What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Choose storage and networking from measured workload behavior, not from the generic label “AI.” Record data volume, file or object counts, read and write patterns, throughput, latency, IOPS, metadata activity, concurrency, checkpoint size and interval, recovery targets, and expected growth. Then assign each data class to an appropriate tier and test the complete path from storage through the network into compute.
Object storage, parallel file systems, block storage, and local NVMe each solve different problems. A durable object store often holds the system of record and inactive artifacts; a POSIX parallel file system serves active, metadata-heavy jobs; and block devices or local NVMe provide provisioned disks or compute-attached scratch. The right design depends on your access semantics, synchronization pattern, failure objectives, and budget.
Start with the workload, not the storage brand
There is no universal “best” storage for AI. NVIDIA states that there is “no one-size-fits-all” storage solution for AI-driven applications. The practical implication is to describe the pipeline in numbers before selecting a service or appliance.
Measure the data set and namespace
- Total logical and physical data size, including replicas, transformed copies, indexes, and checkpoints.
- Number of files or objects, their size distribution, and whether workers open many small files or a few large streams.
- Read/write ratio, sequential versus random access, burst behavior at job start, and whether preprocessing creates temporary data.
- Simultaneous clients, accelerator count, expected scale-out, and synchronized phases such as epoch boundaries or restore.
Define performance in workload terms
Record sustained aggregate bandwidth and per-client throughput, but also tail latency, IOPS, metadata operations per second, and queueing when all workers issue requests together. A storage system that looks fast in a single-client test can stall when thousands of workers enumerate directories or open shards at once. Benchmark representative jobs end to end; theoretical interface limits do not predict training throughput by themselves.
#1 Best Overall
- MEET THE NEXT GEN: Consider this a cheat code; Our Samsung 990 PRO Gen4 SSD helps you reach near max performance with lightning-fast speeds; Whether you’re a hardcore gamer or a tech guru, you’ll get power efficiency built for the final boss
- REACH THE NEXT LEVEL: Gen4 steps up with faster transfer speeds and high-performance bandwidth; With a more than 55% improvement in random performance compared to 980 PRO, it’s here for heavy computing and faster loading
- THE FASTEST SSD FROM THE WORLD'S FLASH MEMORY BRAND: The speed you need for any occasion; With read and write speeds up to 7450/6900 MB/s you’ll reach near max performance of PCIe 4.0 powering through for any use
- PLAY WITHOUT LIMITS: Give yourself some space with storage capacities from 1TB to 4TB; Sync all your saves and reign supreme in gaming, video editing, data analysis and more
- IT’S A POWER MOVE: Save the power for your performance; Get power efficiency all while experiencing up to 50% improved performance per watt over the 980 PRO; It makes every move more effective with less consumption
Write down recovery and growth requirements
For each data class, specify retention, acceptable recovery point (how much work may be lost), recovery time, consistency requirements, and projected capacity growth. Include the time required to repopulate caches or distribute a model to every worker. These requirements determine whether an artifact belongs on an active tier, a durable object store, or an archive tier.
Match the storage interface to the job
Object, file, and block interfaces are not interchangeable. Choose the interface that matches the application’s semantics, then add caching or movement between tiers when the workload changes phase.
| Storage type | Typical role | Strengths | Questions to verify |
|---|---|---|---|
| Object storage | Large source datasets, durable model artifacts, inactive data, and asynchronous or tiered checkpoints | Scalable capacity and durability; lifecycle and retention policies; access through an object API | Can the framework use the object API directly? What are request costs, transfer charges, consistency behavior, and retrieval delays for colder tiers? |
| POSIX parallel file system | Active training data, low-latency reads and writes, synchronous checkpoints, and metadata-concurrent workloads | Shared file semantics, high aggregate throughput, and parallel metadata handling | How does throughput and metadata latency change with client count, small files, synchronized starts, and expansion? |
| Block storage or shared block pools | Provisioned disks, database-like components, shared capacity/throughput pools, and scratch attached to compute | Explicitly provisioned performance and block-device compatibility | Who manages the file system, how many clients can attach, and what happens when capacity or IOPS is exhausted? |
| Local NVMe or SSD | Compute-local scratch, preprocessing spill, caches, and performance-critical temporary data | Very short data path and high local performance | Is data lost when the instance is replaced? Are endurance, cooling, form factor, capacity, and server support suitable for sustained writes? |
Object storage for scale and durability
Object storage is a strong system of record for massive datasets and model artifacts when the application can use an object API or a compatible gateway. It scales capacity independently of compute and supports lifecycle policies that move inactive data to less expensive tiers. It is less suitable when a job requires POSIX locking, frequent metadata mutations, or very small synchronous reads unless a cache or file layer hides those operations.
