Azure Ultra Disk Max Throughput: Limits Explained Azure Ultra Disk is Microsoft's highest-performing managed block storage. Teams pick it for latency-sensitive databases, SAP HANA, and transaction-heavy systems. So the throughput ceiling gets a lot of attention.

But the headline number is one boundary in a chain. The disk configuration, VM size, workload pattern, and deployment constraints each set their own limit, and the lowest one wins.

This article covers the current documented ceiling, why real-world throughput falls short of it, how to validate performance, and what to check before resizing a disk.

Key Takeaways

  • Ultra Disk tops out at 10,000 MB/s per disk, a ceiling and not a promise.
  • The VM's uncached limit often caps throughput well below that.
  • I/O size, queue depth, and IOPS settings shape the throughput you actually use.
  • Test and monitor before paying for more provisioned performance.

What Azure Ultra Disk Max Throughput Represents

Throughput is the data volume moved between a workload and its disk per unit of time. Microsoft's premium storage guidance expresses the relationship as throughput = IOPS × I/O size.

The unit matters. Microsoft defines Ultra Disk MB/s as 10^6 bytes per second, so the maximum is 10,000 MB/s, not 10,000 MiB/s. Don't relabel it.

Four numbers get confused constantly:

  • Provisioned throughput: what you set on the disk
  • Service maximum: 10,000 MB/s per disk
  • VM limit: the uncached remote-storage allowance of the VM size
  • Measured throughput: what your application actually sees

Block size ties it together. At the same IOPS, larger I/O moves more MB/s. A disk running 20,000 IOPS at 8 KB delivers roughly 160 MB/s. The same IOPS at 256 KB is a very different number, still bounded by every published cap.

Factors That Influence Real-World Throughput

VM Limits Come First

Both the disk limit and the VM limit apply, and the VM can be the tighter one. Microsoft's size tables show how far these diverge:

VM size Uncached IOPS Uncached throughput
D2s_v5 (base) 3,750 85 MBps
D96s_v5 (base) 80,000 2,600 MBps
Standard_L80s_v3 80,000 2,160 MBps
Standard_M176s_4_v3 130,000 4,000 MBps
Ebdsv5 E96 (SCSI) 180,000 6,500 MBps
Ebdsv5 E96 (NVMe) 390,000 8,500 MBps
Ebdsv5 E112i (NVMe only) 400,000 10,000 MBps

Only the largest Ebdsv5 NVMe size reaches the disk ceiling. Controller type matters too: the same Ebdsv5 size publishes different limits for SCSI and NVMe.

Microsoft's disk performance documentation states that exceeding either the disk or VM allowance caps application performance. Throughput is also shared across every disk attached to the VM, so a single Ultra Disk may never see its own ceiling.

Azure Ultra Disk throughput bottleneck chain from workload to application

Workload Variables

  • Small random I/O favors IOPS; large sequential I/O favors MBps
  • Higher queue depth raises throughput but also latency (queue depth = IOPS × latency)
  • Read/write mix and concurrency change the request profile
  • Ultra Disk has no host caching and no bursting, so don't expect either to smooth spikes

Guest and Application Layers

Filesystem behavior, database settings, OS configuration, and compute contention can all hold results below what the disk is provisioned for. A disk set for 4,000 MBps will not deliver it to an application issuing small, shallow requests.

Range and Key Technical Properties of Ultra Disk Throughput

Nominal Operating Range

Per Microsoft's disk types page (updated September 15, 2026), the current boundaries are:

Property Documented boundary
Capacity 4 GiB to 64 TiB, 1-GiB increments
IOPS 100 minimum; up to 1,000 IOPS/GiB; 400,000 per disk
Throughput 1 MB/s minimum; up to 0.25 MB/s per provisioned IOPS; 10,000 MB/s per disk

Capacity, IOPS, and throughput are set independently within those bounds. The 2025 Ultra Disk update replaced the older 300 IOPS/GiB formula, so older guides may be out of date.

Published examples show how capacity couples to IOPS and throughput:

  • 4 GiB: 4,000 IOPS, 1,000 MB/s
  • 32 GiB: 32,000 IOPS, 8,000 MB/s
  • 64 GiB: 64,000 IOPS, 10,000 MB/s
  • 512 GiB: 400,000 IOPS, 10,000 MB/s

Azure Ultra Disk capacity IOPS and throughput examples comparison

Boundary Limits and Bursting

Ultra Disk delivers provisioned performance continuously. Managed-disk bursting is not supported, so there are no burst credits to lean on.

Microsoft lets you adjust performance without detaching the disk. Changes are capped at four per 24 hours, and each can take up to an hour to apply.

Safe Operating Margin

Microsoft states no fixed headroom percentage for Ultra Disk. Treat 100% consumed as the cap itself, not spare capacity.

