The defensible way to measure an AI SSD is a controlled before-and-after test: run the same model, data, software, accelerator and workload on the existing storage path, then repeat after changing only the storage configuration. Record an application result—such as training time per epoch, model-load time, inference throughput or tail latency—alongside storage throughput, IOPS, latency and accelerator utilization. A faster SSD metric is useful only when the AI workload is actually limited by the storage path.
What the comparison should prove
“With and without an AI SSD” is not a universal product test. It is a comparison between two complete storage configurations under a defined workload. The baseline might use an existing local disk, a different NVMe drive, a RAID volume or network storage. The candidate might be one or more NVMe SSDs attached over PCIe, a different filesystem or a changed I/O path. If you change the filesystem, caching policy, RAID level, storage software or network at the same time as the drive, your result measures the new configuration, not the isolated effect of the SSD.
As an Amazon Associate I earn from qualifying purchases.
Keep the workload and the rest of the system constant. Use the same model checkpoint, dataset, preprocessing, application version, framework, driver, accelerator, host, network, worker count and run protocol. Document exactly what changed. Storage behavior depends on the entire data path, including PCIe topology, filesystem, software stack and remote-storage network. NVIDIA’s GPUDirect Storage Benchmarking and Configuration Guide and its NVIDIA-Certified Storage documentation both emphasize measuring the path rather than treating a drive’s advertised specification as application performance.
Your conclusion should therefore be workload-specific: did the changed storage path reduce epoch time, improve time to first response, sustain more inference requests, reduce tail latency, raise accelerator utilization or scale better with concurrent workers? If none of those outcomes changed, report that plainly even if a synthetic benchmark shows higher bandwidth.
#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
Define a reproducible baseline and candidate configuration
Write down the baseline
Before installing or reconfiguring anything, describe the current path from the application to its data. Include the storage device or remote source, capacity, filesystem, mount options, RAID or volume layout, interface, PCIe generation and slot, storage protocol, CPU-mediated or GPU-direct transfers, and cache policy. Note whether data is local, shared over a network or staged into a node-local cache.
Change one meaningful variable
The cleanest experiment replaces the storage device while leaving the path otherwise unchanged. A practical deployment comparison may intentionally change several components—for example, moving from network storage to local NVMe—but label it as a configuration comparison. State every change in the article or test log; otherwise readers cannot tell whether an observed gain came from the SSD, a new filesystem, a wider PCIe link or a different software path.
Freeze software and system state
- Record the model, precision, batch size, sequence length, tokenizer and dataset revision.
- Record framework, data-loader, storage-library, driver, firmware, kernel and filesystem versions.
- Use the same accelerator model, host CPU and memory, worker count, container image and power mode.
- Keep network links, switch paths and remote-server settings unchanged when either run uses shared storage.
- Use the same warm-up, measurement interval, checkpoint schedule and stopping rule.
Run the baseline first, then reset the system to the same state for the SSD condition. If the experiment requires a reboot, include an equivalent reboot or settling procedure for both conditions.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteChoose workload-shaped tests instead of a single sequential benchmark
AI pipelines do not all read and write data in the same way. The workload profiles below are starting points from NVIDIA’s certification documentation, not guarantees for a particular application. Replace the representative object sizes with measurements from your own data loader, model server or retrieval engine whenever possible.
