How OpenStack Cyborg Unifies Management of GPUs, FPGAs, ASICs and Other Accelerators

With the rapid rise of high‑performance workloads like NFV, edge computing, HPC and artificial intelligence, cloud platforms need to manage proprietary hardware (GPU, FPGA, ASIC, SoC) more effectively.

Current cloud infrastructure often fails to schedule CPU‑GPU resources efficiently, leading to GPU waste, or cannot manage accelerator cards, causing FPGA‑based smart NICs to remain idle.

To address these challenges, Huawei, together with Intel, Lenovo and others, launched the OpenStack Cyborg project in 2017 within the OpenStack open‑source community.

Cyborg aims to provide a universal framework for managing proprietary hardware, making it easier for infrastructure providers and users to deploy performance‑critical services.

After gaining broad community support, Cyborg became an official OpenStack project in September 2017 and has released the Queens and Rocky versions; development is now active on the Stein release.

Cyborg consists of modules such as API, conductor, agent, and database. By interacting with the Placement service, it reports proprietary hardware resources and collaborates with the Nova scheduler so that virtual machines can be created using these accelerators.

The latest Rocky release adds support for flashing FPGA instances, a new os‑acc hardware‑attach library, a Python client CLI, preliminary multi‑tenant quota, and standardized metadata for south‑bound devices. The upcoming Stein release will deepen Nova integration, broaden hardware support, and provide connectors for PaaS platforms like Kubernetes.

Overall, Cyborg simplifies the management of encryption, transcoding, and neural‑network accelerators for emerging NFV, edge, HPC, and AI workloads, accelerating their cloud adoption.

The project is actively developed by a diverse community that includes contributors from Huawei, Intel, Lenovo, China Mobile, ZTE, Nokia, Tencent, UnionPay, Inspur, 99Cloud, iFlytek, Red Hat, Fujitsu, as well as academic institutions such as Fudan University, Jiangnan Institute, UC Berkeley, Carnegie Mellon, and UIUC.