Azure VM Sizing Complete Guide: Series Selection, Naming, and Cost Optimization

Azure VMs are organized by workload type into the major B / D / E / F / G / H / L / M / N series. B: Burstable (CPU-credit based, dev/test, low-load). D: General Purpose (balanced, web servers). E: Memory Optimized (more RAM, databases and caches). F: Compute Optimized (more CPU, batch processing, game servers). G: High Performance Compute (large CPU + memory). H: HPC (MPI workloads). L: Storage Optimized (large local NVMe, NoSQL DBs). M: Memory Optimized (up to 12 TB RAM, SAP HANA). N: GPU (ML and graphics). Each series also has generations (the v5 in D2s_v5 is the generation), and newer generations offer better performance per cost.

The Azure VM size naming convention is 'series + vCPU count + feature letters + generation'. Example: Standard_D2s_v5 = D series, 2 vCPU, s (Premium Storage capable), v5 (5th generation). Feature letters: a (AMD CPU), s (Premium Storage capable, required for many workloads), d (local temp disk), i (isolated, dedicated hardware), p (Arm-based, Ampere Altra), l (low memory). Example: Standard_D4ads_v5 = D series, 4 vCPU, a (AMD), d (local disk), s (Premium Storage), v5. Once you can read the convention, you can judge VM characteristics instantly from the name, which makes sizing decisions much faster.

Standard sizing process: 1) Identify the workload profile (CPU-bound, memory-bound, I/O-bound, GPU-required). 2) Narrow down to a series (CPU-intensive → F, memory-intensive → E, HPC → H, AI/ML → N). 3) Estimate vCPU and memory (for migrations use current spec + 30% headroom, for greenfield start small and scale up). 4) Pick the s option if you need Premium Storage. 5) Apply cost-optimization options (Spot VM, Reserved Instance, Savings Plan, Hybrid Benefit). 6) Compare generations of the same spec and pick the latest. 7) Estimate monthly cost in the Azure Pricing Calculator. The proven approach is to start at the smallest size, watch resource utilization in Azure Monitor/Insights for 1-2 weeks, and resize as needed. Azure VM resize is online (only a few minutes of restart), so building a culture that avoids over-provisioning is critical.

Spot VMs let you use Azure's surplus capacity at up to 90% off, but they can be evicted at any time with only 30 seconds notice. Good use cases: 1) batch processing (genomics, video rendering, ML training), 2) dev/test environments, 3) stateless workloads (temporary scale-out for web delivery), 4) AKS Spot node pools for cost optimization, 5) Azure Batch worker nodes. Watch outs: 1) never use Spot VMs for stateful production workloads (eviction risk), 2) choose Deallocate or Delete as the eviction policy, 3) set the max price (allowing the pay-as-you-go rate reduces eviction frequency), 4) check availability up front with the Spot Placement Score API. AKS Spot node pools use Tolerations-based pod scheduling so you can design an architecture that survives Spot VM evictions.

Reserved Instance (RI): a 1-year or 3-year reservation for a specific VM size, region, and OS, giving up to 72% off pay-as-you-go. Low flexibility (size changes require RI exchange). Savings Plan: a 1-year or 3-year commitment to a fixed hourly spend, applied flexibly across compute services (VM, Container Instances, Functions Premium, App Service, Container Apps), giving up to 65% off. RIs offer a deeper discount, but Savings Plans win on flexibility. Decision: fixed workloads (e.g. a specific SAP HANA configuration) → RI; variable workloads (auto-scale, containers) → Savings Plan. You can also combine them: the canonical cost-optimization pattern is RI for core workloads + Savings Plan for surplus load + Spot for batch, a three-tier model.

Azure Hybrid Benefit lets you bring existing on-prem Windows Server / SQL Server / Linux subscriptions (with Software Assurance) to Azure and waive the equivalent Azure license cost. Windows Server: waives VM Windows license fees (about 40% cheaper than pay-as-you-go), usable on up to two Azure VMs per license. SQL Server: waives license fees on VMs running SQL Server (Standard/Enterprise) and on Azure SQL DB/MI, with each 4 vCore (Standard) or 1 vCore (Enterprise) license mapping to vCores on the Azure side. Linux RHEL / SUSE: bring a Red Hat or SUSE Cloud Access Subscription to switch from Azure marketplace pricing to BYOS (Bring Your Own Subscription) pricing. If your organization already holds licenses, you should always evaluate Hybrid Benefit. Savings are often very large, regularly in the tens of millions of yen per year.

Azure Managed Disk options: 1) Standard HDD (S series): cheapest, low performance, for backup/archive. 2) Standard SSD (E series): mid performance, affordable, fine for web servers. 3) Premium SSD (P series): high performance, the default for production databases. 4) Premium SSD v2: GA in 2022, an evolved Premium SSD where IOPS/throughput are tunable independently of disk size. 5) Ultra Disk: ultra-high IOPS (160,000) and sub-millisecond latency, for SAP HANA, Oracle, and high-performance SQL Server. Premium SSD and Ultra Disk require a Premium Storage capable VM size (with the s suffix). The standard picks today: Premium SSD v2 for production databases (best cost-performance), Ultra Disk for extreme load, and Standard SSD when cost is the top priority.

AZ-104 (Administrator) tests VM sizing, Premium Storage selection, and backup in depth in Domain 3 (Compute, 20-25%). AZ-305 (Solutions Architect Expert) covers compute selection within full architectures in Domain 4. AZ-204 (Developer Associate, note: retiring 2026-07) approaches the topic from a developer perspective, comparing VMs to App Service and similar PaaS. AZ-140 (AVD Specialty) covers VM selection for AVD session hosts. AZ-120 (SAP on Azure) covers the M series and HANA Large Instances. In real-world work, VM sizing is a central topic in Azure cost optimization and an essential skill for every Azure engineer.