# An Extensible Software Transport Layer for GPU Networking

arXiv: https://arxiv.org/abs/2504.17307 · HTML: https://mlsanity.com/paper/2504.17307/

## Key facts

- Cited by 10 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 24 Apr 2025.

## Built on by (most cited first, 10 of 10)

- [LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference](https://mlsanity.com/paper/2510.09665/) (2025)
- [UCCL-EP: Portable Expert-Parallel Communication](https://mlsanity.com/paper/2512.19849/) (2025)
- [Collective Communication for 100k+ GPUs](https://mlsanity.com/paper/2510.20171/) (2025)
- [OrderMoE: An expert similarity driven distributed edge MoE inference](https://mlsanity.com/paper/2607.17154/) (2026)
- [RailS: Load Balancing for All-to-All Communication in Distributed Mixture-of-Experts Training](https://mlsanity.com/paper/2510.19262/) (2025)
- [HetCCL: Accelerating LLM Training with Heterogeneous GPUs](https://mlsanity.com/paper/2601.22585/) (2026)
- [TENT: A Declarative Slice Spraying Engine for Performant and Resilient Data Movement in Disaggregated LLM Serving](https://mlsanity.com/paper/2604.00368/) (2026)
- [mKernel: Fast Multi-GPU, Multi-Node Fused Kernels](https://mlsanity.com/paper/2609.13585/) (2026)
- [WeightBridge: An Efficient Weight Transfer Library for Reinforcement Learning](https://mlsanity.com/paper/2609.25442/) (2026)
- [Purlin: Separating Orchestration from the Datapath of Collectives](https://mlsanity.com/paper/2609.36954/) (2026)

## Abstract (arXiv)

Fast-evolving machine learning (ML) workloads have increasing requirements for networking. However, host network transport on RDMA NICs is hard to evolve, causing problems for ML workloads. For example, single-path RDMA traffic is prone to flow collisions that severely degrade collective communication performance. We present UCCL, an extensible software transport layer to evolve GPU networking. UCCL decouples the data path and control path of existing RDMA NICs and efficiently runs the control-path transport on host CPUs. This software extensibility brings in transport innovations that cannot be achieved in hardware for ML workloads, e.g., a multipath transport to resolve flow collisions. ML collectives atop UCCL achieve up to 4.5x higher performance compared to existing RDMA NICs.