Free tools Windows power users keep installed
One-click scans. No signup required.
Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
Estimate AI storage by adding separate allowances for input and processed data, model files, checkpoints, operational files, backups, and growth—not by looking only at the raw dataset or model size. Then check whether the storage can sustain the reads and writes your jobs require: a system can have enough capacity and still run slowly or miss checkpoint deadlines.
What should an AI storage estimate include?
Build the estimate from the workload’s stored files, then add the space needed for simultaneous jobs, retained versions, recovery, and growth. Keep each category separate until you have checked what is already shared or duplicated; the same dataset may occupy storage in more than one place if jobs create local copies or processed versions.
As an Amazon Associate I earn from qualifying purchases.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →- Input and processed data: raw source data, cleaned or sharded copies, feature stores, embeddings, and temporary preprocessing output.
- Model artifacts: model weights, tokenizer and configuration files, adapters, and evaluation artifacts.
- Checkpoints and recovery: one checkpoint’s size multiplied by the number of versions retained, with room for replication and restart needs.
- Operations: logs, traces, experiment metadata, cached batches, and backups.
- Growth and concurrency: simultaneous runs, additional model versions, and expected increases in data volume.
For each category, record the estimated size, how many copies or versions are needed, how long they are kept, and whether the data is shared between jobs. Add the resulting amounts for a capacity estimate. If a file is temporary, include its peak overlap with other files: a preprocessing step can need space for both its inputs and its outputs before the inputs are removed.
Storage capacity is not the same as accelerator memory. Disk or object storage holds datasets and artifacts over time; accelerator memory must also accommodate the model and the live workload, including activations and, for inference, the KV cache. A model’s weight size alone therefore does not determine either total storage or the hardware needed to run it.
#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
How do I estimate dataset storage for training?
Start with the data as it will actually be presented to the workers, not just the source archive. Count raw data, transformed data, any stored features or embeddings, and temporary files that coexist during preprocessing. Then account for the number of jobs that need separate copies and whether your storage setup lets them share a common dataset.
Use published figures as a starting point, not a quota
Google Cloud’s 2026 TPU guidance gives the following initial storage estimates per TPU:
| Workload | Dataset storage per TPU | Checkpoint storage per TPU |
|---|---|---|
| LLM pre-training | 2 TB | 200 GB |
| Multimodal training | 12 TB | 1 TB |
| Inference | 1 TB | 1 GB |
These are planning starting points, not universal requirements or a guarantee that a given job will fit. To form a first-pass estimate, multiply the per-TPU figure by the number of TPUs in use; then adjust for copies, concurrent jobs, retention, and measured data characteristics. Do not treat that multiplication as a complete storage design: it does not replace measuring your actual input representation, preprocessing output, or sharing and replication needs.
Training repeatedly rereads data. NVIDIA describes rereading as the key I/O operation in deep-learning training, which is why an input set that fits comfortably on disk can still be a bottleneck if it cannot be read quickly enough. The amount stored and the rate at which it can be delivered are separate questions.
How do I calculate checkpoint storage and bandwidth?
Estimate one checkpoint first, multiply by the versions you retain, and then calculate how quickly the system must write it. A large storage pool can hold many checkpoints but still be too slow to save one within the interval your training process allows.
Estimate a training checkpoint
Google’s checkpoint method starts with parameter count multiplied by bytes per parameter. Its approximate planning figure is 12–16 bytes per parameter for FP16 plus optimizer state, followed by an approximately 3× buffer for optimizer states and mixed precisions. This is a sizing method for training checkpoints, not a universal formula for every model, framework, precision, or checkpoint format.
Google’s Qwen3-72B example yields approximately 864 GB for a checkpoint, approximately 2.5 TB after the buffer, and about 20 GB/s of required checkpoint bandwidth when saving every two minutes. The bandwidth follows the basic calculation:
Required checkpoint write bandwidth = buffered checkpoint size ÷ checkpoint interval
Rank #2
- BREAKTHROUGH PCIe 5.0 PERFORMANCE: Supercharge your workflow and gaming with PCIe 5.0, boasting up to 14,700/13,400 MB/s¹ sequential read/write speeds. Tackle massive files and power up your gaming with Gen5—twice as fast as the 990 PRO SSD.
- EVERY TASK, TURBOCHARGED: Speed past productivity limits. With random read/write speeds up to 1,850K/2,600K IOPS¹, enjoy fast game loads, seamless AI apps, and efficient multitasking. Virtually no lag, no limits—just nonstop performance.
- THINK FAST, CREATE FASTER: With random read/write speeds of up to 1,850K/2,600K IOPS¹, the 9100 PRO SSD fuels seamless AI content creation, swift loads, and smooth gameplay. Work, play, and create at lightning speed.
- SPEED, WHENEVER YOU NEED: From laptops, desktop PCs, to PS5 with added heatsink devices², experience blazing PCIe 5.0 speeds and up to 8TB of storage. Perfect for video editing, gaming, and creative tasks, with the compatibility to match your device.
