From A10 to M60: An Architect's Journey into Azure GPU VM Sizing for Kubernetes Inference Workloads

Short summary

An architect details how a regional constraint forced a deep dive into Azure GPU VM families for a PyTorch-based Visual Element Detection service on AKS. The article decodes Azure's NV/NC/ND naming scheme and explains why NV-series M60 VMs can be a cost-effective substitute for A10 GPUs for lightweight inference. Key takeaway: always start GPU sizing from workload characteristics—model size, throughput, concurrency—not from the VM catalog.

• •Decodes Azure GPU VM naming conventions (NV/NC/ND families and suffix modifiers)
• •Explains GPU partitioning on NVads_A10_v5 and when M60 VMs are a viable fallback
• •Emphasizes workload-first GPU sizing: model size, throughput, and K8s scaling strategy

Generated with AI, which can make mistakes.