NVIDIA HGX Rubin NVL8 on CoreWeave

Serve a fleet of agents from one node.

• Keep long-running agent sessions in memory with 2.3 TB HBM4 memory
• Run tensor-parallel serving inside the node at 3.6 TB/s per GPU over 6th Gen NVLink
• Isolate tool calls and RL environments next to your models (CoreWeave Sandboxes)
• Trace every step, tool call, and decision (CoreWeave Agent Lens)

Run complex simulations and AI models on the same node.

• Double-precision performance on single-node architecture
• Run your existing Slurm scripts on Kubernetes (SUNK)
• Automate node and InfiniBand health checks before jobs start (CoreWeave Mission Control)
• Find the one slow GPU holding back a job Swith straggler detection (CoreWeave Mission Control)

Move massive datasets without waiting on storage.

• Process large ETL and dataframe jobs with 176 TB/s of memory bandwidth per node
• Feed every GPU from a single global dataset at uUp to 7 GB/s (CoreWeave AI Object Storage)
• Keep multi-node pipelines moving with topology-aware scheduling (CKS and SUNK)
• Spot slow GPUs and unhealthy nodes before they stall a pipeline (CoreWeave Mission Control)

Adopting new compute without the right support means your team absorbs the learning curve. From NVIDIA Ada Lovelace to Hopper to Blackwell, CoreWeave has operationalized major NVIDIA platforms, so your team hits the ground running.

NVIDIA Vera Rubin sets a new bar for performance. CoreWeave is engineered for efficiency and reliability at scale, so your workloads realize every bit of performance possible.

Production AI workloads can break in ways you can’t always predict. CoreWeave experts help you manage that operational complexity at factory scale, so your largest clusters keep running when it matters most.