| Workload | Typical storage behavior to reproduce | Representative object sizes | Application outcomes to capture |
|---|---|---|---|
| Training | Large sequential dataset reads, many parallel streams and periodic large sequential checkpoint writes | 256 KiB–4 MiB | Time per epoch, samples per second, checkpoint duration and accelerator utilization |
| Fine-tuning | Repeated reads of base weights with iterative small or medium writes for adapter deltas and checkpoints | 64 KiB–1 MiB | Time to complete a fixed number of steps, validation cadence and checkpoint time |
| Inference | Hot reads of model weights, embeddings and prompt/context features, with metadata-heavy lookups | 4 KiB–256 KiB | Requests or tokens per second, time to first response, per-request latency and tail latency |
| KV cache | Random reads and writes with temporal locality as active contexts change | 1–5000 KiB | Context-load latency, sustained token rate and cache-hit or eviction behavior |
| Query and retrieval | Random approximate-nearest-neighbor or vector lookups plus small ingestion and update writes | 1–120 KiB | Queries per second, p95/p99 latency, ingestion lag and freshness time |
Do not force a workload into a simplistic “sequential versus random” label. Samsung’s MLPerf Storage Benchmark Suite for NVMe SSDs describes cases such as 3D U-Net, ResNet50 and CosmoFlow where sample processing can generate small sequential I/O directed at repeatedly random addresses. Record the actual request sizes, queue depth, read/write mix and access pattern if your tooling exposes them.
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
Measure application impact and storage behavior together
Application-level measures
Select at least one outcome that matters to the user of the AI system. Training may use time per epoch, total time for a fixed number of steps, samples per second or checkpoint completion time. Inference may use requests per second, tokens per second, time to first response, inter-token latency and p95 or p99 latency. Retrieval may use query rate, tail latency and ingestion-to-availability time. Also record accelerator utilization over the measurement window; long idle periods while data is being fetched can reveal a storage bottleneck.
Storage-level measures
- Throughput: bytes per second for reads and writes, with the read/write mix stated.
- IOPS: operations per second at the tested request size and concurrency.
- Latency: an average is insufficient; capture a distribution and relevant percentiles.
- I/O shape: request size, queue depth, parallel streams, access pattern and synchronization behavior.
Throughput, IOPS and latency describe different properties. A drive can deliver high sequential bandwidth while providing little benefit to metadata-heavy inference or small random retrieval reads. Conversely, lower aggregate bandwidth may still improve tail latency for a lookup service.
System context
Log host CPU utilization, memory pressure, GPU utilization, PCIe link behavior where available and network utilization. A CPU-bound data decoder, saturated network link or narrow PCIe slot can hide an SSD improvement. If GPU-direct storage is used, identify that path; if transfers are staged through host memory, say so. The NVIDIA guide distinguishes application metrics from generic storage metrics for this reason.
Control page cache and the I/O path
Operating-system page cache can turn a storage test into a memory test. NVIDIA states: “Any performance evaluation of the storage technology requires bypassing the page cache to ensure that actual storage performance is being measured, and not the page cache read/write performance.” Its GDS guidance documents O_DIRECT for bypassing page cache on that I/O path; use the mechanism appropriate to your application and verify that alignment and buffering requirements are met.
Separate two legitimate but different questions:
- Cold-cache storage behavior: clear or bypass cache so each condition exercises the storage path. Record the cache-clearing or direct-I/O method and any warm-up reads.
- Warm-cache application behavior: leave the cache populated when that reflects production, such as a model server that keeps weights resident. Report how the cache was populated and how much data fits in memory.
Never mix a cold-cache baseline with a warm-cache SSD result. Record whether each transfer uses CPU-mediated I/O or a GPU-direct path, along with filesystem and storage-library settings. A change in path can be valuable, but it must be attributed to the path change rather than silently credited to the drive.
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]
Run the experiment in layers
- Inventory the system. Record drive model and capacity, drive count, interface and PCIe topology, server, CPU, memory, accelerator, filesystem, mount options, storage software, kernel, drivers and firmware.
- Capture the application I/O profile. Measure or log request sizes, read/write ratio, concurrency, queue depth, file counts, metadata operations and network traffic during a representative run.
- Prepare identical datasets. Use the same bytes, directory layout, file ordering and checkpoint contents on both configurations. Verify checksums where practical.
- Establish a stable baseline. Reboot or settle the system according to a documented procedure, apply the chosen cache policy, run warm-up iterations and then collect the baseline measurements.
