Skip to main content

AMD Accelerators in CloudBender Nodes

· One min read

proxiML now supports AMD accelerators as a compute backend for CloudBender nodes. Workloads can run on AMD GPUs through ROCm, expanding the hardware options available to CloudBender deployments beyond NVIDIA.

What Changed​

Previously, CloudBender nodes only supported NVIDIA GPUs with CUDA. With this update, a node's compute backend is determined at registration time and can be either CUDA (NVIDIA) or ROCm (AMD). The platform detects the backend automatically and routes workloads accordingly.

What This Means​

CloudBender users can now provision nodes with AMD GPUs and run training, notebook, and inference jobs on them through the proxiML platform. The scheduler considers backend-specific GPU types when matching jobs to available resources, and GPU reservations track the backend type to ensure compatible placement.

See the CloudBender Reference for more details.