A Leading Hyperscale Platform
Distributed GPU Cluster Interconnection
Client Background
A major internet and AI platform was scaling its GPU compute infrastructure across multiple geographies to support large-scale model training and real-time inference workloads. The explosive growth in AI data volumes meant that concentrating all GPU capacity in a single data centre was no longer viable — the organisation needed to distribute clusters across multiple facilities for resilience, power availability and regulatory reasons, while ensuring that distributed nodes operated as a single, unified compute fabric with minimal inter-site latency.
Key Challenges
- GPU clusters distributed across 4 data centres in different markets needed to behave as one logical compute domain
- Large-scale model training required deterministic, high-throughput inter-site data transport with minimal jitter
- Inference workloads demanded consistent sub-millisecond response times between frontend and backend GPU pools
- Existing public internet and commodity transit paths introduced unacceptable latency variance and packet loss
- Rapid capacity scaling was needed as new GPU racks were deployed quarter by quarter
StarCloud Solution
StarCloud designed and deployed a purpose-built, high-capacity DCI fabric connecting the client's four data centre sites via dedicated wavelength and dark-fibre paths. The architecture employed diverse physical routes for resilience and was engineered to deliver the deterministic latency and ultra-low jitter required for GPU-to-GPU communication across sites — effectively making the distributed clusters appear as a single, co-located compute environment to the training and inference orchestration layer.
Results & Outcomes
The distributed GPU clusters now operate as a single logical compute domain, enabling the platform to run large-scale AI training jobs across sites with no perceptible performance penalty compared to co-located infrastructure. The deterministic, high-availability DCI fabric also supports the platform's CDN and content delivery workloads, while the scalable architecture allows new GPU capacity to be onboarded rapidly as the platform's AI ambitions grow.
Ready to Build Your Solution?
Discover how StarCloud can deliver similar results for your organisation.