- Run a storage microbenchmark. Use request sizes, read/write mixes and concurrency levels that resemble the application. This confirms that the device and path behave as expected; it does not prove that the AI task will be faster.
- Run the production-like AI workload. Keep model, data, software, accelerator settings, worker count and measurement interval fixed. Collect application, storage and system telemetry simultaneously.
- Repeat the baseline. Multiple baseline runs expose drift caused by thermal state, background jobs or remote-server variability. Do not use only the single best result.
- Switch to the candidate SSD configuration. Change the documented storage component or configuration, verify the same dataset and settings, and repeat the same warm-up and cache procedure.
- Repeat the candidate runs. Use the same number of repetitions and report the spread, median and relevant percentiles for each condition.
- Test expected scale. If deployment uses multiple workers, clients, GPUs or nodes, vary concurrency or node count and plot the resulting application and storage curves.
NVIDIA’s certification tests vary node and thread counts, I/O sizes, access modes and repeated runs. That approach is more informative than a single queue-depth result because it shows where scaling stops and which resource becomes dominant.
Free tools Windows power users keep installed
One-click scans. No signup required.
Design concurrency and scale tests deliberately
Start with one worker or client to establish a clear per-path result. Then increase concurrency in steps that match the intended deployment—for example, the number of data-loader workers, inference clients or nodes. Keep the workload per client fixed when measuring scale, and record aggregate and per-client throughput as well as latency. A result that improves at one queue depth but collapses at higher concurrency may indicate a host, filesystem, network or shared-server limit rather than an SSD limitation.
For each level, capture accelerator utilization and network use. If the GPU reaches full utilization while storage throughput continues to rise, storage may no longer be on the critical path. If aggregate I/O plateaus while clients queue and tail latency expands, identify the saturated link or shared component. Present the scaling curve instead of quoting only the peak drive bandwidth.
Calculate and present the result without overstating it
For a metric where higher is better, calculate relative change as (SSD result − baseline result) / baseline result × 100. For a metric where lower is better, such as epoch time or p99 latency, use (baseline result − SSD result) / baseline result × 100. State whether the values are means, medians or percentiles and include the number of runs and spread.
| Observed pattern | What it supports | What it does not support |
|---|---|---|
| Storage throughput and IOPS rise; epoch time falls and GPU idle time shrinks | The tested training path was at least partly storage-limited, and the candidate configuration improved the measured outcome. | A claim that every model or dataset will gain the same percentage. |
| Storage metrics rise; inference latency and throughput are unchanged | The application was limited elsewhere under those conditions, or data was already cached. | A claim that the SSD had no value in every deployment. |
| Average latency improves; p99 worsens at high concurrency | Typical requests got faster but tail behavior or queueing became a problem. | A claim of universally better user experience. |
| One-node results improve; multi-node results flatten at the same network rate | The network or shared storage service became the scaling limit. | A per-drive limit or a guarantee for a different topology. |
| Cold-cache gains are large; warm-cache gains are negligible | The SSD accelerates misses and loading, while the intended steady state is memory-resident. | A claim that normal warm-cache requests will be proportionally faster. |
Report negative or neutral results. A storage upgrade can still reduce startup time, improve checkpoint windows or provide capacity even when steady-state inference does not change, but those are separate outcomes and should be measured separately.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
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
Published figures that need careful qualification
Published demonstrations illustrate why scope matters; they are not universal uplift claims.
