS=INIT
environment variables.
| Component |
Version Logging |
| nccl |
NCCL version 2.30.4+cuda13.0 |
| nccl-plugin-gib |
NET/gIB : Initializing gIB v1.1.2 |
Features and enhancements
Environment plugin
- Introduces the Google Cloud NCCL environment plugin that automatically applies
the best NCCL environment variables based on the runtime platform. The
environment plugin is automatically loaded when you use NCCL 2.28.7 or later.
gIB net plugin
Bug fixes
- Fixes a bug in the gIB net plugin when handling large
isend() requests
- Fixes an issue where the gIB net plugin might not be compatible with NCCL 2.27
or later
- Fixes stability issues in CoMMA
CoMMA (profiler plugin):
gIB tuner plugin
- Automatically applies optimal tuning configs on any detected Google Cloud
platform.
nccl.conf
- gIB writes to /etc/nccl.conf when the file is write-accessible. As a result,
when you install the nccl-gib or nccl-gib-plugins packages, you don't need to
change any environmental variable settings, including
LD_LIBRARY_PATH.
New package
nccl-gib-plugins is a new package that only contains Google's NCCL plugins.
To use this package, you must install NCCL from NVIDIA repositories, and
install the APT package nccl-gib-plugins package from Google's repository.
The Docker installer, diagnostic container, and the nccl-gib package include
a pre-compiled NCCL and are still available.
NCCL/gIB v1.1.0
Release date: Jul 14, 2025
Prebuilt binaries
The following sections list the prebuilt release and diagnostic binaries.
Release
Install these binaries for production workloads. For installation instructions,
see Run NCCL on Compute Engine instances.
Docker installer
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib:v1.1.0 |
| SHA256 |
81c7165f1b41ee096d86f4cfe5a61f5c788546cf6ddc88aba3f0b88a5902af49 |
| Base |
Alpine 3.21.0 |
| Built in |
Ubuntu 20.04 + CUDA 12.8 |
.deb package
| Repository |
https://packages.cloud.google.com/apt gpudirect-gib-apt |
| Package |
nccl-gib (1.1.0) |
| Base |
Ubuntu 20.04 |
| Built in |
Ubuntu 20.04 + CUDA 12.8 |
.rpm package
| Repository |
https://packages.cloud.google.com/yum/repos/gpudirect-gib-rpm |
| Package |
nccl-gib (1.1.0-1) |
| Base |
Rocky Linux 9 |
| Built in |
Ubuntu 20.04 + CUDA 12.8 |
Diagnostic
Install these binaries to test or troubleshoot network performance.
Docker container
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib-diagnostic:v1.1.0 |
| SHA256 |
ee9682122220105b83187bd6eecc8cafbe739dcfb631a54b6059a8f1141038f9 |
| Base |
Ubuntu 20.04 + CUDA 12.8 |
Version compatibility
| Machine series |
Bundled NCCL version |
Supported NCCL versions |
Minimum GPU driver version |
Minimum CUDA runtime version |
| A3 Ultra |
2.27.5 |
>= 2.23 |
550.90.07 |
12.6 |
| A4 |
2.27.5 |
>= 2.25 |
570.124.06 |
12.8 |
Component breakdown
| Component |
Version |
| rdma-core (not included in .deb and .rpm packages) |
2307mlnx47 |
| nccl |
2.27.5 |
| nccl-plugin-gib |
1.1.0 |
Version logging
To see the version logs, set the NCCL_DEBUG=INFO and NCCL_DEBUG_SUBSYS=INIT
environment variables.
| Component |
Version Logging |
| nccl |
NCCL version 2.27.5+cuda12.8 |
| nccl-plugin-gib |
NET/gIB : Initializing gIB v1.1.0 |
Features and enhancements
Overall
- GA of always-on NCCL telemetry
- This version provides all telemetry necessary for straggler detection
CoMMA (profiler plugin) GA
- Stability improvements
- Adds the ability to track sender-side wait for receivers
- You can find the open-source version in the
CoMMA GitHub repository
For more information, see CoMMA overview.
