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1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsBandwidth (or achieved throughput) tells you how much data storage can move per second, IOPS tells you how many I/O operations it can complete per second, and latency tells you how long an individual operation takes. None is a universal “AI storage score.” Large training reads and checkpoint writes usually expose throughput limits, while retrieval, metadata-heavy inference, and agentic workloads often depend more on IOPS and latency. The correct choice comes from matching the metric to your I/O size, access pattern, concurrency, and measurement point.
A drive’s headline specification is not the same as application performance. The result can be limited by queue depth, PCIe lanes, host settings, networking, shared storage, or the way a storage service splits and counts requests. Use the workload-specific guidance and comparison method below instead of choosing an SSD from one peak number.
As an Amazon Associate I earn from qualifying purchases.
What the three storage metrics actually measure
Bandwidth and throughput: data moved per second
Bandwidth is a data-rate ceiling, normally expressed in bytes per second. NVIDIA describes it as the theoretical maximum attainable data rate. Throughput is the rate a system actually achieves, such as measured sequential reads from an NVMe SSD. Product pages often use the terms loosely, so first determine whether a figure is a theoretical limit or an observed result. The distinction is important: a drive can advertise a high bandwidth ceiling while an application achieves much less because of its I/O pattern or the rest of the storage path. See the definitions and NVMe context in the NVIDIA GPUDirect Storage Benchmarking and Configuration Guide.
Bandwidth becomes visible when requests are large enough and sufficiently concurrent to keep the device busy. Sequential model reads, dataset scans, and large checkpoint transfers are typical examples. A sequential benchmark that reports several gigabytes per second does not predict the response time of one small metadata lookup.
#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
IOPS: operations completed per second
IOPS means input/output operations per second: the number of read or write operations storage completes in one second. IOPS does not specify how many bytes are in each operation. A result measured with 4 KiB requests and a result measured with 1 MiB requests can both be reported as IOPS while moving radically different amounts of data. Micron explains the definition and the factors that affect achievable IOPS in its IOPS glossary.
That missing request-size detail is why an IOPS headline is incomplete. As a rough relationship, data rate equals operations per second multiplied by bytes per operation:
data rate ≈ IOPS × I/O size
The approximation is bounded by device firmware, interface bandwidth, queueing, protocol overhead, and the host or network. For illustration, 100,000 operations per second at 16 KiB each would represent about 1.53 GiB/s before those limits and overheads; the same 100,000 IOPS at 4 KiB would represent about one quarter as much data. The arithmetic is useful for sanity-checking specifications, not for predicting an application by itself.
Free tools Windows power users keep installed
One-click scans. No signup required.
Latency: time for one operation
Latency is the time required to complete an operation. Always identify where it is measured: at the NAND device, an appliance, or end to end from the client. AWS defines EBS latency as the time from sending a request until receiving acknowledgement that the operation is complete; that is a service-specific, request-level definition documented in Amazon EBS I/O characteristics and monitoring.
Latency is usually reported in milliseconds or microseconds. Averages can hide stalls, so compare a distribution or percentile when the workload is sensitive to pauses. A system may sustain high aggregate throughput while an individual request waits behind a queue, and a device with excellent median latency can still produce problematic outliers under load.
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
Why the metrics cannot be substituted for one another
- High bandwidth does not guarantee a fast response to a small, isolated request.
- High IOPS may describe tiny requests that deliver modest total throughput.
- Low device latency does not prove low client-to-storage latency when software, PCIe, network, or shared infrastructure adds delay.
- Increasing queue depth can raise utilization and IOPS while also increasing waiting time.
NVIDIA’s guidance for small I/O loads uses under 500 kB as an example: in that situation, IOPS and latency are generally more informative than throughput because requests may not consume the available data-rate ceiling. Large sequential transfers make achieved throughput more visible. The complete path can still be constrained elsewhere, including host configuration, PCIe, networking, or shared storage (NVIDIA GPUDirect Storage guide).
How AI workloads stress storage differently
“AI storage” is not one access pattern. NVIDIA’s workload descriptions below provide typical object-size ranges and patterns; they are guidance, not universal limits or performance guarantees. Use them to decide which measurements to collect, then validate with your own application.
