// Platform

The managed stack for GPU workloads.

Every layer — orchestration, storage, networking, observability — tuned for training-scale performance.

// Deep dive

Every layer, in detail.

01 // Orchestration

Kubernetes and Slurm, side-by-side.

Run production inference on managed Kubernetes and long-horizon training on Slurm — from the same control plane, on the same fabric.

  • Managed control plane with CRDs for GPU jobs and topologies
  • Slurm with pyxis + enroot for container-native training
  • Bring-your-own scheduler via Kueue / Volcano / Ray
  • Multi-tenant namespaces with per-team quotas
Scheduler
K8s 1.30 · Slurm 24.05
Container
OCI · enroot · pyxis
Autoscaling
Cluster + workload
02 // Scheduling

Topology-aware, gang-scheduled by default.

Jobs land on GPUs that share an NVLink domain and rack — no fragmented placements, no silent perf cliffs at scale.

  • Gang scheduling for tightly-coupled MPI / NCCL jobs
  • Topology hints for NVLink islands and IB leafs
  • Preemption tiers with checkpoint-aware eviction
  • Backfill for short jobs against reserved capacity
Placement
NVLink + rack-aware
Preemption
Checkpoint-aware
Queueing
Priority + fair-share
03 // Observability

Per-GPU truth, per-job lineage.

DCGM metrics, NCCL traces, and job lineage streamed to Grafana and your OTLP endpoint — nothing gated behind support tickets.

  • DCGM per-GPU utilization, ECC, thermals, power
  • NCCL and IB counters correlated to job IDs
  • OpenTelemetry export to your stack
  • Prebuilt Grafana dashboards per workload
Metrics
DCGM · NCCL · IB
Export
OTLP · Prometheus
Retention
13 months hot
04 // Storage

Parallel NVMe where training data lives.

Lustre and WEKA parallel filesystems co-located with the compute fabric — plus S3-compatible object storage and a cold tier for retention.

  • Lustre / WEKA at 400 GB/s aggregate per cluster
  • S3-compatible object store with signed URLs
  • Cold tier for datasets and checkpoints at rest
  • Same VPC, same fabric — zero egress inside cluster
Parallel FS
400 GB/s / cluster
Object
S3-compatible
Egress in-cluster
Free
05 // Networking

Non-blocking InfiniBand, NVLink islands.

3.2 Tbps non-blocking InfiniBand between nodes, NVLink domains inside them — the fabric collectives assume, delivered.

  • 3.2 Tbps NDR InfiniBand, non-blocking
  • NVLink / NVSwitch islands up to 72 GPUs (GB200)
  • SR-IOV, GPUDirect RDMA, GPUDirect Storage
  • Private VPC peering and dedicated egress tiers
Fabric
NDR IB · 3.2 Tbps
NVLink domain
up to 72 GPUs
RDMA
GPUDirect · GDS
06 // Security

Isolated tenancy, encrypted end-to-end.

Single-tenant nodes, private endpoints, and full-fabric encryption — with SOC 2 Type II and ISO 27001 controls behind them.

  • Single-tenant physical nodes by default
  • AES-256 at rest, TLS 1.3 in transit, mTLS optional
  • Private endpoints and VPC peering, no public IPs required
  • SOC 2 Type II · ISO 27001 · GDPR · DPDP
Tenancy
Single-tenant nodes
Encryption
AES-256 · TLS 1.3
Compliance
SOC 2 · ISO 27001
// Architecture

How it fits together.

  ┌──────────────────────────────────────────────────────────────┐
  │                       Control Plane                          │
  │        Scheduler · Observability · Identity · Billing        │
  └───────────────────────────┬──────────────────────────────────┘
                              │
  ┌───────────────────────────┴──────────────────────────────────┐
  │                     Compute Fabric                           │
  │   NVIDIA H100 · H200 · GB200  ·  NVLink Islands              │
  │   ────────────────────────────────────────────────────────   │
  │           Non-blocking InfiniBand (3.2 Tbps)                 │
  │   ────────────────────────────────────────────────────────   │
  │         Parallel NVMe · Object Storage · Cold Tier           │
  └──────────────────────────────────────────────────────────────┘