NCCL/gIB v1.0.6
Release date: Jun 16, 2025
Prebuilt binaries
The following sections list the prebuilt release and diagnostic binaries.
Release
Install these binaries for production workloads. For installation instructions,
see Run NCCL on Compute Engine instances.
Docker installer
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib:v1.0.6 |
| SHA256 |
b0d5e1f29ca89c33b5b6252ca635a8c3889f553de3a382ba07c9b1f43fcf452f |
| Base |
Alpine 3.21.0 |
| Built in |
Ubuntu 20.04 + CUDA 12.8 |
.deb package
| Repository |
us-apt.pkg.dev/projects/gce-ai-infra/gpudirect-gib-apt |
| Package |
nccl-gib (1.0.6.1) |
| Base |
Debian 12 |
| Built in |
Ubuntu 20.04 + CUDA 12.8 |
.rpm package
| Repository |
us-yum.pkg.dev/projects/gce-ai-infra/gpudirect-gib-rpm |
| Package |
nccl-gib (1.0.6-1) |
| Base |
Rocky Linux 9 |
| Built in |
Ubuntu 20.04 + CUDA 12.8 |
Diagnostic
Install these binaries to test or troubleshoot network performance.
Docker container
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib-diagnostic:v1.0.6 |
| SHA256 |
a9ad0984f28e0c2ab40ea21f2c4e861c4ff9fffe1ba4a73f5efa2ec911ad7ca4 |
| Base |
Ubuntu 20.04 + CUDA 12.8 |
Version compatibility
| Machine series |
Bundled NCCL version |
Supported NCCL versions |
Minimum GPU driver version |
Minimum CUDA runtime version |
| A3 Ultra |
2.26.6 |
2.23, 2.24, 2.25, 2.26 |
550.90.07 |
12.6 |
| A4 |
2.26.6 |
2.25, 2.26 |
570.124.06 |
12.8 |
Component breakdown
| Component |
Version |
| rdma-core (not included in .deb and .rpm packages) |
2307mlnx47 |
| nccl |
2.26.6 |
| nccl-plugin-gib |
1.0.6 |
Version logging
To see the version logs, set the NCCL_DEBUG=INFO and NCCL_DEBUG_SUBSYS=INIT
environment variables.
| Component |
Version Logging |
| nccl |
NCCL version 2.26.6+cuda12.8 |
| nccl-plugin-gib |
NET/gIB : Initializing gIB v1.0.6 |
Features and enhancements
Overall
- Major updates to NCCL profiler
- A4 performance tuning optimization
- NCCL shim fixes for NCCL 2.26
NCCL profiler
- Adds v3 API support (without CUDA kernel tracking)
- Profiler plugin loads by default in the guest environment
- Updates the tuning scripts for the A4 platform up to the 64-node scale
- Applies specific collective tuning for all-gather, all-reduce, reduce-scatter,
and broadcast
Miscellaneous
- NCCL shim: fixes the NCCL 2.26 gIB loading issue
- NCCL guest config: removes the enforcement on intra-node P2P chunk sizes
- NCCL guest config: removes the enforcement of NCCL_IB_GID_INDEX
- NCCL core: updates to 2.26.5
- Removes the pci-utils dependency
NCCL/gIB v1.0.5
Release date: Mar 12, 2025
Prebuilt binaries
The following sections list the prebuilt release and diagnostic binaries.
Release
Install these binaries for production workloads. For installation instructions,
see Run NCCL on Compute Engine instances.