Parallel file systems for active POSIX work
A managed parallel file system fits jobs that need shared POSIX paths, low-latency access, high aggregate bandwidth, or heavy metadata concurrency. It can hold a hot working set while durable copies remain in object storage. Validate namespace behavior, file-count limits, client scale, and performance during synchronized checkpoints rather than relying on a headline throughput figure.
Block storage and local scratch
Block pools are useful when you need provisioned capacity, throughput, or IOPS for compute and scratch. Local NVMe or SSD can be the fastest tier for temporary data and caches, but it generally requires rebuilding after host failure or replacement. Treat local capacity as disposable unless the platform explicitly provides persistence and redundancy.
Separate active data from inactive data
Keep the data needed by a running job close to compute, and move data that is no longer on the critical path to a durable, lower-cost tier. Microsoft’s Azure guidance recommends a file system for active job-specific (“hot”) data and moving inactive data to Blob Storage with lifecycle policies for Cool or Archive tiers.
Rank #2
- Ideal for high speed, low power storage
- Gen 4x4 NVMe PCle performance
- Up to 6,000MB/s read, 4,000MB/s write
- Includes Acronis cloning software
- 5-year limited warranty
A practical tier layout
- System-of-record tier: Store original datasets, approved model weights, and retained checkpoints in durable object storage with versioning or an equivalent recovery mechanism.
- Active tier: Stage the current training set, shards, and files requiring POSIX semantics on a parallel file system or provisioned block pool.
- Scratch tier: Use local NVMe/SSD or ephemeral block space for preprocessing output, shuffle data, and caches that can be recreated.
- Lifecycle tier: Apply explicit retention and transition rules to completed runs, old checkpoints, and logs. Document retrieval time and charges before placing recovery-critical artifacts in an archive class.
Do not leave every artifact on the fastest tier. Compare the cost of storage, requests, transfers, cache fills, and operator time with the cost of a slower restore or a rerun.
Design checkpointing as a storage-and-network problem
Checkpoints combine large writes, consistency requirements, and synchronized reads during recovery. A synchronous write can pause every worker while data is flushed. A restore can overload a shared link when many accelerators request the same large artifact at once.
Estimate checkpoint bytes
Count model parameters, optimizer state, gradients or momentum, scheduler state, random-number state, data-loader position, and any sharding metadata that must be persisted. The byte count depends on precision and parallelism; a sharded checkpoint may be smaller per worker but still large in aggregate. Measure the actual serialized output rather than multiplying parameter count by a single assumed factor.
Choose a write pattern
- Synchronous persistence: Simpler recovery semantics, but the training step waits for storage completion. Measure the pause and its effect on utilization.
- Asynchronous persistence: A background process or host writes a consistent snapshot while training continues. Ensure the snapshot is complete and identifiable before marking it recoverable.
- Hierarchical persistence: Write first to fast local or file-system storage, then copy to durable object storage. Define how long the local copy is trusted and how failures during promotion are handled.
Microsoft gives “every 500 iterations” as an example checkpoint interval; it is guidance, not a universal rule. Set the interval from the value of lost work, checkpoint duration, storage budget, and failure rate. Test interrupted writes and restart from the newest known-good checkpoint.
Control restore fan-out
When many workers load the same model, avoid making every accelerator independently fetch a large object across a shared external link. A staged design can have leader nodes read the durable checkpoint and distribute it across the accelerator cluster. AWS describes this hierarchy as a way to reduce direct external-storage load. Its examples, including a modeled 125 TB checkpoint and a 4,000-accelerator scenario, are calculations for defined restore goals, not industry-wide benchmarks.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Design two network planes
Training data access and worker-to-worker exchange are separate networking concerns. AWS’s training guidance distinguishes traffic from storage to nodes from east-west communication among training nodes. Size and monitor both.
Rank #3
- SPEED UP PROJECTS. Launch creator applications fast with uncompromising PCIe 4.0 read speeds up to 7,100MB/s,[2] (1TB and 2TB[1] models) and write speeds up to 6,700MB/s[2] (1TB[1]-4TB[1] models).
- CREATE AND STORE MORE. Make more room for your 4K videos and high-resolution images with capacities from 500GB[1] up to 4TB[1] on M.2 2280 built with our trusted 8th generation SANDISK BiCS QLC 3D CBA NAND.
- IT GOES WHERE YOU GO. With an all-new power efficient design, your drive delivers high performance with low power, giving you more time to be productive while on the go.
- UNCOMPROMISED RELIABILITY. With up to 1,200 TBW[3] (4TB[1] model) endurance rating, your drive is designed for creators.
