
Introduction
Pick the wrong Azure VM family and you pay twice: once in hourly rates, and again in performance problems. Families bundle compute, memory, storage, and networking in different ratios, so the choice shapes latency, availability, and cloud spend.
Choosing by vCPU count alone is the classic mistake. You can end up with a memory-starved database, a disk-throttled application server, or a burstable VM that runs out of credits.
Over-provisioning to avoid those problems is no better. Flexera's 2026 State of the Cloud report estimates that 29% of cloud spending is wasted.
This guide explains the main Azure VM families, maps them to workload patterns, and gives you a practical framework for evaluating performance, cost, availability, and growth.
Key Takeaways
- Choose the family first, then the series, generation, and size.
- B- and D-series cover burstable and balanced workloads; E- and M-series cover memory-heavy ones.
- F-, L-, N-, and H-series target compute, local storage, GPU, and HPC needs.
- Weigh CPU, memory, disk limits, network, region, pricing, and measured behavior together.
- Validate with monitoring data after deployment, then keep reviewing.
What Are Azure VM Families?
An Azure VM family is a group of predefined configurations built around a particular balance of vCPUs, memory, storage capability, and networking. The terms nest like this:
- Family: the broad workload-oriented group, such as general purpose or memory optimized
- Series: sizes sharing similar hardware and features, such as Dv5 or Ev5
- Size: a specific configuration, such as
Standard_D8ds_v5 - Instance: the individual VM you deploy
One naming trap: the v5 in a series name is a series version. It is separate from Generation 1 (BIOS boot) versus Generation 2 (UEFI) VMs.
What to compare
Compare these practical characteristics:
- CPU performance and memory per vCPU
- Managed-disk and local-storage capability
- Data-disk limits, IOPS, and throughput caps
- Network bandwidth and GPU support
- Regional availability
Check current limits in Microsoft's documentation, not static third-party lists.
For a concrete sense of scale, Standard_D8ds_v5 offers 8 vCPUs, 32 GiB RAM, up to 16 data disks, 12,800 uncached IOPS, 12,500 Mbps of network bandwidth, and a 300 GiB temporary disk.
Common family groups
General purpose (A, B, D). B-series is burstable. Below its baseline CPU, a VM banks credits; above it, it spends them. When credits run out, the VM is throttled to baseline.
D-series is the balanced choice for production web and application servers. Av1 (Basic and Standard A) retired on August 31, 2024, so check Microsoft's lifecycle pages before considering any A-series size.
Memory optimized (E, M). E-series gives 8 GiB per vCPU versus 4 for D-series. D8_v5 has 8 vCPUs and 32 GiB; E8_v5 has 8 vCPUs and 64 GiB. Same vCPUs, double the memory, which is why databases, caches, analytics, and Azure Virtual Desktop (AVD) session hosts often land on E.
M-series scales from 128 GiB to 30,400 GiB for very large in-memory systems.
Compute and storage optimized (F, L). F-series suits CPU-bound processing. Newer Falsv6 sizes give full-core vCPUs without SMT.
L-series (Lsv3) pairs 8 GiB per vCPU with local NVMe, one 1.92 TB drive per 8 vCPUs, for replicated NoSQL stores like Cassandra and MongoDB. Local NVMe does not remove the need to pick the right managed-disk tier for persistent data.
Specialized (N, H). N-series adds GPUs: NC for AI/ML and rendering, ND for deep learning, NV for graphics and visualization. H-series (HB, HC, HX) targets HPC and simulation.

Both need stronger justification because of cost, capacity, driver requirements, and lifecycle changes. HC and HBv2 retire May 31, 2027.
Why the right family pays off
A good match shows up in application latency, query performance, batch completion time, user density, and total infrastructure cost. The goal is performance headroom without paying for capacity nobody uses.
What to Consider When Choosing the Right Azure VM Family
There is no universal ranking. The best family depends on workload behavior, operating system, performance targets, data-access patterns, availability needs, and budget. Follow a simple process:
- Baseline the current or expected workload.
- Identify the dominant bottleneck.
- Shortlist one or two families.
- Test under representative load.
- Review actual performance and cost after deployment.
Workload profile and utilization pattern
Is the workload steady, bursty, CPU-bound, memory-bound, I/O-heavy, GPU-dependent, or short-lived? B-series fits low-baseline workloads with occasional spikes, such as dev/test or small internal tools. Sustained production load belongs on D, E, or a specialized family, where performance is consistent.
CPU-to-memory balance
Compare CPU demand with working-set size, cache needs, and concurrency. D-series suits balanced application tiers. E or M suits databases and in-memory platforms. Use real telemetry and benchmarks, not nominal specs.