Docker installer
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib:v1.0.5 |
| SHA256 |
19eb592ec6ba4c5fae0257bbff5a29af5a7a94a0a59f7b0afc1ace488b69dc39 |
| Base |
Alpine 3.21.0 |
| Built in |
Ubuntu 22.04 + CUDA 12.8 |
.deb package
| Repository |
us-apt.pkg.dev/projects/gce-ai-infra/gpudirect-gib-apt |
| Package |
nccl-gib (1.0.5) |
| Base |
Debian 12 |
| Built in |
Ubuntu 22.04 + CUDA 12.8 |
.rpm package
| Repository |
us-yum.pkg.dev/projects/gce-ai-infra/gpudirect-gib-rpm |
| Package |
nccl-gib (1.0.5-2) |
| Base |
Rocky Linux 9 |
| Built in |
Ubuntu 20.04 + CUDA 12.8 |
Diagnostic
Install these binaries to test or troubleshoot network performance.
Docker container
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib-diagnostic:v1.0.5 |
| SHA256 |
e8edbfe3cf6695d578dd38d7dda6ecdc8a357a1ed56f7f8593fc38cacd9fa05a |
| Base |
Ubuntu 22.04 + CUDA 12.8 |
Version compatibility
| Machine series |
Bundled NCCL version |
Supported NCCL versions |
Minimum GPU driver version |
Minimum CUDA runtime version |
| A3 Ultra |
2.25 |
2.23, 2.24, 2.25 |
550.90.07 |
12.6 |
| A4 |
2.25 |
2.25 |
570.124.06 |
12.8 |
Component breakdown
| Component |
Version |
| rdma-core (not included in .deb and .rpm packages) |
2307mlnx47 |
| nccl |
2.25.1 |
| nccl-plugin-gib |
1.0.5 |
Version logging
To see the version logs, set the NCCL_DEBUG=INFO and NCCL_DEBUG_SUBSYS=INIT
environment variables.
| Component |
Version Logging |
| nccl |
NCCL version 2.25.1+cuda12.8 |
| nccl-plugin-gib |
NET/gIB : Initializing gIB v1.0.5 |
Features and enhancements
Overall
- New: added NCCL/gIB .deb and .rpm packages
Profiler plugin
- Experimental: added Google's CoMMA profiler plugin implementation
- Improves all_reduce and all_to_all performance on A4 VMs
- Improves large scale collective and all_to_all performance on A3 Ultra VMs
Miscellaneous
- Automates NCCL configurations based on detected machine shape
- Recommends forcing the use of the gIB net plugin
Known issues
- When using NCCL 2.26 or later, setting
NCCL_NET=gIB might cause NCCL to
crash. This was fixed in NCCL/gIB v1.0.6.
NCCL/gIB v1.0.4
Release date: Feb 24, 2025
Prebuilt binaries
The following sections list the prebuilt release and diagnostic binaries.
Release
Install these binaries for production workloads. For installation instructions,
see Run NCCL on Compute Engine instances.
Docker installer
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib:v1.0.4 |
| SHA256 |
400bfb76ad4f3c153028a4e744ed3c4e924d2dae46e09b0efb2056804b8ed7cf |
| Base |
Alpine 3.21.0 |
| Built in |
Ubuntu 22.04 + CUDA 12.8 |
Diagnostic
Install these binaries to test or troubleshoot network performance.
Docker container
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib-diagnostic:v1.0.4 |
| SHA256 |
fd7b1999e175306d4ac62b1cb8039b1b3386dbb8a529fae03256714f3284e460 |
| Base |
Ubuntu 22.04 + CUDA 12.8 |
Version compatibility
| Machine series |
Bundled NCCL version |
Supported NCCL versions |
Minimum GPU driver version |
Minimum CUDA runtime version |
| A3 Ultra |
2.25 |
2.23, 2.24, 2.25 |
550.90.07 |
12.6 |
Component breakdown
| Component |
Version |
| rdma-core (not included in .deb and .rpm packages) |
2307mlnx47 |
| nccl |
2.25.1 |
| nccl-plugin-gib |
1.0.4 |
Version logging
To see the version logs, set the NCCL_DEBUG=INFO and NCCL_DEBUG_SUBSYS=INIT
environment variables.
| Component |
Version Logging |
| nccl |
NCCL version 2.25.1+cuda12.8 |
| nccl-plugin-gib |
NET/gIB : Initializing gIB v1.0.4 |
Features and enhancements
Overall
- Rebased to NCCL core 2.25 to support A4 and A4X VMs
- Added arm64 builds
GPUViz
- Added GPUViz to the release build
Miscellaneous
- Allowed NCCL to automatically detect the socket interface to use during
bootstrapping
Bug fixes
- Minor bug fixes to GPUViz
Known issues
- When using NCCL 2.26 or later, setting
NCCL_NET=gIB might cause NCCL to
crash. This was fixed in NCCL/gIB v1.0.6.
