Biagio Cosenza

Multi-node GPU systems are the foundation of modern, high-performance computing and deep neural networks training. While vendor-specific collective communication libraries such as NVIDIA’s NCCL deliver superior performance over traditional MPI, they sacrifice portability by tightly coupling applications to a single hardware platform. The oneAPI Collective Communications Library (oneCCL), as part of the oneAPI ecosystem, aims to provide a vendor-agnostic communication API for heterogeneous GPU clusters, but its support for non-Intel hardware remains limited. This paper presents a lightweight integration layer that bridges oneCCL’s portable API with vendor-optimized backends, mapping oneCCL collective calls to native implementations such as NCCL on NVIDIA GPUs and RCCL on AMD GPUs. This design preserves source-level portability through the SYCL programming model and the oneCCL API while delivering communication performance on par with the underlying native library. We evaluate our approach on NVIDIA A100 GPUs and AMD MI250x GPUs, comparing vensor specific libary such as NCCL and RCCL with our oneCCL integration across allreduce and alltoall collectives, with multiple data types and message sizes, in both intra-node and inter-node configurations. Experimental results show that oneCCL achieve near-identical performance against native backends without losing cross-platform portability.