RESEARCH CLUSTER
Example reference architecture for universities and research groups.
- Multi-user scheduling with Slurm or Kubernetes
- Shared high-performance storage
- Quota and project-based allocation
- Mixed training and experimentation workloads
02 Build · Infrastructure You Own
GPU infrastructure built around your workload.
Peregrine designs, specifies, sources, integrates and commissions GPU infrastructure that your organisation owns — then hands it over, or operates it for you.
Lifecycle
From discovery to handover.
A fixed-price project with defined stages. Each stage produces documentation your team keeps.
DISCOVERY
Workloads, constraints, site
ARCHITECTURE
Compute, network, storage
PROCUREMENT
Specify and source
COMMISSIONING
Rack, power, cabling
PLATFORM INSTALLATION
Drivers, CUDA, K8s / Slurm
TESTING
Burn-in and validation
HANDOVER
Documentation or Operate
Deployment options
Where the infrastructure lives is your decision.
YOUR PREMISES
Infrastructure installed in your own facility where power, cooling and physical security requirements can be met.
AUSTRALIAN DATA CENTRE
Peregrine can arrange hosting in Australian data-centre facilities suited to GPU power densities.
COLOCATION
Your hardware in a colocation facility you select or we recommend, with Peregrine handling integration.
HYBRID
Owned infrastructure for steady workloads, complemented by Peregrine capacity or public cloud for peaks.
Potential components
Specified per workload, not from a catalogue.
Architecture is determined per workload. The components below are potential elements of a build — not a mandatory list — and are selected once requirements are understood.
Peregrine designs, specifies, sources, integrates and commissions infrastructure from established hardware vendors.
- NVIDIA GPUs
- CPU
- RAM
- NVMe storage
- Shared storage
- High-speed Ethernet
- InfiniBand where required
- Rack infrastructure
- Power
- Cooling requirements
- Kubernetes
- Slurm
- Drivers
- CUDA
- Monitoring
GPU Platforms
Built on NVIDIA Blackwell and Hopper systems.
Blackwell Ultra
NVIDIA HGX B300
Large-scale training and high-throughput inference for reasoning models.
View platform
Blackwell
NVIDIA HGX B200
Balanced Blackwell platform for training and production inference.
View platform
Blackwell · PCIe
NVIDIA RTX PRO 6000 Blackwell Server Edition
Cost-effective per-GPU memory for inference fleets, vision, rendering and digital twins.
View platform
Hopper
NVIDIA HGX H200
Mature ecosystem with larger memory for inference of large models and training.
View platformExample reference architectures
Illustrative, not inventory.
These examples show how components can combine for common objectives. Actual node counts, accelerators and interconnect are determined by your Compute Blueprint.
INFERENCE CLUSTER
Example reference architecture for serving production models.
- GPU nodes sized for model memory footprint
- Model serving with autoscaling endpoints
- Low-latency networking to applications
- Observability for latency and utilisation
TRAINING CLUSTER
Example reference architecture for large training runs.
- Tightly coupled GPU nodes with high-speed interconnect
- Parallel file system or fast shared storage
- Checkpointing and job orchestration
- Power and cooling planned for sustained load
How it works
Five steps, one accountable partner.
ASSESS
Understand workloads and constraints
DESIGN
Architecture and specification
DEPLOY
Source, install, commission
CONNECT
Networking, storage, platform
SCALE
Expand or add capacity
- Commercial model
- Fixed-price project
- At completion
- HANDOVER or PEREGRINE OPERATE
- You own
- Hardware, software configuration, documentation
- Optional
- Managed AI Platform (Operate)
Discuss your GPU build.
Share your workload, site and timeframe. We will recommend a deployment option and scope a fixed-price project.