openucx/ucc

Unified Collective Communication Library

What it solves

UCC provides a unified API and library for collective communication operations, ensuring high performance and scalability for AI/ML, High-Performance Computing (HPC), and I/O workloads across different hardware and programming models.

How it works

It acts as a flexible communication layer that supports nonblocking operations and a relaxed ordering model. It integrates with various transports and hardware collectives, treating hardware-accelerated collectives as first-class citizens. It is built on top of UCX and supports multiple GPU backends including CUDA and HIP (AMD).

Who it’s for

Developers working on large-scale AI/ML training and HPC applications that require efficient data exchange between multiple processing units (CPUs, GPUs, and DPUs).

Highlights

  • Broad Hardware Support: Compatible with InfiniBand, ROCE, CUDA, NCCL, RCCL, and SHARP.
  • Scalable Performance: Designed specifically for high-scale AI/ML and HPC workloads.
  • Flexible Programming: Supports repetitive collective operations (init once, invoke multiple times) and nonblocking operations.
  • Multi-Runtime Integration: Works with Open MPI and OpenSHMEM.

Related

  • Project
    ROCm/rccl

    A library of standard collective communication routines for GPUs that enables high-bandwidth data synchronization across multiple GPUs and nodes.

  • Project
    uccl-project/uccl

    An efficient, portable GPU communication library that provides high-performance collectives, P2P transfers, and expert-parallel communication across heterogeneous GPU and NIC hardware.

  • Project
    microsoft/mscclpp

    A GPU-driven communication stack that provides customizable, 0-copy inter-GPU communication primitives to improve performance for scalable AI applications and LLM serving.

  • Project
    NVIDIA/cccl

    A unified collection of core CUDA C++ libraries (Thrust, CUB, and libcudacxx) providing high-performance parallel algorithms and standard library abstractions for GPU development.

  • Project
    uxlfoundation/oneDAL

    A C++ and DPC++ library that provides hardware-accelerated machine learning routines for tabular data across CPUs, GPUs, and distributed systems.