NVIDIA Enhances Multi-GPU Communication with NCCL 2.26 Release

Darius Baruo
Jun 18, 2025 17:23

NVIDIA’s NCCL 2.26 introduces performance enhancements, improved monitoring, and quality of service features, optimizing multi-GPU and multinode communications for AI and HPC applications.

NVIDIA has announced the release of version 2.26 of its Collective Communications Library (NCCL), a pivotal update aimed at enhancing multi-GPU and multinode communication capabilities. NCCL 2.26 introduces significant performance improvements, advanced monitoring capabilities, and enhanced quality of service (QoS) for users, according to NVIDIA’s blog post.

Key Features and Enhancements

The new release of NCCL, part of NVIDIA’s Magnum IO suite, is designed to optimize the performance of inter-GPU and multinode communications, crucial for AI and high-performance computing (HPC) applications. The update introduces several key features:

• PAT Optimizations: Enhancements to the Parallel All-Reduce Tree (PAT) algorithm to improve execution efficiency, notably in large-scale operations.
• Implicit Launch Order: New functionality to prevent deadlocks and ensure synchronized operation launches across multiple communicators.
• Profiler Support: Expanded support for GPU kernel and network profiling, allowing for detailed performance analysis at the kernel and network levels.
• QoS Control: Introduction of communicator-level QoS controls to manage network resource allocation efficiently.
• RAS Improvements: Stability and diagnostic enhancements for more reliable and informative collective operations.

Detailed Feature Analysis

The PAT optimization separates computation and execution processes, allowing multiple warps to execute steps concurrently, thus enhancing performance in scenarios with numerous parallel trees. The implicit launch order feature, controlled via NCCL_LAUNCH_ORDER_IMPLICIT, reduces the risk of deadlocks by automatically managing kernel launch dependencies.

Profiler enhancements include new kernel profiler infrastructure and network-defined event support, which provide a comprehensive view of NCCL’s performance. The network plugin QoS support introduces a trafficClass field, enabling applications to prioritize critical network communications, thereby improving end-to-end performance in overlapping communications scenarios.

Bug Fixes and Minor Updates

NCCL 2.26 also addresses several bugs and introduces minor features, such as Direct NIC support, enhanced diagnostic message timestamping, and improved memory usage with NVLink SHARP. These updates contribute to better performance and reliability across various systems.

For more details on the NCCL 2.26 release, visit the NVIDIA blog.

Image source: Shutterstock