- KEEP YOUR DRIVE UPDATED. Monitor your SSD’s performance and check for updates with the downloadable SANDISK Dashboard application.[5]
Storage-to-compute traffic
Measure aggregate read and write bandwidth from the storage service to all clients, per-node throughput, tail latency, and behavior during cold starts. Include cache-fill traffic: a job may appear fast after a warm cache but fail its first-run service-level objective when every worker reads the same shards.
Worker-to-worker traffic
Collective operations, gradient exchange, parameter synchronization, and distributed checkpoint coordination use the east-west network. Check link speed, topology, oversubscription, congestion control, and placement across zones or racks. A storage upgrade will not fix a training job limited by worker collectives.
Account for placement and transfer
Keep compute and active storage in the intended region and, where applicable, zone or availability domain. Cross-zone or cross-region transfers can add latency and cost and may consume the same links needed for collectives. Document which traffic is allowed to leave the failure domain and what happens when a preferred zone is unavailable.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteProvider patterns you can adapt
The following examples illustrate role separation, not a universal vendor ranking. Product limits, configuration requirements, prices, and regional availability change; verify current terms before procurement.
Microsoft Azure
Azure’s AI storage guidance recommends a file system for active job-specific data. It identifies Azure Managed Lustre when minimized transfer time and latency are the priority, Azure NetApp Files when enterprise features, existing NetApp investments, hybrid use, or customization matter, and local NVMe/SSD file systems when performance is the overriding objective. It recommends moving inactive data to Azure Blob Storage and applying lifecycle policies for Cool or Archive tiers.
Use the “every 500 iterations” checkpoint example only as a starting point. Measure the time to write and validate a checkpoint, the amount of work lost at your chosen interval, and the restore time under concurrent load.
Rank #4
- HUGE SPEED BOOST: Get random read/write speeds that are 40%/55% faster than 980 PRO; Experience up to 1400K/1550K IOPS, while sequential read/write speeds up to 7,450/6,900 MB/s reach near the max performance of PCIe 4.0*
- BREAKTHROUGH POWER EFFICIENCY: Use less power and get more performance; Enjoy up to 50% improved performance per watt over 980 PRO, plus optimal power efficiency with max PCIe 4.0 performance**
- SMART THERMAL CONTROL: Samsung's own nickel-coated controller delivers effective thermal control; With its slim size, 990 PRO is a perfect fit for desktops and laptops that meet the PCI-SIG D8 standard***
- THE CHAMPION MAKER: Up to 65% improvement in random performance enables faster loads for an ultimate gaming experience on PS5 and DirectStorage PC games****
- SAMSUNG MAGICIAN SOFTWARE: Get the most out of your SSD with Samsung Magician's advanced yet intuitive optimization tools; Monitor drive health, protect valuable data, and receive important updates for your 990 PRO
Google Cloud
Google Cloud’s AI Hypercomputer documentation separates Cloud Storage object storage, Managed Lustre POSIX storage, and Hyperdisk pools. Cloud Storage is positioned for large datasets and model artifacts; Managed Lustre for low-latency, metadata-concurrent activity such as synchronous checkpointing; and Hyperdisk pools for shared capacity, throughput, and IOPS used by compute and scratch.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →The same documentation describes Rapid Bucket and Rapid Cache options. It states that Rapid Bucket can deliver up to 15 TB/s and sub-millisecond latency for open files in its described zonal context. That is a provider-stated product limit, subject to configuration and availability, not an independently measured cross-cloud benchmark. Confirm the current documentation and the exact workload conditions before using the figure in a design target.
AWS
AWS describes linking S3 training data to FSx for Lustre for low-latency, high-throughput file access, or accessing S3 directly through VPC gateway endpoints. Choose the file-system path when the application needs file semantics and a hot working set; direct object access can be simpler when the framework already streams from S3 and the access pattern tolerates object semantics.
AWS also treats inter-node exchanges as a separate training-networking problem and describes hierarchical checkpoint distribution in which leader nodes read a checkpoint and distribute it across the accelerator cluster. Size the external link, leaders, and internal fabric for simultaneous restores rather than for one reader.
Step-by-step selection process
- Characterize the pipeline. Record dataset size, file/object count and size distribution, read/write ratio, batch behavior, preprocessing, client concurrency, checkpoint contents and interval, and cold-start requirements. Benchmark a representative job from source data through compute.
- Classify each data set. Mark data as system-of-record, active hot data, scratch, or retained archive. Identify which components require POSIX paths, which can use an object API, and which can be recreated.
- Select the active tier. Evaluate object storage for scalable datasets, a managed parallel file system for low-latency POSIX and metadata-heavy activity, and local NVMe/SSD or block pools for scratch and provisioned performance.