- STAY COOL, RUN FAST: Push limits, not temperatures. A 5nm controller boosts power efficiency up to 49% over the 990 PRO SSD³, while advanced thermal control keeps performance smooth and reliable.
For that example, a buffer of about 2.5 TB divided by 120 seconds is roughly 20 GB/s, allowing for rounding and unit conventions. Use the size that must actually be written in the interval—not merely the unbuffered checkpoint figure—when assessing whether a storage path can keep up.
Convert one checkpoint into a retention estimate
If you retain several recovery points, multiply the space for one saved checkpoint by the number of retained versions. Add any separate copy kept for replication or recovery, and consider whether a new save temporarily overlaps with older files. The retention count is a policy choice: more versions provide more rollback points but consume more capacity.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsAlso consider restart and recovery behavior. A checkpoint is useful only if the job can find and read it when needed; estimate read performance as well as write performance, particularly if a recovery must restore a large file before training resumes.
How much disk space does a 70B model need?
There is no single disk-space figure implied by a model’s parameter count. AWS’s approximate model-weight figures show how strongly precision affects weight memory, but they are memory estimates for weights, not a complete disk-space estimate for an AI workload.
| Model size | FP16 model-weight memory | FP8/INT8 model-weight memory | INT4/NVFP4 model-weight memory |
|---|---|---|---|
| 7B | 14 GB | 7 GB | 3.5 GB |
| 13B | 26 GB | 13 GB | 6.5 GB |
| 70B | 140 GB | 70 GB | 35 GB |
For a 70B model, the table’s weight-memory estimates range from 35 GB at INT4/NVFP4 to 140 GB at FP16. A deployment or training setup may also need tokenizer and configuration files, adapters, evaluation artifacts, other model versions, checkpoints, and the data and operational files identified above. The table does not establish the exact on-disk size of any particular download or checkpoint.
For training, do not substitute the model-weight estimate for checkpoint sizing. A checkpoint may contain optimizer state and other training state; Google’s approximate 12–16 bytes per parameter method and its additional buffer address that broader checkpoint estimate. For inference, weight placement is only one part of live memory demand; context length and concurrent requests affect KV-cache needs.
How much storage does inference need for long context and high concurrency?
Inference sizing depends on more than the weight files. AWS says to include model weights, activations, KV cache, runtime overhead, request lengths, concurrency, and latency objectives when sizing storage and memory. Longer contexts and more simultaneous requests increase KV-cache demand.
AWS gives the following KV-cache relationship:
KV-cache size = 2 × KV-data-type bytes × number of layers × number of KV heads × head dimension × context length × batch size
Rank #3
- 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
The factor of 2 accounts for the key and value data in the cache. The other terms make the workload dependencies explicit: model architecture, cache precision, context length, and batch size all matter. AWS’s Mistral-7B example estimates 1.95 GB of KV cache for one concurrent request at a 16,000-token context, and 7.81 GB for four concurrent requests. Those are example KV-cache figures, not complete memory or disk totals for serving the model.
When estimating inference storage, keep persistent model files and operational data separate from working memory. A longer context or higher concurrency may require more live cache capacity even if the stored weight files do not change. Include the expected request lengths and concurrency in capacity planning for the runtime, then separately size the durable storage used for model versions, logs, backups, or other retained artifacts.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Do I need NVMe, a NAS, object storage, or a parallel filesystem?
Choose storage by access pattern rather than by a single label. Durable storage is suited to retained datasets and artifacts; local NVMe is useful for fast cache or staging; shared high-throughput storage is valuable when multiple workers need concurrent access or the dataset is too large for local cache.
| Storage type | Useful role in an AI workload | Planning consideration |
|---|---|---|
| Object storage | Durable datasets, checkpoints, model weights, backups, and cross-job sharing | Plan for how jobs access data and how copies or retained versions affect capacity. |
| Durable block storage | Persistent volumes and disks attached to active jobs | Match the volume and its performance to the job’s ongoing read and write needs. |
| Managed parallel filesystem | High-throughput shared access when datasets exceed local cache or many workers read concurrently | Assess sustained shared throughput under the intended worker load. |
| Local NVMe | Low-latency cache and staging for training data, checkpoints, and model artifacts | Treat it as a cache or staging tier where appropriate, not automatically as the only durable copy. |
| Repository or bucket platform | Mutable checkpoints, logs, intermediate files, scratch data, and rolling backups | Check file-count, file-size, and version-history limits as well as total capacity. |
Google lists durable block storage, Cloud Storage, and Managed Lustre as storage options for TPU workloads. Hugging Face documents bucket use for checkpoints, logs, preprocessing intermediates, agent scratch data, and rolling backups. It recommends chunking files below 200 GB and keeping total file counts manageable. Those operational recommendations matter when a workload produces many intermediate or checkpoint files, not just when its total byte count is large.
NVIDIA identifies local NVMe as a cache or staging tier in its DGX guidance. That makes it a complement to durable or shared storage in many designs: stage data close to the workers for repeated reads, while retaining needed source data and recovery artifacts in the appropriate durable location.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What storage bandwidth does model training require?
