Start Free

GPU & STORAGE FILE TRANSFER

Send a GPU server the data it needs,
and put the results back in storage.

Transfer the datasets and models held in object storage, on a NAS or in a data centre to your GPU servers.
The checkpoints and results produced during training and inference can travel back to storage as well. Your existing AI infrastructure stays as it is, while file movement between storage and GPU compute is managed on one platform.

Flow canvas60%

DatasetVision Training Set v12GPUGPU-NODE-01GPU-NODE-02GPU-NODE-03

1 dataset, 3 / 3 GPU nodes1 step

AI INFRASTRUCTURE INTEGRATION

Your AI infrastructure stays as it is.
INNORIX moves the data.

The systems that decide compute and workload in your AI infrastructure stay exactly as they are.

INNORIX delivers datasets and models between the source and target those systems name, and sends checkpoints and results back to storage.

Your AI infrastructureINNORIX
GPU provisioningDelivering datasets and models
Workload schedulingThe file transfer itself
Training and evaluation pipelineSending checkpoints and results
Artifact metadata and lineageTransfer state and results

INNORIX does not manage GPU scheduling or the AI pipeline. It moves the files your existing systems decide on.

STORAGE & GPU CONNECTIVITY

One transfer layer for the data path
between storage and GPU compute.

Whether the dataset and the model sit in cloud object storage, in a data centre or in private storage, they can be delivered to the GPU environment the work runs in.

Data moves between different clouds, regions and on-premises environments while the storage and GPU infrastructure stay as they are. The route and the target are set to match the environment — a cloud GPU, a data centre GPU cluster, or your own.

Cloud to GPUObject storage in one region, delivered to a GPU cluster in another.

Connected

SOURCEObject StorageUS West, S3

training-dataset2.4 TB
foundation-model48 GB
image-dataset820 GB
auxiliary-data12 GB
Transferring

68%

1.63 TB of 2.4 TB transferred

TARGETCloud GPUAsia, NVIDIA H100

NVIDIA H100Available
NVIDIA A100Available
NVIDIA L40SAvailable
NVIDIA L4Available
training-dataset.tar2.4 TB, 12,480 filesTransferring
68%

1.63 TB of 2.4 TB transferred

AI DATA LIFECYCLE

From dataset to model, checkpoint and result,
the same data delivery.

AI training starts with getting the dataset and the model into the GPU environment, and carries on through storing the checkpoints it produces and collecting the final result.

INNORIX moves the files each stage needs between storage and the GPU environment, and automates the data movement that keeps recurring. Preparing the files for training, keeping the intermediate results and delivering the final ones are all connected, so file transfer works the same way across the whole AI workflow.

Training run #4821resnet-vision, gpu-cluster-01, started 09:52

  1. DatasetTraining data ready

    training-dataset.tar2.4 TB, 12,480 files

  2. ModelModel delivered

    foundation-model48 GB, 12 files

  3. AI TrainingRunning

    Epoch 64 / 10064%

    GPU Cluster8 GPUs

  4. CheckpointTraining state stored

    checkpoint-064320 GB, 64 files

  5. ResultResult collected

    inference-result120 GB, 3,420 files

Checkpoints repeat, so a long run is never lost

training-dataset.tar2.4 TB, 12,480 filesObject StorageGPU ClusterCompleted100%

DATA READINESS & RECOVERY

Less time with a GPU waiting for data,
and an interrupted transfer carries on.

The data AI training and inference need ranges from single files of several terabytes to millions of small images and documents. INNORIX takes the load that size and count create off the transfer, and delivers large data to the GPU environment reliably.

If the network fails or the transfer stops, what has already completed is kept and only the missing part is sent again. Automatic resume, retries and verification cut out needless repeat transfers and confirm that the data AI needs is ready.

  • Large transfers

    A terabyte dataset in a single transfer

  • Many small files

    Millions of files, handled reliably

  • No lost progress

    What finished is kept; only the rest is sent again

Transfer #4821-01Object Storage → gpu-cluster-01

training-dataset.tar2.4 TB, 12,480 filesCompleted
100%

2.4 TB of 2.4 TB transferred

Parts of this transfer

part 0001–40001.01 TBDone
part 4001–80000.67 TBDone
part 8001–124800.72 TBDone
  • Completed data keptNot sent again after a stop
  • Automatic resumeContinues when the connection returns
  • Verified and losslessData is verified after transfer

ORCHESTRATOR & API

Your AI orchestrator runs the transfer
at the moment it needs to.

