
Introduction
Picking an Azure VM for SQL Server by grabbing the biggest RAM number or the loudest IOPS figure is a reliable way to overspend and still hit a bottleneck.
SQL Server performance depends on memory capacity, CPU-to-memory ratio, data-file IOPS, log latency, throughput, TempDB behavior, disk limits, and licensing, all at once.
Many teams discover this after migration. The VM looks fine on paper, then throttles because a single managed disk or the VM's own storage ceiling caps performance first.
This guide compares the current Azure VM families (D, E, M, L, and constrained-vCPU sizes) and shows how to size them from measured memory and I/O needs, not generic workload labels.
Key Takeaways
- Match the family to the bottleneck: balanced (D), memory-heavy (E), very large (M), local NVMe (L), license-sensitive (constrained).
- Check VM-level IOPS and throughput caps and each managed disk's limits.
- Baseline memory pressure, peak IOPS, latency, CPU, TempDB, and log activity first.
- Verify specs, prices, and regional availability against Microsoft's current documentation and US pricing pages.
Overview of Azure VMs for SQL Server
An Azure VM is an IaaS compute instance. You control the operating system, SQL Server installation, storage layout, patching, and many high-availability decisions.
That control is why SQL Server on Azure VMs still fits lift-and-shift migrations, legacy applications, and workloads that need SQL Server Agent or linked servers. According to Microsoft's Azure SQL overview, Azure SQL Database is PaaS without guest-OS access, SQL Server Agent, or linked servers. Azure SQL Managed Instance is also PaaS, with service-specific constraints on Agent and linked servers.
The sizing challenge splits in two:
- Memory drives buffer-pool efficiency and query execution.
- IOPS, throughput, and latency determine how fast SQL Server reads data, flushes the log, and serves TempDB.
Best Azure VM Families for SQL Server by Memory and IOPS
Compare each family on the same criteria: memory-to-vCPU ratio, VM-level IOPS and throughput ceilings, compatible disks, local temporary storage, licensing impact, and total cost.
D-series (Ddsv6): the balanced starting point
Ddsv6 suits small-to-medium SQL Server workloads, mixed transactional apps, and dev/test. It runs from D2ds_v6 (2 vCPUs, 8 GiB) to D192ds_v6 (192 vCPUs, 768 GiB), a fixed 4 GiB per vCPU.
- Memory: Microsoft's SQL guidance notes this ratio is below the 8:1 recommended for production SQL workloads. Consistent physical reads or memory-grant waits are your cue to move up.
- IOPS and throughput: D192ds_v6 reaches 260,000 uncached IOPS and 12,000 MB/s with Premium SSD (~400,400 with Ultra or Premium SSD v2). Smaller sizes top out far lower, so match disk performance to the VM ceiling.
- Caveats: Separate data, log, and TempDB. Use ReadOnly host caching on supported data disks and None on logs.
E-series (Edsv6, Ebdsv5): the memory-intensive default
Most Edsv6 sizes give 8 GiB per vCPU, spanning 2 to 192 vCPUs and 16 to 1,832 GiB. That makes E-series the first candidate for memory-heavy OLTP, reporting, and analytics. More RAM per core often helps more than adding CPU.
- Memory: Move from D to E when buffer-cache hit ratio drops below Microsoft's 90% target, Page Life Expectancy collapses, or Memory Grants Pending climbs with moderate CPU.
- IOPS and throughput: Edsv6 series maxima are 360,000 IOPS on Premium SSD and 400,000 on Ultra or v2. They are series maxima, not guarantees for every size.
- Controller matters: E112ibds_v5 (NVMe, 672 GiB) hits 400,000 IOPS with Ultra or v2; E96bds_v5 (SCSI, same RAM) tops out at 180,000.
- Licensing: Bigger VMs raise SQL core costs. Consider constrained-vCPU sizes (below).
M-series: very large memory
Msv3 medium memory spans 12 to 176 vCPUs and 240 to 3,892 GiB. Msv3 High Memory reaches 416 to 832 vCPUs and up to 15,200 GiB. Mbdsv3 covers 128 to 3,892 GiB with stronger storage, up to 650,000 IOPS on Ultra or v2 for M176bds_4_v3.