NCCL/gIB v1.0.3
Release date: Dec 18, 2024
Prebuilt binaries
The following sections list the prebuilt release and diagnostic binaries.
Release
Install these binaries for production workloads. For installation instructions,
see Run NCCL on Compute Engine instances.
Docker installer
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib:v1.0.3 |
| SHA256 |
04e288a184d5eac1eae7feedb275ddb1e8d7df9c1c0c5f794824151a47eafe90 |
| Base |
Alpine 3.21.0 |
| Built in |
Ubuntu 22.04 + CUDA 12.2 |
Diagnostic
Install these binaries to test or troubleshoot network performance.
Docker container
| Image |
us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib-diagnostic:v1.0.3 |
| SHA256 |
29f7a13a7a55fd09bfb266275fabec60835e0bde9d5deecffd676005e514a6f5 |
| Base |
Ubuntu 22.04 + CUDA 12.2 |
Version compatibility
| Machine series |
Bundled NCCL version |
Supported NCCL versions |
Minimum GPU driver version |
Minimum CUDA runtime version |
| A3 Ultra |
2.23 |
2.23, 2.24, 2.25 |
550.90.07 |
12.6 |
Component breakdown
| Component |
Version |
| rdma-core (not included in .deb and .rpm packages) |
2307mlnx47 |
| nccl |
2.23.4 |
| nccl-plugin-gib |
1.0.3 |
Version logging
To see the version logs, set the NCCL_DEBUG=INFO and NCCL_DEBUG_SUBSYS=INIT
environment variables.
| Component |
Version Logging |
| nccl |
NCCL version 2.23.4+cuda12.2 |
| nccl-plugin-gib |
NET/gIB : Initializing gIB v1.0.3 |
Features and enhancements
gIB net plugin
- Adds logging for gIB net plugin environment variables
- Enables guest-assisted multi-pathing by default
- Used more QPs for more sophisticated P2P patterns by default
gIB guest config checker
- Gives a warning when guest config checker cannot find a config file
- Updates guest config to match the recommended environment variables for gIB
v1.0.3
gIB tuner plugin
- Adds support for the Parallel Aggregated Tree (PAT) algorithm
- Updates tuner configurations for world-level all-reduce at 9 or more nodes for
gIB v1.0.3
- Added tuner configs for rail-aligned collectives.
GPUViz
- Added GPUViz to the diagnostic build.
Miscellaneous
- Important, non-backwards compatible change: the NCCL/gIB
container is now split into two container images, release and diagnostic
- Release
- ~230 MB without compression, ~100 MB with compression
- Only contains the strictly necessary files for a minimal
installer
- Non-backwards compatible change:
The shell and daemon subcommands have been removed from this
image
- Diagnostic
- ~10 GB without compression, ~5 GB with compression
- Designed for debugging and contains many debugging utilities and
benchmarks like NCCL tests and perftest
- Updated the script to set recommended NCCL environment variables for gIB v1.0.3
- Diagnostic: Installed ping in the diagnostic build
- Diagnostic: Added neper to the diagnostic build
- Diagnostic: Added a script that dumps RDMA-related system state
Bug fixes
- Fixed an issue where NCCL core symlinks are not copied optimally.
Known issues
- When using NCCL 2.26 or later, setting
NCCL_NET=gIB may cause NCCL to crash. This was fixed in NCCL/gIB v1.0.6.