Storage performance and capacity
VMs and disks each impose IOPS and throughput limits, and the lower one wins. Microsoft's own example: three P30 disks provide 5,000 IOPS each (15,000 combined), but a Standard_D8s_v3 caps uncached I/O at 12,800, per Microsoft's disk guidance. A VM with plenty of vCPU and memory can still stall on storage.

Evaluate these separately from VM capacity:
- Disk type: Ultra Disk, Premium SSD v2, Premium SSD, Standard SSD, or Standard HDD
- Disk count and performance tier
- Temporary storage, which is not persistent and can be lost on redeployment or stop
- Backup requirements
Network, scaling, and availability
Check inbound, outbound, and east-west traffic, load balancing, scale-set design, and availability zones. Confirm your preferred family and generation exist in your required US region (for example East US 2) and your subscription. az vm list-skus --location eastus2 --zone filters by region and zone, and --all includes sizes your subscription cannot use. Accelerated Networking cuts latency and jitter, but support varies by size and OS.
Operating system and software constraints
Confirm Windows or Linux compatibility, licensing, processor architecture (x86-64 versus Arm64 options such as Bpsv2), application certification, GPU drivers, and InfiniBand/RDMA drivers for HPC. A cheaper family is a bad deal if it creates licensing or compatibility risk.
Cost and pricing model
Compare compute price against storage, networking, licensing, monitoring, backup, and scaling costs. Match the model to the workload:
- Pay-As-You-Go: unpredictable or short-lived demand
- Reservations: one- or three-year commitments to a specific configuration
- Savings Plans: a fixed hourly spend across eligible services and regions
- Spot VMs: interruptible, no SLA, 30 seconds' eviction notice, and not available for B-series
- Azure Hybrid Benefit: covers the Windows Server license (with qualifying Software Assurance), though base compute is still billed
Always pull current regional prices from Azure's pricing pages.
Validate and revisit
Use Azure Monitor for host-level CPU, network, and disk data, and add the Azure Monitor Agent for guest metrics. Azure Load Testing helps expose bottlenecks before launch. Azure Advisor's resize recommendations look at seven days of data by default and target P95 CPU and network at or below 40% and P99 memory at or below 60% for user-facing workloads.
Set a review cycle tied to application releases, seasonal demand, and sustained utilization changes. The first choice is not permanent.
How Lucidity Can Help
Lucidity is a cloud block storage management platform. It complements, rather than replaces, workload profiling and Microsoft's VM-family documentation. It covers the storage side of your Azure deployments.
- Visibility: real-time insight into every volume, so you can connect disk utilization and performance to VM and managed-disk decisions
- Lumen: evidence-backed tiering recommendations (for example, Premium SSD to Standard SSD) based on IOPS, throughput, latency, and cost
- Idle disk detection: flags unattached, reserved, unmounted, and zero-I/O disks
- AutoScaler: expands, shrinks, and optimizes disks with zero downtime and no infrastructure or application code changes
- Multi-cloud and governance: works across Azure, AWS, and Google Cloud, is application-agnostic, and is SOC 2 compliant and FinOps certified
Lucidity's free, self-serve Assessment scans environments in about five minutes and supports Azure. It is a quick way to find storage waste and performance issues in your estate.
Conclusion
Choosing an Azure VM family starts with workload evidence, not the largest size or the lowest hourly rate. The aim is to align CPU, memory, storage, network, specialized hardware, availability, and pricing with what the application actually needs.
Those needs shift as traffic, data volume, and software versions change. Keep reviewing telemetry, costs, and disk utilization, and adjust when the data supports it.
Frequently Asked Questions
What are the different types of Azure VMs?
Microsoft groups sizes into general purpose (including burstable B), compute optimized, memory optimized, storage optimized, GPU accelerated, FPGA accelerated, and high-performance compute. Each group contains multiple series and sizes.
What is Azure VM used for?
Azure VMs provide cloud-hosted compute for web and application servers, databases, development environments, analytics, virtual desktops, batch jobs, AI training, and HPC simulations.
What are Azure B series VMs?
B-series VMs are burstable instances for workloads with a low baseline CPU need and occasional spikes. They earn credits below baseline and spend them above it. When credits run out, the CPU drops to baseline, so sustained high load needs another family.
What are the key differences between Azure D-series and E-series virtual machines?
D-series balances compute and memory (4 GiB per vCPU) for general workloads. E-series offers 8 GiB per vCPU for databases, caches, and analytics. Compare storage, network, and price only for the specific sizes you are evaluating.
What are the key differences between VDI and AVD?
VDI delivers desktops from a cloud server or datacenter. Azure Virtual Desktop is Microsoft's managed service for desktops and RemoteApps on Azure. Pilot session-host sizing (up to 6 light or 1 power user per vCPU), storage, identity, and scaling first.