CoreWeave adds guided self-service to its SUNK AI training system
CoreWeave is extending SUNK, its Slurm on Kubernetes training system, with a guided self-service preview for deploying and managing clusters through the CoreWeave console. The company says the operator-powered experience is designed to reduce operational overhead while preserving control over large AI training workloads.
- CoreWeave is previewing guided self-service for SUNK through its console.
- SUNK runs Slurm on Kubernetes for large-scale AI training.
- The system includes topology-aware scheduling and automated health management.
- CoreWeave reports up to 96% goodput and approximately 90 seconds to restart failed jobs.
- IBM and Cursor AI are cited as SUNK users.
A unified system for Slurm-based AI training
SUNK runs Slurm with Kubernetes, combining Slurm's interface for managing large-scale AI training resources with Kubernetes-based lifecycle management. CoreWeave says the system is designed to provide cloud-native agility, operational ease and scalable management as training infrastructure becomes more complex.
The company describes SUNK as an integrated production-grade system rather than a collection of separate infrastructure layers. Its capabilities include topology-aware scheduling for large, synchronous training workloads, automated health management, failure mitigation, operational visibility and lifecycle management for keeping clusters consistent and upgradeable.
CoreWeave says this approach is intended to help teams move beyond stitched-together infrastructure stacks. SUNK also includes SUNK User Provisioning, or SUP, which automates onboarding and permissions and provides isolated login environments. According to CoreWeave, teams can begin working in minutes while organizations retain access controls and governance.
Guided self-service comes to the CoreWeave console
The newly announced guided self-service experience is available as a preview. It provides an opinionated workflow for deploying and managing SUNK from the CoreWeave console, powered by a Kubernetes operator.
CoreWeave presents the feature as an extension of SUNK's operational model. Rather than requiring teams to assemble and maintain each part of a training environment themselves, the system is intended to provide a more streamlined way to engage with a production-ready cluster while maintaining the operational discipline needed for long-running jobs.
The announcement does not specify a general availability date or other release terms for the preview.
Designed for failures and long-running workloads
CoreWeave says frontier AI workloads can span tens of thousands of GPUs for days or weeks. In that environment, hardware events, GPU stragglers and small inefficiencies can reduce useful training time. SUNK is designed to maintain repeatable behavior under sustained, synchronous load across multi-thousand-GPU clusters.
In production deployments running long-duration training workloads, CoreWeave reports up to 96% goodput, 10x longer Mean Time to Failure for thousand-GPU clusters, and approximately 90 seconds to restart a job when failures occur. The company gives the benchmark comparison for Mean Time to Failure as approximately 3.66 days compared to roughly 8 hours.
SUNK works with CoreWeave Mission Control, which evaluates node health, tracks GPU performance variance and detects stragglers. Users can drain unhealthy nodes, reschedule workloads and isolate failure domains, according to CoreWeave. The system also connects infrastructure telemetry with job-level information such as communication bottlenecks, GPU variance, memory pressure and distributed training state.
Customer examples and external recognition
CoreWeave says IBM and Cursor AI use SUNK to scale AI training without building and maintaining bespoke infrastructure. IBM Senior Technical Staff Member Brian Belgodere said the system supports clusters with thousands of GPUs, while automated researcher onboarding enables access to be granted and revoked in under a minute.
Cursor AI's Sam Kottler said SUNK provided shared file systems, automated user provisioning and customizable environments. He also cited Mission Control's node health checks and remediation as reducing operational overhead.
CoreWeave says SemiAnalysis recognized SUNK in its ClusterMAX 2.0 rating as the only premium, unified training system available on the market at that time. The company also says it was recently named one of the first NVIDIA GB200 Exemplar Clouds after exceeding NVIDIA's training performance targets under real-world conditions.
