Compute & GPU

📅Published 2026-08-25·🔄Updated 2026-08-26·⏱1 min read

This section covers the compute layer that utilizes existing GPU assets as hybrid nodes. It guides the 3-tier architecture combining on-premises GPU, cloud GPU, and Bedrock, and high-performance networking configuration for DGX H200-class equipment.

---

⚡

GPU Workloads & SR-IOV Networking

3-Tier Cascade (on-premises GPU, cloud GPU, Bedrock) architecture, DGX H200 SR-IOV VF configuration, MLNX_OFED and systemd orchestration

🎛️

GPU Scheduling & Cloud Fallback

GPU taint isolation at node registration time, hybrid-only NVIDIA Device Plugin nodeSelector, Karpenter-based cloud GPU fallback NodePool