Nodes in groups
Group hosts by data center, rack, or purpose. Connect a node and its GPUs report in automatically.
Real-time health
Per-GPU utilization, VRAM usage, and temperature across the fleet, refreshed continuously.
No oversubscription
A per-GPU budget guarantees the sum of active slices never exceeds the physical VRAM.
Features
Built to run a GPU fleet
Fine-grained VRAM slicing
Carve a GPU into arbitrary-sized slices (e.g. 80 GB → 40 + 20 + 20) and run isolated workloads side by side.
Node grouping & onboarding
Create groups and connect new hosts; the node agent dials out and reports inventory and health.
Allocation tracking
Every slice records who holds it and what runs on it, backed by an append-only audit trail.
Health & operational alerts
Alerts for offline GPUs, overheating, downed nodes, and exhausted capacity.
Use cases
One fleet, many teams
Consolidate small services
Pack many small inference services onto shared GPUs instead of dedicating one card each.
Internal multi-team sharing
Allocate slices to teams and track usage for transparent internal chargeback.
Maximize expensive cards
Keep H100 and A100 capacity working by filling unused VRAM with additional pods.
Capacity at a glance
See free versus allocated VRAM across every node and GPU in one view.
FAQ
Common questions
How does VRAM slicing work?
Each GPU has a VRAM budget; you carve slices of any size as long as their sum stays within the total. Workloads run with a memory cap per slice. Stronger hardware isolation (MIG/MPS) is on the roadmap.
How do I add a node?
Create or pick a group, then connect the host. Its agent dials out to the control plane and reports the GPUs it sees, along with live health.
Can multiple workloads share one GPU?
Yes — that is the point. Several slices can run on one GPU at once, as long as their combined VRAM stays within the card’s budget.
What happens if a GPU goes offline?
Heartbeats stop, the node is marked offline, and an operational alert is raised so you can act before workloads are affected.