Supported Cloud Matrix
| Cloud | Status |
|---|---|
| GCP | Full |
| Azure | Full |
| OCI | Best effort |
| Lambda | Best effort |
| on-prem bare metal | Full |
Per-cloud Notes
AWS
- p4d / p4de / p5 / p5e instance types.
- EFA + libfabric best-effort (rely on TCP retransmit eBPF).
- IAM auth via SPIFFE trust.
GCP
- a3-highgpu / a3-megagpu.
- VPC-native + multi-NIC, IB is exposed via NCCL.
Azure
- ND H100 v5.
- NDm A100 v4.
- Accelerated networking via MANA.
OCI / Lambda
- BM.GPU.H100.8 / H200.8.
- Best effort.
on-prem
- The control plane supports air-gapped install on bare metal.
- HSM-rooted KMS via PKCS#11.
Kubernetes Operators on Cloud
We integrate with:
- Karpenter (AWS / GCP)
- Cluster Autoscaler (legacy)
- Volcano (on-prem)
- KubeRay (all)
Test Coverage
Each cloud provider ships an integration test that uses a 4-GPU spot instance + a CP-VM in the same VPC to verify the round trip.