InfiniBand networking
We've partnered with NVIDIA to design and deploy a SHARP-enabled InfiniBand network that provides faster and highly performant multi-node interconnect.
With up to 3200Gbps of one-to-one non-blocking interconnect, your teams can get GPUs communicating at massive scale with sub-millisecond latency. That unlocks higher performance from GPUs and accelerates training time.
Storage
At CoreWeave, we built our storage services to help enable enhanced performance from GPU clusters.
Feed data into your GPUs and handle massive datasets with reliability and ease, accelerating time-to-train.
Customers can utilize our AI Object Storage services with Local Object Transport Accelerator (LOTA) or leverage Dedicated Storage Clusters. LOTA gives your teams up to 2GB/s per GPU read speeds, while Dedicated Storage Clusters support storage backends of your choice.
Plus, CoreWeave Storage helps your teams recover quickly from job interruptions. With fast checkpointing and recovery of intermediate results, your teams can quickly pick up their training jobs close to where they left off.
CoreWeave Kubernetes Service (CKS)
CoreWeave Kubernetes Services delivers an AI-optimized managed Kubernetes environment with a focus on performance, efficiency, scale and ease of use.
With CKS, your teams can get the benefits of bare metal performance within the flexibility of the cloud. We've eliminated the hypervisor layer completely, enabling your teams to operate directly on bare metal nodes. This helps ensure optimal performance, reduced latency, and quicker time to market.
Plus, CKS grants heightened visibility into cluster health and performance—down to the bare metal node performance. Nip interruptions in the bud and catch early warning signs to keep training jobs on track.
SUNK
We built Slurm on Kubernetes (SUNK) to combine the benefits of Slurm’s job scheduling with Kubernetes’ orchestrating services. With SUNK, your teams can run training jobs with the combined flexibility of Kubernetes and familiarity of Slurm for a superior experience.
Observability
Our observability platform provides visibility into essential cluster metrics, allowing your teams to efficiently monitor nodes and quickly identify the root cause of any interruptions. That means not only recovering from jobs quickly but also preventing them before they even happen.
This helps enable continuous high performance and minimizes downtime. That means more time spent training and less time spent firefighting or handling interruptions and issues.
Mission Control
Our Mission Control service helps enable enhanced cluster health management, providing your teams with more resilient and reliable AI infrastructure.
Mission Control also helps keep nodes at peak performance with two essential features: Node Lifecycle Controller and Fleet Lifecycle Controller.
When issues arise, our Node Lifecycle Controller swiftly replaces unhealthy nodes, reducing the frequency, duration, and cost of interruptions. Meanwhile, Fleet Lifecycle Controller helps ensure node health from deployment throughout its entire lifecycle.