There is no single bandwidth target for every model or storage system. The required rate depends on how quickly workers consume training data and how often checkpoints must be written; measure sustained throughput for the intended workload rather than relying on a device’s headline or peak figure.
Recommended Free Tools
For checkpoint writes, divide the checkpoint size by the allowed save interval. For input reads, account for repeated passes through the data and the number of workers reading at once. A shared store must serve the aggregate demand, while a cache can reduce repeated remote reads only if the needed data fits and is actually reused.
NVIDIA’s B200 guidance gives these reference targets:
Rank #4
- GROUNDBREAKING READ/WRITE SPEEDS: The 990 EVO Plus features the latest NAND memory, boosting sequential read/write speeds up to 7,250/6,300MB/s. Ideal for huge file transfers and finishing tasks faster than ever.
- LARGE STORAGE CAPACITY: Harness the full power of your drive with Intelligent TurboWrite2.0's enhanced large-file performance—now available in a 4TB capacity.
- EXCEPTIONAL THERMAL CONTROL: Keep your cool as you work—or play—without worrying about overheating or battery life. The efficiency-boosting nickel-coated controller allows the 990 EVO Plus to utilize less power while achieving similar performance.
- OPTIMIZED PERFORMANCE: Optimized to support the latest technology for SSDs—990 EVO Plus is compatible with PCIe 4.0 x4 and PCIe 5.0 x2. This means you get more bandwidth and higher data processing and performance.
- NEVER MISS AN UPDATE: Your 990 EVO Plus SSD performs like new with the always up-to-date Magician Software. Stay up to speed with the latest firmware updates, extra encryption, and continual monitoring of your drive health–it works like a charm.
| System configuration | Workload tier | Read target | Write target |
|---|---|---|---|
| Single-system-unit | Standard | 40 GB/s | 20 GB/s |
| Single-system-unit | Enhanced | 125 GB/s | 62 GB/s |
| Four-unit system | Standard | 160 GB/s | 80 GB/s |
| Four-unit system | Enhanced | 500 GB/s | 250 GB/s |
These are B200 reference targets, not general requirements for all AI workloads or a guarantee that a particular storage system will deliver them. Compare them with your measured read and write rates under the expected worker count, file sizes, and concurrency. Also consider recovery time: a system that writes checkpoints adequately may still restore them too slowly for operational needs.
How do I build a practical estimate?
Use a small worksheet with explicit assumptions. Replace starting estimates with measured sizes and observed access rates as soon as you have representative workload data.
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 →- Define the workload: record whether it is pre-training, multimodal training, or inference; the number of accelerators and concurrent jobs; and the model versions and data formats involved.
- Inventory persistent inputs: count raw data plus cleaned or sharded copies, feature stores, embeddings, and retained preprocessing output.
- Inventory artifacts: include weight files, tokenizers, configurations, adapters, evaluation files, and any additional model versions required at once.
- Size checkpoints: estimate one checkpoint using the format and training state in use, then multiply by retained versions and add any separately stored recovery copy.
- Account for operational overlap: include logs, traces, experiment metadata, cached batches, backups, and temporary output that coexists with its inputs.
- Add concurrency and growth: include simultaneous jobs, planned model versions, expected data growth, and replication requirements.
- Check capacity and throughput separately: calculate total stored bytes, sustained reads for repeated training data access, and checkpoint write bandwidth for the save interval.
- Choose tiers: place shared durable data and retained recovery artifacts appropriately, and use local cache or staging where it helps serve workers at the required rate.
- Validate with a representative run: measure actual data sizes, read rates, checkpoint duration, and recovery behavior under the worker count and concurrency you expect.
A useful estimate states its assumptions next to the totals: data-copy count, retention period or version count, concurrent job count, checkpoint interval, and the growth allowance used. If any of those assumptions changes, update the relevant bucket rather than applying an arbitrary multiplier to the entire storage pool.
Frequently Asked Questions
Can I estimate AI storage from parameter count alone?
No. Parameter count can help estimate model-weight memory and, with additional assumptions, training checkpoint size, but it says nothing by itself about datasets, retained checkpoints, logs, preprocessing output, backups, or simultaneous jobs.
Does the dataset need to fit entirely on local NVMe?
Not necessarily. The storage design can combine local cache or staging with durable object or block storage, or shared parallel storage when workers need high-throughput access to data larger than the local cache.
Why can a workload run slowly when there is plenty of free disk space?
Free capacity does not show how quickly data can be read or checkpoints written. Training repeatedly rereads data, and checkpoint saves have an interval deadline, so sustained throughput and concurrent access can constrain a job even when total capacity is ample.
Crashes, 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 minuteWindows 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 reinstallShould I count cache files as permanent storage?
Count them when they coexist with other files and contribute to peak capacity, but distinguish cache from data you must retain for recovery or later runs. The distinction helps prevent a temporary staging copy from being mistaken for a durable backup.
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.
