Assessment & discovery
Capture business goals, workloads, data sensitivity, current platforms, operating model, and deployment constraints.
Private GPU infrastructure, designed on NVIDIA reference architecture
From facilities readiness and cluster design through Kubernetes, GPU Operator, vLLM, monitoring, and handover — sized on NVIDIA best practice from a single Hopper or Blackwell node to a multi-rack SuperPOD, with a clear expansion path toward Rubin.
Rack, fabric, and platform — one design from node to SuperPOD
The solution is layered so a single validated rack and a multi-rack SuperPOD share the same facility, fabric, and platform design — only the scale changes.
| Layer | What it covers |
|---|---|
| Facility layer | Rack space, power distribution, direct liquid cooling, cabling, and physical security sized for high-density Hopper, Blackwell, or Rubin racks. |
| Compute layer | GPU nodes or NVL72 rack systems with GPU, CPU, and memory sized per workload, forming one NVLink domain per rack (8 to 72 GPUs). |
| Fabric layer | Non-blocking InfiniBand or Spectrum-X Ethernet, BlueField DPUs, and SuperNICs linking racks into one low-latency, multi-rack GPU domain. |
| Platform layer | Kubernetes or RKE2, NVIDIA GPU Operator, MIG, and topology-aware scheduling so jobs place along the fastest NVLink and fabric paths. |
| Serving layer | vLLM and model-serving endpoints for training, fine-tuning, and inference workloads, exposed through governed APIs and ingress. |
| Management & security layer | Prometheus, Grafana, DCGM metrics, IAM/RBAC, backup, audit logging, and the runbooks used at operational handover. |
From assessment to operational handover
Workload sizing
Estimate GPU, CPU, memory, storage, network, and availability requirements for training, inference, VDI, HPC, or mixed workloads.
GPU architecture design
Design the GPU cluster topology, node profiles, isolation model, capacity plan, and expansion path.
Procurement support
Support bill of materials, vendor coordination, delivery planning, and commercial alignment for the selected architecture.
Sized on NVIDIA generations, from Hopper to Rubin
Architecture decisions follow NVIDIA reference designs for each GPU generation, so today’s node design has a validated path to rack-scale and multi-rack SuperPOD deployment instead of a rebuild.
| Generation | Representative systems | Scale-out unit | Best-fit deployment |
|---|---|---|---|
| Hopper (2022) | H100 / H200 SXM, 4th-gen NVLink | Single node, 8-GPU NVLink domain | Departmental training, fine-tuning, and high-throughput inference |
| Blackwell (2024) | B200, GB200 / GB300 NVL72, 5th-gen NVLink | NVL72 rack, 72-GPU NVLink domain | Rack-scale generative AI and frontier-scale training |
| Rubin (2026) | Vera Rubin NVL72, 6th-gen NVLink, ConnectX-9, BlueField-4 | NVL72 rack, scale-out via Spectrum-X / InfiniBand | Agentic AI, reasoning, and long-context inference at SuperPOD scale |
Data center, network, and storage readiness
Data center readiness
Review rack space, power, cooling, cabling, physical security, and operational constraints for high-density GPU systems.
Network design
Plan management, storage, cluster, inference, ingress, and east-west traffic patterns with adequate throughput and segmentation.
Storage design
Define object, file, block, local, and high-throughput storage patterns for datasets, checkpoints, models, vectors, and backups.
A phased path from one rack to a SuperPOD
High-end Hopper, Blackwell, and Rubin deployments are validated in phases against NVIDIA reference architecture, so scale-out stays predictable as GPU count grows from one rack to many.
Rack-scale power and liquid cooling
Plan direct liquid cooling, coolant distribution, redundant power feeds, and rack density for high-power NVLink rack systems such as NVL72, ahead of delivery.
NVLink domain sizing
Size the GPU-to-GPU NVLink domain per rack, then design the scale-out fabric that links multiple racks into one addressable SuperPOD.
Non-blocking fabric design
Design leaf-spine InfiniBand or Spectrum-X Ethernet fabrics with BlueField DPUs and SuperNICs for non-blocking, low-latency, multi-rack GPU communication.
Topology-aware scheduling
Configure Kubernetes, GPU Operator, MIG, and NCCL topology awareness so training and inference jobs place workloads along the fastest NVLink and fabric paths.
Reference-architecture validation
Validate delivered racks against NVIDIA DGX SuperPOD / DGX BasePOD reference designs with burn-in, NCCL bandwidth tests, and acceptance benchmarks before handover.
Phased capacity buildout
Start with a single validated rack or proof-of-concept cluster, then expand rack by rack toward a multi-rack SuperPOD as workload demand and budget allow.
GPU software stack and operations
Kubernetes / RKE2 deployment
Deploy a production-oriented Kubernetes foundation with namespaces, RBAC, ingress, storage classes, and platform policies.
NVIDIA GPU Operator
Install and validate GPU drivers, device plugin, DCGM integration, MIG support where applicable, and GPU scheduling.
vLLM / LLM deployment
Deploy inference endpoints for Qwen, Llama, or selected models using GPU-backed serving patterns.
Monitoring stack
Implement Prometheus, Grafana, DCGM metrics, alerting, dashboards, and capacity indicators.
Handover and operations guide
Provide runbooks, operational guide, knowledge transfer, and platform validation evidence for customer teams.
Frequently asked questions
When is on-premises GPU infrastructure appropriate?
On-premises infrastructure is appropriate for persistent utilization, dedicated control, private data paths, facility-specific requirements, or internal policies that require owned environments.
What is assessed before deployment?
Assessment covers workloads, facilities, rack density, power, cooling, cabling, storage, networking, security, Kubernetes, operations, support model, and expansion path.
Which NVIDIA GPU generation should we plan for — Hopper, Blackwell, or Rubin?
Hopper (H100/H200) fits departmental training and inference on a single node. Blackwell (GB200/GB300 NVL72) is built for rack-scale generative AI with a 72-GPU NVLink domain per rack. Rubin (Vera Rubin NVL72) extends that design with 6th-generation NVLink and faster scale-out networking for agentic AI and long-context inference. The right starting generation depends on workload, budget, and facility readiness — architecture design should still leave a validated path to the next generation.
Can a GPU On-Prem deployment scale into a SuperPOD?
Yes. Deployments are designed against NVIDIA reference architecture from day one — NVLink domain sizing, non-blocking InfiniBand or Spectrum-X fabric, liquid cooling, and topology-aware Kubernetes scheduling — so capacity can expand rack by rack from a single validated cluster to a multi-rack SuperPOD without re-architecting.
Can Indosat Business help with handover?
Yes. Handover can include validation evidence, architecture records, runbooks, monitoring dashboards, access model, and knowledge transfer.