Blog Post

Azure High Performance Computing (HPC) Blog
7 MIN READ

Matching Storage Performance to every stage of your AI Workloads on Azure

wolfgangdesalvador's avatar
Nov 17, 2026

Great AI performance starts with a well designed storage architecture. Optimizing how data moves to and from accelerators at every step, from training to inference, keeps the fleet working at its full potential.

Every minute a GPU spends waiting for data is a minute you pay for and get nothing back. As accelerators get faster and clusters grow, that idle time adds up quickly: training jobs stall waiting for the next batch, checkpoints can pause thousands of GPUs at once, and inference fleets sit idle while model weights load. Keeping GPUs productive comes down to one question: can your storage deliver the right data, at the right speed, at every stage of the workload?

On Azure, the answer is one architecture that combines durable core storage with workload-specific acceleration. Accelerator storage can bring data closer to compute or provide an accelerated access path within the durable storage foundation itself. In this post, we show how that works in practice, with real benchmark  results, and what it means for the time your GPUs spend computing instead of waiting.

One AI workload, many data paths

A single AI application places very different demands on storage across its lifecycle. Data preparation must enumerate and filter enormous numbers of objects. Training requires sustained high-throughput reads, while checkpointing creates bursts of write bandwidth. Model loading drives simultaneous reads of large files across many nodes, inference and agentic workloads add latency-sensitive access and the need to persist and reuse context.

Because these requirements vary so widely, the sweet spot is usually one durable data estate with a data path tuned to each stage. The storage solutions may differ, but they must work together so data, artifacts, and state remain available across the AI lifecycle.

One continuous data path on Azure

Azure uses Blob Storage as the durable data foundation, with multiple performance paths for workloads that require higher throughput, lower latency, file-system access, or compute-local caching. Core storage provides the foundation where data and state live across the AI lifecycle. Accelerator storage describes how performance is delivered, not only where a separate storage layer sits: it can place performance closer to compute or accelerate access to data within core storage itself. These are choices within one architecture, not a mandatory chain: each stage can use the path that best matches its throughput, latency, interface, and reuse requirements.

  • Core storage is the durable, massively scalable system of record for the whole AI lifecycle. On Azure, that foundation is Azure Blob Storage: one home for raw and curated datasets, checkpoints and model weights, built for the scale and durability AI data demands. Azure Blob Storage has a performance profile that in several scenarios can be smoothly plugged directly in the specific AI scenario and this performance can be augmented through accelerator storage. 
  • Accelerator storage provides the workload-specific speed that improves core storage performance when needed for compute. It can take the form of a storage or caching layer closer to compute, or an accelerated access path within the core storage service itself. In either case, the goal is to deliver the throughput and latency each stage needs, from sustained training reads and checkpoint write bursts to model-loading fan-out and fast access to reusable context, so GPUs spend more time computing and less time waiting.

Because core storage remains the system of record, acceleration can be matched to each job’s performance needs without turning every workload into a new data silo. Where a separate accelerator layer is used, its capacity can be sized to the job; where acceleration is delivered within core storage, higher performance can be applied to the data that needs it. Applications can retain familiar file system interfaces or existing object storage APIs, while data, artifacts and state remain part of one durable data estate.

Acceleration in action: performance where the workload needs it

Turbo tier: accelerating Azure Blob Storage

Turbo tier brings SSD-backed performance to your Azure Blob data while the remainder can stay on lower-cost HDD-backed tiers in the same storage account. It brings SSD-backed performance to performance-demanding data while retaining HDD-backed economics for the rest, within one standard Azure Storage account. Applications continue using the same account, namespace, Blob APIs, SDKs and tooling. Turbo is an access tier, not a cache or a separately managed high-performance storage account: customers can accelerate the subset of data that requires it rather than move an entire dataset into a separate SSD storage architecture.

Customers define a Turbo capacity target for the high-performance working set. Smart placement moves eligible objects between Turbo and HDD-backed tiers as access patterns change, while explicit tiering remains available when applications need deterministic control. The objective is to apply higher throughput, higher transaction rates and lower latency where they create workload value, without requiring every object to carry the cost of SSD-backed storage.

In a synthetic 4 KiB random-read test against the same dataset, Turbo tier delivered 16.8× the transaction rate, reaching 10.1 million transactions per second compared with 601,000 on Standard Blob with HDD-backed tiers. Turbo tier used 256 Standard D96s v6 worker VMs, while Standard Blob used 72 to stay within HDD-backed storage account limits. In a separate read-bandwidth benchmark, Turbo tier sustained more than 20 Tbps, relevant for checkpoint recovery, model loading and inference scale-out.