| Figure | Scope and limitation |
|---|---|
| About 270 GB/s | Samsung Memory Solutions Lab reported this for a particular 10-storage-node disaggregated-storage rack. The report describes six servers in two configurations, DSS v0.6 and NFS v4 with their respective operating systems and kernels. It is a system-level result, not per-drive bandwidth, and the article notes that testing with actual GPU servers was still needed. See the published test report. |
| 25× I/O, 3.4% inference-time overhead and 3× higher throughput | SNIA summarized these figures in 2023 for a Samsung PM9A3 under a specified challenging workload and compared with the MLPerf inference implementation. They should not be generalized to another SSD, workload or topology. See SNIA’s “Benchmarking Storage with AI Workloads.” |
When citing a published number, include the system scale, software, protocol, workload, date and whether it is a vendor or industry demonstration. Do not place a rack-level result beside a single-drive result without labeling the difference.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What to disclose so another engineer can reproduce it
| Category | Details to publish |
|---|---|
| Storage hardware | SSD model, firmware, capacity, number of drives, endurance class if relevant, RAID or volume layout and temperatures during runs |
| Topology | Server model, CPU and memory, accelerator model, PCIe generation and slot, NUMA placement and network link speed |
| Software path | Operating system and kernel, filesystem and mount options, storage library, drivers, container image, direct-I/O or GPU-direct setting |
| Workload | Model and version, dataset revision and size, file layout, object sizes, read/write mix, batch or sequence settings, checkpoint schedule and preprocessing |
| Execution | Client, worker and node count, concurrency, queue depth, warm-up, cache state, run duration, repetition count and system settling procedure |
| Results | Application metrics, storage throughput/IOPS/latency, accelerator and host utilization, network use, averages or percentiles, variability and failed runs |
Include the geography or deployment context when it affects the path—for example, a local server versus a remote disaggregated rack. A result from a single server does not establish behavior for a shared cluster.
Common failure modes and fixes
The SSD benchmark is fast but training time is unchanged
Check whether the data already fits in page cache, the accelerator is compute-bound, preprocessing is CPU-bound or the network is limiting throughput. Re-run with an explicitly documented cold-cache condition and inspect accelerator idle time. If the application remains unchanged, the correct conclusion is that storage was not the bottleneck for that training setup.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, 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 minuteResults vary widely between repetitions
Look for thermal throttling, background jobs, garbage collection, remote-server contention, changing cache state and asynchronous checkpoint activity. Standardize system settling, record temperatures and background load, increase repetitions and report the distribution instead of selecting a favorable run.
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).
Only one concurrency level was tested
Add a stepped concurrency test that reflects deployment. A drive can look excellent for one client yet provide little additional application throughput when the PCIe link, filesystem, network or shared storage service is saturated.
Sequential bandwidth was used as the headline result
Replace or supplement it with workload-shaped I/O and application outcomes. Inference, retrieval and KV-cache paths often use small or mixed requests; training checkpoints and streaming reads may behave differently. State request size and read/write mix for every storage number.
The baseline and SSD use different cache or I/O paths
Repeat both conditions with the same cache policy and transfer path. If the purpose is to compare paths, keep the comparison but name it accurately—for example, “network storage with host-buffered I/O versus local NVMe with GPU-direct I/O”—rather than attributing the entire difference to the SSD.
Recommended Free Tools
A compact reporting format
A useful result page can begin with a one-sentence verdict, followed by a table that separates application and storage evidence:
Workload: [model, dataset, fixed steps or requests]
Baseline: [device/path, filesystem, cache and I/O mode]
Candidate: [device/path, filesystem, cache and I/O mode]
Unchanged: [host, accelerator, software, workers, network, protocol]
Runs: [count; median and spread]
Application: [epoch time or throughput; latency percentiles]
Storage: [read/write throughput, IOPS, latency, request size and mix]
System: [GPU, CPU, memory and network utilization]
Verdict: [measured change and the bottleneck supported by the evidence]This format prevents a high synthetic score from being mistaken for an end-to-end AI speedup and makes clear whether the test measured a drive, a complete storage path or a changed system configuration.
Bottom line for an AI SSD comparison
Use the same workload and system, change only the declared storage configuration, control cache behavior, measure both application outcomes and storage metrics, and repeat at realistic concurrency. Treat NVIDIA’s workload profiles as test-design references, not promises; treat vendor and industry figures as scoped examples. The meaningful result is the change in the AI task under your stated conditions—not the largest sequential number printed on an SSD specification sheet.
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.