Running near a boundary invites queue buildup, higher latency, throttling, and slow transactions when traffic spikes. Leave room for growth and for shifts in I/O size or concurrency.

Latency and Cost Trade-offs

Ultra Disk is designed for submillisecond latency, but that isn't a guarantee under every queue depth or throttled state. Average MB/s can hide tail latency that hurts databases.

More IOPS won't raise MB/s if the workload uses small I/O. Larger I/O may raise throughput without improving per-transaction latency.

Billing covers provisioned capacity, IOPS, and throughput hourly, plus a vCPU reservation charge. Check the current Azure pricing page before comparing against a larger VM, multiple disks, or another disk type.

How Throughput Is Specified, Measured, and Validated

Throughput is both a deployment spec and an operational measurement. Check it at the disk, VM, guest OS, and application layers.

Specification

Use Microsoft Learn for rated values, supported VM sizes, and regional availability. Ultra Disk supports data disks only (not OS disks), requires LRS, and has no availability set support. Keep rated values separate from tested values, and don't treat third-party figures as the Azure maximum.

Measurement

Azure Monitor exposes these metrics:

  • Data Disk Read/Write Bytes/Sec (per LUN)
  • Data Disk Read/Write Operations/Sec
  • Data Disk Queue Depth and Data Disk Latency
  • Data Disk Bandwidth Consumed Percentage and IOPS Consumed Percentage (100% means the disk cap)
  • VM Uncached Bandwidth Consumed Percentage and VM Uncached IOPS Consumed Percentage

Compare the disk and VM percentages to see which allowance is binding. One caveat: Microsoft's documentation has labeled the latency metric SCSI-only, so don't assume it covers NVMe.

Testing

Microsoft's benchmarking guide uses DiskSpd on Windows and fio on Linux, with parameters for block size, queue depth, concurrency, and duration. Those are tool examples, not Ultra results.

Record I/O size, read/write mix, queue depth, duration, concurrency, VM SKU, and disk settings so you can reproduce a run. Combine controlled tests with production observation. One synthetic peak proves little.

Azure Ultra Disk performance testing and validation workflow

Operating Outside the Range and Common Misreadings

Hit a disk or VM cap and the chain is predictable: queues grow, latency rises, throttling starts, and the application slows down. A larger disk allowance doesn't remove the separate VM cap.

Common mistakes:

  • Treating 10,000 MB/s as guaranteed application throughput
  • Confusing IOPS with MB/s
  • Assuming a bigger disk fixes every bottleneck
  • Ignoring VM aggregate limits and block size

Design Review Checklist

  1. Confirm the current Microsoft limit.
  2. Verify the VM's uncached ceiling and controller type.
  3. Match IOPS and throughput settings to the workload.
  4. Test under representative load.
  5. Monitor after deployment.
  6. Compare cost against utilization.

Six-step Azure Ultra Disk design review checklist process

Visibility helps with step 5 and 6. Lucidity's Lumen analyzes IOPS, throughput, latency, and cost metrics for every disk across Azure, AWS, and Google Cloud, which helps teams right-size provisioned storage. It complements Azure's controls and can't bypass disk or VM limits. Confirm Ultra Disk coverage with Lucidity before relying on Lumen for that tier.

Conclusion

Ultra Disk's 10,000 MB/s is a governing boundary, not a result every application will see. Reliable sizing means weighing the disk configuration, VM limits, workload shape, measurement method, and operating margin together.

Let current Microsoft documentation and real telemetry drive performance and cost decisions, not a single headline spec.

Frequently Asked Questions

Can I shrink my Azure disk size?

Not directly. Microsoft documents no in-place downsizing for managed disks. The workaround is to create a smaller data disk and copy the data over, which excludes OS disks.

What is the maximum throughput of an Azure Ultra Disk?

10,000 MB/s per disk (decimal units). Provisioned IOPS and the VM's uncached limit can lower the effective figure.

What limits Azure Ultra Disk throughput?

Provisioned IOPS and throughput, the VM size and controller, aggregate limits across attached disks, I/O size, queue depth, and guest or application configuration.

What is the difference between throughput and IOPS?

Throughput is data moved per second; IOPS is completed operations per second. I/O size connects them: throughput = IOPS × I/O size.

Can I change Ultra Disk performance without downtime?

Yes. You can adjust IOPS and throughput without detaching, up to four changes per 24 hours, taking up to an hour. Online capacity expansion is documented for NVMe-attached Ultra Disks in supported regions.

How do I monitor Ultra Disk throughput?

Use Azure Monitor disk and VM metrics, including bytes/sec, queue depth, and consumed-percentage metrics. Add guest OS and database telemetry, and compare against both disk and VM limits.