File transfer connects to your existing AI orchestrator, scheduler and pipeline without changing them. Name the files and the target from Kubeflow, Airflow, SLURM or your own system through the API, and INNORIX runs the transfer and returns its progress and its result.

Your AI system schedules and runs the work exactly as it does today, while INNORIX handles the file movement between GPU clusters, cloud storage and the data centre. Completion callbacks and run results feed back into the workflow, so data preparation, training and result collection carry on automatically.

  • Simple to connect

    A standard API, connected straight away

  • Nothing to change

    Used without touching the AI system

  • Results you can see

    Status, logs and a completion callback

Transfer APItr_8f3a2e6b, called by airflow-dag-vision

Callback sent
  • AI OrchestratorKubeflow, Airflow, Custom
  • SchedulerSLURM, PBS, LSF
  • AI ApplicationTraining, Inference, Fine-tuning
INNORIXRuns the transfer
  • High-Speed Transfer
  • Resume & Retry
  • Verification
  • GPU ClusterNVIDIA H100, A100
  • Cloud StorageS3, GCS, Azure Blob
  • Data CenterNAS, Object Storage

POST/api/v1/transferscURL

1{2  "source": "s3://training-dataset",3  "target": "gpu-cluster-01",4  "files": "dataset/*.tar",5  "callback": "https://ai.company.com/webhook"6}

RESPONSEJSON

1{2  "transferId": "tr_8f3a2e6b",3  "status": "completed",4  "files": "12,480",5  "size": "2.4 TB",6  "progress": "100%",7  "callback": "delivered"8}

Run status

  1. Requested10:14
  2. Transferring10:15
  3. Completed10:33
  4. Callback sent10:33

UNIFIED OPERATIONS

Run every file transfer between GPU and storage
from one screen.

Manage transfers as different as dataset delivery, checkpoint collection, model movement and result gathering in INNORIX Studio.

Check what is running and what has finished, and follow up on a failed transfer over only the range that needs it.

Recent Runs

  • Training Dataset DeliveryCompleted
  • Checkpoint CollectionTransferring
  • Model DistributionCompleted

Faster than an explanation: try it yourself.

Connect the devices and build the distribution flow.

Add flowCancelSave draftSave & active

Flow canvas

Daily · 02:00HQHQ-BUILD-01TOKYO-EDGE-07EdgeTOKYO-EDGE-01OSAKA-EDGE-02NAGOYA-EDGE-03SAPPORO-EDGE-04FUKUOKA-EDGE-05TOKYO-EDGE-06

0 devices · 1 step

Configuration

Start when

NowSchedule

Every day · 02:00

Source

HQ-BUILD-01

Target

Choose target

GET STARTED

Now put it to work in your application.

TECHNICAL FAQ

Frequently asked questions

Can files be transferred between a GPU server and object storage?

Yes. Connect supported object storage and GPU servers to transfer datasets, models and other files. The connection conditions for the storage and the GPU servers have to be confirmed for your environment.

Can an on-premises GPU server be connected to cloud storage?

Supported on-premises servers and cloud object storage can be set as source or target.

Can data on a NAS or an ordinary file server be sent to a GPU server?

A supported file server or NAS environment can be connected to deliver data to a GPU server. The actual connection method depends on the operating system, the file system and the network configuration.

Can models and checkpoints be transferred as well as datasets?

Yes. Datasets, model weights, checkpoints, evaluation data, results and other files used or produced by AI work can be transferred.

Can it handle large image datasets or TB-scale files?

INNORIX supports both very large files and large numbers of small files. The transfer performance you actually get depends on the file composition, the storage API, the network and the device specifications.

Can checkpoints produced on a GPU server be collected automatically?

Checkpoint files can be set up to go to storage on a schedule, on an event or at the request of an external system. Whether a checkpoint exists at all, and which files are selected, is decided by the existing AI system and the transfer settings.

Does INNORIX perform GPU scheduling or AI training itself?

No. GPU provisioning, scheduling, training and the AI pipeline are handled by existing infrastructure. INNORIX carries out the file movement that work needs.

Every one of these capabilities,on a single platform.

The capability on this page is one of the many file transfer features INNORIX Platform provides. From file transfer to automation and unified operations, try every feature free.

Start Free

Looking at data delivery for AI?

We'll go through the transfer setup you need for the storage and GPU environment you already run.

  • Review of storage and GPU environment connections
  • Setting up large dataset transfers
  • Connecting to an existing AI pipeline

We use these details only to answer your inquiry. See our Privacy policy.