Use it for very large instances, consolidated estates, or working sets that won't fit in E-series.
- NUMA limit: Microsoft states SQL Server is unsupported with more than 64 virtual cores per NUMA node. Disabling SMT is the documented mitigation.
- Write Accelerator: Supported on Msv3 HM and Mbdsv3, recommended for transaction logs, not data files.
- Caveats: Ultra and v2 support is documented for specific sizes, so check yours. Size failover nodes too, and remember unused memory is expensive.
L-series (Lsv3, Lasv3): local NVMe, not durable storage
L-series offers 8 GiB per vCPU across 8 to 80 vCPUs. L80as_v3 carries 19.2 TB local NVMe at up to 3.8M IOPS and 20,000 MB/s. Remote Premium SSD, by contrast, caps at 80,000 uncached IOPS.
Local NVMe is ephemeral. Deallocation, redeployment, and some host events can wipe it.
- Good for: TempDB and scratch data. TempDB rebuilds at SQL startup; the SQL IaaS Agent extension can recreate folders and permissions at VM start.
- Never for: user databases, transaction logs, or backups.
- Note: Lsv3 doesn't support Premium storage host caching.
Constrained-vCPU E and M sizes
Constrained sizes keep the parent's memory and storage while limiting active vCPUs. E32-8s_v5 keeps 256 GiB, 32 data disks, and the parent's storage limits with 8 active vCPUs. M128-64bds_v3 and M176-88bds_v3 work the same way.
- Price: VM compute pricing stays the same as the parent. Savings come from SQL licensing on fewer cores.
- Floor: Per-VM licensing has a 4-core minimum.
- Hybrid Benefit: Needs active Software Assurance or a qualifying subscription. Enterprise SA covers up to 4 Standard SQL VM vCPUs per core, not a 4:1 Enterprise VM entitlement.
- Test first: Check availability, quotas, and resize limits, then validate under realistic concurrency.
How We Chose the Best Azure VM
Four mistakes show up again and again: sizing on RAM alone, sizing storage by capacity instead of performance, using average I/O instead of peak, and ignoring licensing or deallocation behavior.
Establish a workload baseline
Capture peak CPU, memory, IOPS, bandwidth, latency, and waits, as Microsoft's baseline guide recommends.
- Memory: Available MBytes, Total/Target Server Memory, Lazy Writes/sec, Page Life Expectancy, Memory Grants Pending
- I/O waits: PAGEIOLATCH_SH for page reads, WRITELOG for log flushes
- Latency: consistent waits above roughly 10-15 ms deserve investigation
Configure max server memory to leave room for Windows, and enable Query Store.
Translate I/O into VM and disk requirements
Measure read and write IOPS separately for data, log, and TempDB, plus latency, queue depth, block size, and throughput.
Check both layers. Per-disk ceilings land roughly here:
- Premium SSD: 20,000 IOPS and 900 MB/s
- Premium SSD v2: 80,000 IOPS and 2,000 MB/s
- Ultra Disk: 400,000 IOPS and 10,000 MB/s
The VM has its own aggregate cap. Adding disk IOPS never bypasses a VM ceiling.
Watch the Azure Monitor metrics for VM and disk IOPS and bandwidth consumed percentage. Microsoft's diagnostic threshold is 95% or higher for 5 consecutive minutes.

Match storage architecture to the workload
- Premium SSD: general default for data and logs
- Premium SSD v2 / Ultra: submillisecond latency and independently tuned IOPS; no host caching
- Local NVMe: TempDB only
Keep data, log, and TempDB on separate paths. Storage Spaces column count is fixed, so adding disks may add capacity without adding IOPS unless you recreate the pool. Validate SQL IaaS Extension behavior before resizing.
Validate performance and cost before production
Benchmark with representative data, concurrency, backups, failover, and peak writes. DiskSpd tests disks, but synthetic results don't equal SQL transaction throughput.
Separate the cost levers next:
- SQL licensing: pay-as-you-go, Hybrid Benefit, constrained vCPUs
- Compute: 1- or 3-year reservations or savings plans (these exclude SQL software and storage)
- Storage and ops: managed disks, HA/DR replicas, and day-to-day operations

Use the official Azure pricing pages for SQL Server Standard, Enterprise, and managed disks.
Decision checklist:
- Memory headroom confirmed
- Peak IOPS, throughput, and latency fit VM and disk limits
- TempDB placement defined
- Licensing and edition limits checked (Standard caps buffer pool at 128 GB)
- Resize path, zone or region design, monitoring, and rollback plan ready
Conclusion
The best Azure VM for SQL Server meets measured needs for memory, IOPS, throughput, latency, availability, and licensing without paying for idle capacity.
Reassess after migration and when demand shifts. Monitoring data will show whether the next move is query tuning, a disk change, or a resize.
Once the VM and disk architecture are set, Lucidity can add real-time visibility and autonomous optimization for Azure block storage, helping ITOps and FinOps teams cut spend on over-provisioned disks without infrastructure changes.
Frequently Asked Questions
What is a virtual machine in Microsoft Azure?
An Azure VM is an on-demand compute instance. SQL Server runs on it as IaaS, so you manage the operating system, database configuration, and storage.
What is the difference between SQL Server 2019 and 2022?
SQL 2022 adds Parameter Sensitive Plan optimization, DOP and CE feedback, and ledger. It also enables Query Store by default on new databases. 2019 mainstream support ended February 28, 2025; 2022's runs to January 11, 2028.
Can I use MSSQL in Azure?
Yes. Choose SQL Server on Azure VMs, Azure SQL Database, or Azure SQL Managed Instance. A VM fits when you need full compatibility or OS-level control.
How much memory does SQL Server need on an Azure VM?
It depends on working set, concurrency, and query patterns. Measure memory pressure, and set max server memory to leave room for Windows and other services.
How do I calculate IOPS for SQL Server on Azure?
Measure peak read and write IOPS, latency, throughput, block size, and queue depth separately for data, log, and TempDB. Compare them with both disk and VM-level limits.
Which Azure VM series is best for SQL Server?
D suits balanced workloads, E memory-heavy ones, M very large memory, and L local-NVMe TempDB. Eligible constrained-vCPU sizes help license-sensitive deployments.