- Select the durable tier. Keep long-retention data in durable object storage or an appropriate archive. Add lifecycle rules, version handling, and a documented restore procedure instead of leaving every artifact on the active tier.
- Plan checkpoint movement. Calculate real checkpoint bytes, measure write pauses and restore duration, and test asynchronous or hierarchical persistence when the framework supports it. Include validation, naming, retention, and cleanup.
- Map both network planes. Measure storage-to-node throughput, east-west collective traffic, oversubscription, topology, congestion, and zone/region transfer. Test synchronized starts and simultaneous restores.
- Validate economics and failure behavior. Compare cost per completed training run or inference request, including requests, transfers, cache fills, and operator effort. Simulate node loss, an interrupted checkpoint, stale or inconsistent data, a full scratch tier, and a partial network failure.
How to size capacity and throughput
Capacity envelope
Build a capacity envelope from the system-of-record copy, active working set, scratch expansion, retained checkpoints, replicas, and growth horizon. Include temporary space created during conversion, shuffling, decompression, and failed-job cleanup. A design that fits the source dataset but not two concurrent preprocessing jobs will exhaust its active tier at the worst time.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsThroughput target
Set a target from the job’s consumption rate and concurrency. If each worker consumes a measured rate, multiply by the number of concurrent workers, then add headroom for synchronized bursts, checkpoint writes, and cache misses. Verify that the storage service, client network, host interfaces, and file-system settings can sustain the target simultaneously; the narrowest component controls end-to-end performance.
Best Value
- This product has been replaced by our latest generation. Please search for the SANDISK Optimus GX 7100 NVMe SSD
- HIGH-OCTANE GAMING. Experience speeds up to 7,250MB/s read and 6,900MB/s write (1-2TB models), with up to 35% faster performance than previous generation.
- PURPOSE-BUILT. Designed for serious on-the-go gamers, with a PCIe Gen4 interface and SANDISK’s next generation TLC 3D NAND.
- MORE TIME TO CLEAR THAT CHECKPOINT. Built with laptops and handheld gaming devices in mind, with up to 100% more power efficiency over the previous generation.
- DO MORE WITH DASHBOARD. Ensure your drive is optimized for prime performance with the downloadable WD_BLACK Dashboard (Windows only).
Latency and metadata target
For small-file or metadata-heavy pipelines, count open, stat, create, rename, and directory operations, not only bytes per second. Measure tail latency during job launch and checkpoint commit. If the application can shard data into larger objects or prebuild an index, compare that change with buying a faster tier.
NVMe and SSD compatibility checklist
Local NVMe/SSD is a performance tier, not a guarantee that any consumer drive will work in a GPU server. Before specifying a device, verify:
- Interface and protocol support, including the host’s PCIe generation and available lanes.
- Form factor, carrier, hot-swap requirements, and physical clearance.
- Endurance rating for sustained checkpoint and preprocessing writes.
- Capacity after formatting, overprovisioning, and any RAID or redundancy layout.
- Cooling and thermal throttling behavior in the target chassis.
- Server, firmware, hypervisor, and operating-system support.
- Whether the data is disposable and how it is rebuilt after host replacement.
Common failure modes and fixes
“The benchmark is fast, but training is slow”
Check client concurrency, metadata latency, cache warmth, CPU preprocessing, and network oversubscription. Re-run with the real file count and synchronized worker start. A single-stream test does not establish aggregate performance.
“Workers pause at every checkpoint”
Measure serialization separately from the storage flush. Reduce unnecessary state, use sharded or asynchronous persistence where correctness permits, and stage copies before promoting them to durable storage. Do not increase the interval blindly if the resulting recovery point exceeds your loss tolerance.
“Restore overwhelms the network”
Throttle or stage fan-out, use leader-based distribution, and reserve capacity for east-west collectives. Confirm that the durable store, file system, and internal fabric can handle all readers at once.
“The hot tier is growing without bound”
Set lifecycle rules for completed runs, delete reproducible intermediates, and keep an auditable durable copy before eviction. Track cache occupancy and age so cleanup does not remove the newest known-good checkpoint.
“A zone or host failure loses the scratch data”
Assume local NVMe and ephemeral disks are lost unless the platform explicitly guarantees persistence. Rebuild scratch from the durable source and ensure the job can resume from a validated checkpoint.
Recommended Free Tools
Decision rule
Choose the least expensive architecture that meets measured throughput, latency, metadata, concurrency, durability, and recovery targets. Use object storage as the durable foundation when its API and access pattern fit; add a POSIX parallel file system for active shared work; use block pools or local NVMe for provisioned or disposable scratch; and engineer the two network planes independently. Re-test whenever model size, worker count, checkpoint format, region, or service configuration changes.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
