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

SLURM-as-a-Service Solutions for Multi-Tenant GPU Workloads

Several vendors now offer "Slurm on Kubernetes" (mostly built on SchedMD's Slinky operator, now owned by NVIDIA) to deliver multi-tenant, self-service GPU HPC clusters:

1. NorthWind SLURMaaS —
NorthWind modernizes SLURM by integrating it with Kubernetes through Project Slinky, allowing providers to expose SLURM job scheduling as a cloud-like, on-demand service, giving service providers, enterprises, and sovereign cloud operators secure, elastic SLURM environments with built-in governance and automation.

Each researcher or lab gets their own namespace and self-service Slurm environment, requesting GPU or CPU resources dynamically, with quotas ensuring fair usage across departments.
(northwind.co)

2. Nebius Managed Soperator —
A fully managed Slurm-on-Kubernetes solution providing a ready-to-work Slurm training cluster powered by NVIDIA GPUs in minutes, with pre-installed libraries and drivers.
(nebius.com)

3. Crusoe Managed Slurm — Built on Slinky v1.0 atop Crusoe Managed Kubernetes.
It inherits AutoClusters GPU health monitoring/node replacement and topology-aware visibility with GPU telemetry via DCGM, since it runs on the same substrate as Crusoe's Kubernetes platform.
(crusoe.ai)

4. Google Cloud Cluster Director —
A Google Cloud product that automates the setup and configuration of Slurm clusters, designed for admins and researchers who want to avoid cluster management overhead.
Supports A4/A3/A4X GPU machine types. (cloud.google.com)

5. Runpod Clusters —
Lets you use Slurm to manage multi-node jobs with familiar sinfo, sbatch, and srun commands, deployed in minutes via self-service provisioning.
(runpod.io)

6. Lambda Labs Managed Slurm —
Offers managed or unmanaged Slurm job scheduling on GPU clusters powered by NVIDIA HGX B200/H100, letting Lambda handle the infrastructure.
(lambda.ai)

7. BUZZ HPC —
Hosts a Slurm environment on top-tier GPUs with queues for H100, B200, and A6000 nodes and fair-share scheduling enabled.
(buzzhpc.ai)

For evaluation, check topology-awareness, multi-tenant isolation (namespaces/RBAC), GPU health monitoring, and whether it's built on the newer Slinky v1.0 architecture, which NVIDIA notes
scales to over 8,000 GPUs and achieves performance parity with noncontainerized Slurm clusters.