Recommended Free Tools
| Workload | Typical I/O pattern | Object size listed by NVIDIA | Metrics to investigate first |
|---|---|---|---|
| Training | Large sequential model and dataset reads, periodic large sequential checkpoint writes, parallel streams | 256 KiB–4 MiB | Sustained achieved throughput, read/write mix, concurrency, and checkpoint behavior |
| Fine-tuning | Frequent base-weight reads plus iterative small or medium writes for adapter deltas and checkpoints | 64 KiB–1 MiB | Read throughput, write behavior, and latency under the actual mixed workload |
| Inference | Hot model-weight reads and metadata-heavy lookups in the serving graph | 4 KiB–256 KiB | Weight-loading throughput plus lookup IOPS and latency |
| KV cache | Random reads and writes, temporal locality, and possible persistence or eviction | 1–5000 KiB | Mixed-operation performance, latency, IOPS, locality, and cache behavior |
| Query and retrieval | Random reads for approximate-nearest-neighbor or vector lookup, with small ingestion and update writes | 1–120 KiB | Random-read IOPS and latency, including writes and updates |
| Agentic workflows | Many small state, checkpoint, memory, trace, tool-output, and artifact operations across concurrent sessions | 1 KiB–4 MiB | Latency and IOPS under concurrency, aggregate throughput, and predictable behavior |
These ranges and patterns come from NVIDIA’s NVIDIA-Certified Storage guidance. They should shape a test plan, not replace one.
Training: sustained movement and checkpoint interference
Training commonly streams large portions of a dataset and model, then writes checkpoints in large sequential operations. Measure sustained read throughput and the read/write mix while the training job is active. Checkpoint timing matters: a device that looks fast during reads may create a pause when a large write competes for bandwidth or queue capacity. Report concurrency and whether multiple workers are reading in parallel.
Fine-tuning: reads plus iterative writes
Fine-tuning repeatedly reads base weights while writing adapter deltas and checkpoints. A read-only sequential test can miss write-induced latency or contention. Test the mixed ratio and request sizes produced by the training framework, and observe both throughput and tail latency during checkpoint creation.
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]
Inference: separate model loading from serving lookups
Loading model weights can be a throughput task, but steady-state serving may issue smaller metadata and lookup requests. Measure startup or model-reload time separately from live request handling. For the serving phase, include random or metadata-heavy operations, IOPS, and latency at the client-visible measurement point.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →KV cache: locality and eviction are part of performance
KV-cache activity can combine random reads and writes with temporal locality. Persistence and eviction add another layer of behavior. A single sequential number is a poor proxy; test the mixed operation pattern, cache hit and miss behavior, and latency while eviction is active.
Query and retrieval: small random reads dominate
Approximate-nearest-neighbor or vector retrieval typically performs random reads in the 1–120 KiB range listed by NVIDIA, along with small ingestion or update writes. Random-read IOPS and latency should be primary measurements, but include updates so a write burst does not unexpectedly degrade query response.
Agentic workflows: concurrency and consistency matter
Agents can generate many small state, memory, trace, tool-output, checkpoint, and artifact operations at once. The important question is not only peak IOPS; it is whether latency remains predictable as concurrent sessions increase. Track aggregate throughput and latency percentiles while the full mix is running.
Choosing the metric that matters for your workload
| If your dominant activity is… | Start with… | Do not omit… |
|---|---|---|
| Large sequential dataset or model movement | Sustained achieved throughput | Read/write mix, worker concurrency, and queue depth |
| Large checkpoint writes | Sustained write throughput | Training slowdown and latency during checkpoint overlap |
| Small random retrieval or metadata lookups | Random-read IOPS and latency | Request size, tail latency, and update writes |
| Mixed cache reads and writes | Latency and IOPS under the mixed pattern | Locality, hit rate, and eviction behavior |
| Many concurrent agent sessions | Latency distribution and IOPS at target concurrency | Aggregate throughput and outlier behavior |
For a workload that changes phases, use a phase-specific scorecard rather than one weighted average. For example, record training read throughput, checkpoint write time, and the latency seen by any concurrent service. A single “maximum IOPS” line cannot represent all three.
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
How to compare benchmark results fairly
A comparison is meaningful only when the workload and measurement point are clear. Hold the following variables constant or report them next to every result.
- Access pattern and mix: State whether operations are sequential or random and the read/write percentage.
- I/O size: Record the request size and whether the storage layer merges, splits, or otherwise transforms operations.
- Concurrency and queue depth: Document worker count and outstanding requests. More queued work can increase utilization, but it can also increase latency.
- Measurement point: Say whether the result is device-level, appliance-level, or end to end from the application client.
- Duration and distribution: Distinguish a short peak from sustained behavior, and report latency percentiles when outliers affect users or job progress.
- Application outcome: Where possible, include training time per epoch, checkpoint duration, inference throughput, or query response time alongside component metrics.
Use subsystem tests, then validate the full application
Tools such as fio are useful for isolating a storage subsystem. NVIDIA’s GPUDirect Storage documentation describes microbenchmarks as building blocks, not substitutes for the full application: they may not capture workload diversity, concurrency, stress, and full-stack interactions. Run a controlled test to find a bottleneck, then replay a representative application phase to confirm that the bottleneck changes the outcome.
