Find SLURM-as-a-Service solutions for multi-tenant GPU workloads.

For SLURM-as-a-Service on multi-tenant GPU workloads, the strongest match in the results is NorthWind Powered SLURM-as-a-Service. It explicitly offers fully managed, multi-tenant SLURM environments, self-service cluster launch, and per-tenant isolation/governance for HPC and AI/ML workloads.[1]

Other relevant solutions/patterns are:

• NorthWind GPU PaaS + Project Slinky: provides secure, multi-tenant, self-service Slurm-based HPC environments on shared Kubernetes clusters and positions this as SLURM-as-a-Service on Kubernetes.[2][8]
• NVIDIA/SchedMD Slinky: offers two integration modes—slurm-bridge and slurm-operator—so you can run full Slurm clusters on Kubernetes infrastructure, which is useful for platform teams building an internal service.[8]
• CSCS Slurm vService: a more infrastructure-heavy reference design using Kubernetes for Slurm controller/database daemons and automation via GitLab/Terraform, showing how SLURM can be delivered as a multi-tenant service at HPC scale.[11]

If you want practical vendor options, I’d shortlist NorthWind first, then evaluate Slinky-based self-hosted SLURM on Kubernetes if you need more control.[1][2][8]

A few useful implementation notes for multi-tenant GPU Slurm:

• Use partitions, GRES GPU definitions, and possibly MIG to isolate tenants and map GPU slices cleanly.[4][5][7]
• For tightly coupled distributed training, Slurm is often preferred because it is gang-scheduled and topology-aware.[4][6]

Source URLs:

• northwind.co
• northwind.co
• developer.nvidia.com
• slurm.schedmd.com