Yes, several AI workstations can share one Synology NAS, and for a small lab it is the sensible design. One protected copy of each dataset and model replaces a copy on every machine, results land in one place, and backup is done once. It works well provided three things are planned: a network fast enough for several machines pulling large files at once, a folder structure that separates shared material from each person's work, and permissions that stop one project overwriting another. The NAS remains storage. Each workstation still does its own computing on its own fast internal drive.
What does a small AI lab look like?
| Component | Role | Typical choice for 3 to 6 people |
|---|---|---|
| AI workstations | Training, fine-tuning, inference | GPU workstations or compact AI desktops, one per person or shared |
| NAS | Datasets, model library, results, backups | 5 to 8 bays with redundancy |
| Switch | Connects everything | Multi-gigabit or 10GbE |
| Backup target | Second copy of the NAS | Second NAS or backup appliance, ideally elsewhere |
| UPS | Clean shutdown during a power cut | Sized for NAS, switch and at least one workstation |
| Firewall | Separates the lab from the office and internet | Business firewall with a dedicated segment |
How fast does the network need to be?
Size it for the busy moment, not the average. CALC. Four workstations each copying a dataset at 200 MB/s need 4 x 200 = 800 MB/s from the NAS at the same time. A single 2.5GbE link carries about 280 MB/s, so each machine would get roughly 280 / 4 = 70 MB/s. A 10GbE link carries up to about 1,100 MB/s in theory, which covers the 800 MB/s if the drives can supply it.
The practical guidance:
- The NAS uplink is the shared resource. Give the NAS the fastest port in the lab, 10GbE where possible.
- Workstations can be slower than the NAS. 2.5GbE per workstation with a 10GbE NAS uplink is a balanced starting point.
- Hard drives set a ceiling. An array of hard drives may deliver a few hundred MB/s to several readers. An SSD cache or SSD volume helps with many small files.
- Keep it wired. Wi-Fi is not suitable for dataset traffic.
These are working figures for planning. See 10GbE in a small office and 2.5GbE as the sweet spot.
How should the folders be organised?
| Shared folder | Contents | Access | Snapshots |
|---|---|---|---|
| datasets | Curated, versioned datasets | Read for all; write for the data owner | Daily |
| models | Approved base models | Read for all; write for one maintainer | Weekly |
| projects | One subfolder per project: code, adapters, checkpoints | Project members only | Several times a day |
| scratch | Temporary working files | All; no guarantees | None; auto-cleaned |
| results | Evaluation outputs and reports | Write for projects; read for all | Daily |
| backups | Workstation configuration and images | Administrators | By backup job |
Read-only access to datasets and base models is the single most useful rule. It prevents a stray script from altering the copy everyone depends on, and it limits what a compromised workstation can damage. A controlled models folder also supports the vetting described in AI model supply chain security.
Which protocol: NFS or SMB?
Linux AI workstations usually mount shares with NFS; Windows and Mac machines use SMB. A Synology NAS can offer the same folder over both. Two cautions: restrict NFS exports to the workstations' addresses, and keep user identities consistent across machines so file ownership means the same thing everywhere. Joining the NAS to the company directory, where there is one, keeps accounts in one place.
Sharing one NAS does not mean sharing one folder. The lab works when everyone reads from the same place and writes only to their own.
What should stay off the shared NAS?
- Active training reads. Copy the dataset for the current run to the workstation's NVMe; read speed from a shared hard-drive array will hold back the GPU.
- Live vector indexes and databases. Run them locally and back them up to the NAS.
- Two jobs writing to the same checkpoint folder. Give each run its own directory.
- Models being served. Load from local storage. See NAS vs local SSD for AI models.
How do you keep it fair and tidy?
- Quotas per user or project, so one experiment cannot fill the volume.
- A checkpoint retention rule, for example keep the last three and the best. See backing up models, datasets and checkpoints.
- Scheduled heavy jobs. Run backups and integrity checks outside working hours.
- One owner. Someone is responsible for the models folder, the quotas and the restore test.
If the lab shares GPUs as well as storage, see GPU partitioning with MIG and vGPU.
Buying in Singapore
NAS: the Synology DS1525+ 5-bay (S$1,800) or DS1825+ 8-bay (S$2,148); in a rack, the RackStation RS2423+ 12-bay (S$4,002). 10GbE for the NAS on models with a compatible slot: the Synology E10G30-T2 dual-port card (S$387.60). Switch: the HPE Networking Instant On 1960 12-port 10GBASE-T (S$2,444.40). Workstations: the NVIDIA DGX Spark 4TB (S$9,450) or a workstation with the NVIDIA RTX PRO 6000 Blackwell Workstation Edition (S$29,100). Segmentation: the Fortinet FortiGate 90G (S$3,796.80) (listed 11 Nov 2026; prices change; ask for a current quotation). To start, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Can multiple AI workstations share one Synology NAS? Yes. Several workstations can mount the same shared folders for datasets, models and results. It works well with a fast NAS network link, a clear folder structure and read-only access to shared material.
How fast should the network be for a shared AI NAS? Give the NAS a 10GbE uplink where possible, since it serves every machine at once. Workstations on 2.5GbE or 10GbE is a balanced design for a small lab.
Should training data be read directly from the shared NAS? For light work it can be. For sustained training, copy the dataset to the workstation's local NVMe drive so the GPU is not waiting on shared network storage.
How do you stop users overwriting each other's work? Use separate project folders with member-only access, keep datasets and base models read-only, set quotas and enable snapshots so earlier versions can be restored.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