Why queue depth changes the answer
At low queue depth, a device may not receive enough work to reach its provisioned or advertised IOPS and throughput. At very high queue depth, utilization may improve while latency rises. The useful operating point is the one that matches your application’s concurrency and service-level target, not automatically the highest queue depth in a benchmark.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.The AWS EBS example: why request accounting matters
AWS EBS illustrates why storage numbers must be read with service rules attached. For its SSD volumes, an individual I/O is capped at 256 KiB. Larger requests may be split, while sequential smaller requests may be merged; random smaller requests are counted separately. Consequently, the IOPS and throughput reported by EBS depend on request size and access pattern, not just the amount of data your program asks for.
AWS also notes that the appropriate queue length varies by workload. If the host does not issue enough concurrent requests, a volume may never reach its provisioned IOPS or throughput. These are EBS behaviors, not universal rules for NAND or every SSD, but they demonstrate why a benchmark must document the storage service’s operation-counting and request-transformation rules. See the complete EBS I/O characteristics documentation.
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).
Reading an NVMe SSD specification for AI use
Separate the headline maximum from your path
For local direct-attached NVMe SSDs, vendor specifications commonly list throughput, IOPS, and latency. NVIDIA cautions that theoretical performance also depends on PCIe bandwidth and the configured lane count (GPUDirect Storage configuration guide). Verify the host’s PCIe generation, lane allocation, firmware, cooling, filesystem, and driver path before treating a drive’s maximum as attainable application throughput.
Match the specification’s test conditions
Look for the request size, read/write mix, queue depth, and whether the number is peak or sustained. A random-read IOPS figure measured with tiny requests is relevant to a retrieval test only if your retrieval layer issues similar requests. A sequential-read bandwidth figure is relevant to training only if the dataset path can supply enough parallel, contiguous work.
Choose by workload, not by the largest number
When evaluating an NVMe SSD for a local AI node, write down the dominant phases and their requirements:
- Required sustained read and write rates for data movement and checkpoints.
- Random IOPS and latency targets for retrieval, metadata, or cache activity.
- Concurrency and queue depth generated by the number of workers, GPUs, and services.
- Whether the result must be measured at the SSD, host filesystem, or application client.
- How performance behaves after the workload runs long enough to expose thermal, cache, or steady-state effects.
Do not label one drive “best for AI” without workload-specific evidence. Two systems with identical SSDs can produce different application results if their PCIe topology, host settings, software stack, or shared path differs.
A practical test plan
- Capture the real trace: Measure representative request sizes, sequential versus random access, read/write mix, worker count, and queue depth during each AI phase.
- Define success in application terms: Set targets such as epoch duration, checkpoint completion time, inference requests per second, or a query-latency percentile.
- Run a controlled subsystem test: Use a tool such as
fiowith the same I/O size, mix, and queue depth. Record whether the test is device-level or end to end. - Stress the transition points: Test startup, checkpoint overlap, cache eviction, ingestion updates, and increasing concurrent sessions rather than only steady sequential reads.
- Compare distributions: Report sustained throughput, IOPS, and latency percentiles over a stated interval. Keep peak and average values labeled.
- Validate with the application: Repeat the workload and confirm that a subsystem improvement changes the end result. If it does not, another part of the path is the bottleneck.
Common interpretation mistakes and their fixes
- Mistake: treating IOPS as a complete performance number. Fix: always pair IOPS with I/O size, queue depth, and read/write pattern.
- Mistake: using sequential bandwidth to predict random retrieval. Fix: run random-read tests at the request sizes your index or serving layer uses.
- Mistake: quoting device latency as user-visible latency. Fix: identify the measurement point and include filesystem, network, and software time when the application is remote.
- Mistake: benchmarking with too little outstanding work. Fix: match concurrency to production and state the queue depth; insufficient request supply can prevent a system from reaching its provisioned limits.
- Mistake: selecting the maximum queue depth because it produces the highest IOPS. Fix: choose the queue depth that meets latency and throughput targets together.
- Mistake: trusting a short peak result. Fix: run long enough to observe sustained behavior and report outliers.
- Mistake: assuming all AI jobs have the same storage profile. Fix: classify each phase as primarily sequential, random, mixed, or concurrency-sensitive, then test accordingly.
The decision rule
Use throughput when your job moves large, mostly sequential objects and total data volume controls runtime. Use IOPS plus latency when it issues many small or random requests, especially for retrieval, metadata, cache, and agent state. Use all three when phases mix large transfers with small interactive operations, and judge the result at the application’s measurement point.
The most defensible comparison reports the workload, request size, queue depth, concurrency, measurement location, sustained duration, latency distribution, and final application outcome. That context turns a storage specification into an engineering decision instead of a misleading peak-number contest.
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.
