Juniper AI Data Center Deployment Services Datasheet

Data centers built to run AI workloads have requirements distinctly different from non-AI data centers. If network inefficiencies delay job completion times (JCTs), the time and resource costs to train your AI models multiply. Juniper AI Data Center (AIDC) Deployment Services help you avoid that problem.

By delivering leading-edge deployment and production optimization expertise and assistance, our AIDC Deployment Services help you achieve enhanced network performance, smoother operations, and unlock the full potential of your AI workloads. We tailor these services to your specific AI training model to accelerate time to value and maximize ROI.

Juniper AIDC Deployment Services provide a turnkey solution to deploy, monitor, and optimize key network performance metrics for your AIDC implementation. The services are based on Juniper Validated Designs (JVDs) for AIDC deployments using Ethernet, ensuring top-quality deployments that are compatible with GPU types from NVIDIA (fixed services) and all vendors (custom service). The services include one each of up to three types of greenfield networks: frontend compute, backend compute, and backend storage with Juniper AIDC-qualified devices.

The services have two phases: deployment and production optimization. The deployment phase ends with network acceptance validation followed by a 30-day production optimization phase.

The services are available with or without Apstra. For customers with Juniper Apstra, AI Data Center Deployment with Apstra creates blueprints for all AIDC fabrics and then auto-provisions all devices. This saves time and eliminates potential operator errors during the deployment phase by automating the high-level design (HLD) and low-level design (LLD) processes. AI Data Center Deployment—the standard deployment service without Apstra—requires the manual creation of an HLD and LLD prior to actual deployment.

Both services begin with a workshop run by the Juniper project manager and consultant engineer who will collaborate with the customer to develop a mutual understanding of the overall requirements, deployment plan, and outcomes for the production optimization phase. This leads to the design and actual deployment of the fabrics.

Once all devices are deployed, the GPU network interface cards will be configured and validated to use Ethernet by Juniper. The final step in the deployment phase is acceptance testing and validation, in which Juniper will run continuous collective communications library functions to achieve high bandwidth and low latency across the GPU fabrics.

After the networks have been turned over to the customer, Juniper experts continue to engage for the 30-day production optimization phase. During the 30 days, Juniper closely monitors key parameters and Ethernet tuning through multiple iterations of the customer’s training model cycles, including deployment of trained models into the frontend inference network. This phase provides the customer with the key advanced analytics and final adjustments to Ethernet interfaces that minimize JCT at maximum bandwidth.