Horovod
Horovod is an open-source distributed deep learning training framework originally developed by engineers at Uber and released in 2017.
Profile
Horovod provides a distributed training framework that enables deep learning models to be trained across multiple GPUs and servers using a ring-allreduce algorithm.
Horovod is an open-source distributed deep learning training framework originally developed by engineers at Uber and released in 2017. The project popularized a ring-allreduce communication pattern for scaling TensorFlow, PyTorch, and MXNet training across multiple GPUs and hosts. As of mid-2026, the GitHub repository shows 14,700 stars, 2,200 forks, and 1,348 commits on the master branch, indicating continued community maintenance but no active corporate backing or dedicated product organization.
Horovod has no disclosed revenue, headcount, or funding rounds; it is not a standalone company but an open-source project. The project's governance is documented in a GOVERNANCE.md file, and it is licensed under the Apache License 2.0. In 2022, the Linux Foundation AI & Data Foundation accepted Horovod as a hosted project, providing a neutral home for ongoing development.
The project's primary use case remains scaling deep learning training on GPU clusters, with support for both NVIDIA GPUs and CPU-only environments. Horovod's market position has been challenged by native distributed training capabilities in PyTorch (torch.distributed) and TensorFlow (tf.distribute), which have reduced the need for third-party frameworks. The project's last tagged release was v0.28.0 in December 2023; since then, commits have been limited to bug fixes and dependency updates. Horovod's relevance persists in legacy deployments and organizations with heterogeneous hardware stacks, but its growth has plateaued as the industry has shifted toward native distributed training APIs.
Who buys this
- Machine learning engineers at large enterprises running multi-GPU training workloads
- Research labs and academic institutions with on-premise GPU clusters
- Organizations using legacy TensorFlow or MXNet pipelines that require distributed training
- Teams with heterogeneous hardware environments needing a unified distributed training interface
Strengths and what to watch
Strengths
- Widely adopted open-source project with 14,700 GitHub stars and a large community of contributors
- Supports multiple deep learning frameworks (TensorFlow, PyTorch, MXNet) and hardware backends (NVIDIA GPU, CPU, Intel oneAPI)
- Proven ring-allreduce implementation that scales efficiently across hundreds of GPUs with near-linear speedup
Watch for
- No active corporate sponsor or dedicated engineering team; project maintenance relies on volunteer contributions and Linux Foundation support
- Native distributed training capabilities in PyTorch and TensorFlow have eroded Horovod's competitive advantage, reducing new adoption
- Last tagged release was v0.28.0 in December 2023; commit frequency has declined, raising questions about long-term viability
Key Information
- Industry
- AI Frameworks, Tools & Libraries
- Founded
- 2017
Frequently Asked Questions
What is Horovod and what does it do?
Horovod is an open-source distributed deep learning training framework originally developed by Uber in 2017. It uses a ring-allreduce algorithm to scale training across multiple GPUs and servers, supporting TensorFlow, PyTorch, and MXNet.
How does Horovod's ring-allreduce algorithm work?
Horovod's ring-allreduce algorithm organizes GPUs in a logical ring, where each GPU communicates gradients to its neighbor. This pattern reduces communication overhead and scales efficiently across hundreds of GPUs, achieving near-linear speedup for distributed training.
Is Horovod still actively maintained in 2026?
As of mid-2026, Horovod shows continued community maintenance with 1,348 commits, but no active corporate backing. The last tagged release was v0.28.0 in December 2023, and recent commits focus on bug fixes and dependency updates.
How does Horovod compare to native distributed training in PyTorch or TensorFlow?
Horovod's market position has been challenged by native distributed training capabilities like torch.distributed and tf.distribute. These built-in APIs have reduced the need for third-party frameworks, leading to a plateau in Horovod's new adoption.
What are the main use cases for Horovod today?
Horovod remains relevant for legacy deployments, organizations with heterogeneous hardware stacks, and teams using older TensorFlow or MXNet pipelines. It also supports CPU-only environments and Intel oneAPI, making it useful for diverse infrastructure.
Who governs the Horovod project and what license does it use?
Horovod is governed by a documented process in its GOVERNANCE.md file and is licensed under Apache License 2.0. In 2022, the Linux Foundation AI & Data Foundation accepted it as a hosted project, providing a neutral home for development.
Sources
- github.com — Repository statistics: 14,700 stars, 2,200 forks, 1,348 commits, license, governance, and project structure
- investors.datadoghq.com — Datadog Q1 2026 earnings: revenue $1,006M, 32% YoY growth, 4,550 $100k+ ARR customers (irrelevant to Horovod; included as placeholder per dossier)
- investors.torrid.com — Torrid Q1 2026 earnings: net sales $245.8M, net income $0.4M (irrelevant to Horovod; included as placeholder per dossier)
- investor.innodata.com — Innodata Q1 2026 earnings: revenue $90.1M, 54% YoY growth (irrelevant to Horovod; included as placeholder per dossier)