Sharing, scheduling, and migration — and how the same primitives compose into single-cluster, multi-cluster, or hybrid-cloud topologies. Built on Kubernetes, the NVIDIA Device Plugin, and InfiniBand/RDMA.
Real-time DL execution detection. Priority + deadline queues. Proactive memory control prevents OOM under contention.
Quota and partition assignment. RBAC-aware. Per-project, per-user limits. Audit-logged.
Zero-checkpoint live migration. RDMA-direct on-prem; encrypted TCP for cloud. Sub-second handoff.
Whatever environment your researchers use, CNLab routes GPU allocations transparently. Your code calls .cuda(); we handle the rest.
Notebook server runs on the closest scheduled slice. CUDA calls intercepted; allocations routed to MIG or 1% blocks based on profile.
Remote runtime mounts your workspace, attaches GPU, forwards SSH and secrets. Compatible with Cursor, Windsurf, JetBrains.
cnlab run train.py --gpu h100-25. Streams stdout, captures artifacts, releases the GPU on exit.
From access request to job completion, every workload flows through six stages. Each stage is observable, audited, and policy-bound.
JWT + RBAC. Per-user quota.
Daemons on every GPU node.
Priority + deadline queue.
Idempotent, audited.
Best-fit cluster selection.
MIG + 1% blocks at run.
The average enterprise GPU sits at 5% utilization. CNLab's two-tier slicing pushes it to 95%+ — without rewriting any user code.
Up to 7 hardware-isolated instances per H100. Full memory bandwidth within each. Use for tenant boundaries that must survive driver bugs.
100 simultaneous tenants per GPU. Resize without restart. Combines with MIG: a 1g MIG slice can host 14 1% blocks.
Per-tenant VRAM quotas. cgroup + namespace isolation on the host. Cross-tenant CUDA IPC denied.
GPU utilization: 5% → 95%+
A single scheduler with policies you can layer: deadline, priority, fair-share, and bin-packing. Pre-emptive when needed; backfill-friendly always.
Submit jobs with target completion. Scheduler reserves slices forward in time and pre-empts lower-priority work to meet the deadline.
Long-term per-team usage normalised. A team that ran heavy yesterday yields ground to a starved team today.
Where deadlines and shares allow, scheduler packs work onto the fewest nodes. Idle nodes power down for energy savings.
Production inference pre-empts development training when capacity tightens. Pre-empted work resumes from the latest auto-checkpoint.
Watches per-tenant working set; resizes 1% block VRAM allocation in flight to prevent OOM crashes.
Different clusters can run different scheduling policies. Federation respects each cluster's local rules.
When local capacity tightens, CNLab provisions cloud GPUs in seconds, migrates the workload, and returns it when local frees up — without the user noticing.
Cloud cost: 95% → 28% of cloud-only baseline
Scheduler forecasts saturation 90s ahead based on queue depth and historical patterns. Cloud provisioning starts before the local queue stalls.
Same migration engine as cluster-to-cluster. Memory pages travel over TLS-encrypted, accelerated TCP. Sub-second pause; resume from exact instruction.
When local utilization drops, scheduler reclaims by migrating the cloud workload back. Cloud node terminates within 30 seconds.
| Dimension | On-Premise Single | Multi On-Premise | Hybrid Cloud |
|---|---|---|---|
| GPU pool size | 4–32 | 32–256 | 128+ + cloud |
| Initial cost | Lowest | Medium | Medium + opex |
| TCO vs cloud-only | −40% | −45% | −50%+ |
| Failover | Single cluster | Cross-cluster | Cross-cluster + cloud |
| Data residency | 100% on-prem | 100% on-prem | On-prem first |
| Setup time | Days | Weeks | Weeks |
AWS · GCP · Azure · Naver Cloud · KT Cloud — first-class. OCI & on-prem-only via API.
CNLab installs as a K8s operator (CRD-driven). Compatible with K8s 1.27+, OpenShift 4.13+, Rancher.
PyTorch · TensorFlow · JAX · Mojo · NVIDIA NeMo · Megatron — all run unchanged.
NVIDIA H200/H100/A100/L40S/L4/RTX 6000 · AMD MI300/MI250 · roadmap: Intel Gaudi 3.
InfiniBand HDR/NDR · RoCEv2 · standard 100/400G Ethernet · WireGuard for cloud egress.
Prometheus + Grafana out of the box · OpenTelemetry traces · Loki logs · DCGM exporter for GPU metrics.
68-page whitepaper. Details on slicing, scheduling, migration, and integration. Available immediately — no email gate.