Models
Deploy AI models selected for the work they need to perform.
WF2040 Private AI 384
WF384
Private inference at enterprise scale
High-capacity private AI for larger organisations.
A dedicated rack platform for multiple departments, multiple models, demanding retrieval-augmented generation (RAG) and agent workloads.
WF384 specification
Recommended configuration.
Sized for your workloads, deployment environment and integration requirements.
Compute & memory
- GPUs
- 4 × NVIDIA RTX PRO 6000 Blackwell Server Edition
- Total GPU memory
- 384 GB GDDR7 ECC, aggregate across four GPUs
- Aggregate GPU memory bandwidth
- Approximately 6.39 TB/s
- Peak FP4 Tensor performance
- Approximately 16 PFLOPS aggregate
- CPU
- High-core-count AMD EPYC platform, with single- or dual-socket configurations
- System RAM
- 1 TB ECC DDR5 standard, configurable to 2 TB+
Storage & infrastructure
- Boot storage
- 2 × enterprise NVMe SSD in RAID 1
- AI / data storage
- 30–60 TB usable enterprise NVMe, configurable upward
- Networking
- 2 × 100 GbE recommended; 25 GbE minimum
- Management
- Dedicated BMC / IPMI / Redfish management interface
- Power
- Redundant high-efficiency PSUs, sized for approximately 2.4 kW of GPU load plus system overhead
- Chassis
- 4U / 5U rack server with front-to-back cooling
- Deployment
- On-premises, colocation or private datacentre
Software, integration & control
- Operating system
- Ubuntu LTS / supported enterprise Linux
- Inference
- vLLM / TensorRT-LLM / Triton, as appropriate
- API
- OpenAI-compatible internal API endpoint
- Identity
- Entra ID / Active Directory / OIDC / SAML integration
- Storage integrations
- SharePoint, file shares, object stores and SQL / data platforms
- Observability
- GPU telemetry, request logging, model metrics, utilisation, latency and audit trail
- Isolation
- NVIDIA MIG where appropriate, containers and departmental access policies
GPU figures aggregate four cards, each rated at 96 GB, 1,597 GB/s and 4 PFLOPS FP4. These totals describe hardware capacity and theoretical peak performance; application throughput depends on the model, software and workload. NVIDIA GPU specifications.
Beyond the hardware
Build the environment
around your work.
Knowledge
Connect approved organisational information to your private AI environment.
Agents
Give AI controlled access to tools, systems and defined business processes.
Management
Monitor, maintain and evolve the platform as models and requirements change.
Private AI
Build capability you control.
Private models. Private knowledge. Dedicated infrastructure.