Read more in our blog post Blob Storage Turbo Tier: High performance with object storage scale and economics

For AI training and fine-tuning, this can help feed performance-sensitive training data and support checkpoint writes. For inference, it can provide faster access to models and application data as services start and scale; analytics and HPC pipelines can similarly accelerate repeated processing of the performance-demanding portion of a larger data lake. Preview eligibility is limited to LRS accounts in selected regions, subject to regional capacity.

Azure Managed Lustre: feeding training and protecting checkpoints

Azure Managed Lustre is another example of what the accelerator layer can deliver when AI workloads need performance close to compute. Its integration with Azure Blob Storage connects the high-performance file system used by the job with the durable core storage layer where datasets, checkpoints and model artifacts can remain available across the AI lifecycle. MLPerf Storage v3.0 results, released by MLCommons on September 1, 2026, provides an independent and comparable view of that accelerator performance. Azure Managed Lustre was the first hyperscale cloud storage service submitted at the scale of more than 600 GB/s (4.8 Tbps).

The results illustrate two critical accelerator storage patterns. For training, Azure Managed Lustre sustained the data path needed to keep simulated accelerator utilization above 90% for five consecutive hours, while read bandwidth scaled from 17.4 GiB/s at 3 simulated NVIDIA B200 accelerators to 379.1 GiB/s at 70. For checkpointing, the largest configuration, representing a 1,250B-parameter model, wrote at 642.2 GiB/s and recovered at 489.6 GiB/s, the second-highest peak checkpoint write among all submissions. Overall, Azure performed comparably to purpose-built on-premises AI storage in both categories.

 

Figure 1. MLPerf Storage v3.0 results on Azure. Left: training read bandwidth from 3 to 70 simulated B200 accelerators versus ideal linear scaling. Right: checkpoint write and recovery read throughput by model scale.

Explore the Azure Managed Lustre MLPerf Storage v3.0 results for benchmark methodology, tested configurations, and detailed training and checkpointing results.

DACS: bringing reusable data closer to GPU compute

DACS extends the accelerator pattern to AI workloads on Azure Kubernetes Service. Now in public preview, it is a distributed read-through cache that uses local NVMe disks across nodes to keep frequently read data from Azure Blob Storage close to GPU compute.
For repeated reads after the working set is warm, this can reduce Blob fetches and increase the bandwidth available for model loading and repeated training-data access.

This means reducing warm-model loading times and repeated Blob reads as inference workloads scale, restart or recover, and can accelerate training that repeatedly accesses the same dataset portions. In Standard LRS sequential-read testing with BlobFuse2 across ten workers, colocated DACS reached 47,760.75 MiB/s of warm bandwidth at 16 I/O threads per worker, 84.0% above BlobFuse2 alone and 10.9% above dedicated DACS. Teams can use DACS through Azure Container Storage with Run:ai Model Streamer, BlobFuse2 or the native Python client, while Azure Blob Storage remains the durable source of truth and the fallback when data is not cached.

 

Figure 2. Standard LRS sequential-read bandwidth by I/O thread count for BlobFuse2 alone and dedicated or colocated DACS, with cold and warm cache results.

Choosing the right acceleration for your bottleneck

To optimize your GPU data path, you can choose the best solution for each workload stage:

  • Azure Blob Storage can be accelerated using Turbo tier when an application needs higher-performance object access without changing its Blob account, namespace, or APIs.
  • Azure Managed Lustre can be used when training and checkpointing require a high-throughput parallel file system close to compute, and your applications need POSIX file system semantics.
  • DACS on AKS can be used when repeated reads of Blob data can benefit from distributed node-local caching. Today, DACS supports Run:ai Model Streamer, BlobFuse2 or the native Python client.
  • In all the cases, Azure Blob Storage remains the durable data foundation that can be still accessed directly by all the portion of the workflows that do not require specific acceleration.

The architectural goal is the same: place performance where the workload needs it, so compute spends less time waiting for data without creating new systems of record.

 

See it at SC26

These topics will be covered in our session at SC26. At the Microsoft booth, we’ll walk through the same data path end to end: preparing data, feeding training, protecting checkpoints, loading models and reusing context for inference. You’ll see where each storage layer helps keep GPUs busy and how the layers connect into one architecture. Join Wolfgang De Salvador, Sachin Sheth and Brian Lepore for Storage Innovation for AI on Wednesday 18 November, 2:00–2:20 pm CT, at booth #3411.

  • Session: Storage Innovation for AI
  • Speakers: Wolfgang De Salvador, Sachin Sheth and Brian Lepore
  • Where: Microsoft booth theater, booth #3411, McCormick Place, Chicago
  • When: Wednesday 18 November, 2:00–2:20 pm CT
Updated Nov 17, 2026
Version 1.0
No CommentsBe the first to comment