An NVIDIA DGX Spark gives a small team a private AI computer: 128GB of unified memory on the desk, running language models without sending data outside. What it does not give you is somewhere safe and roomy to keep everything around the model: the source documents, the datasets, the library of model files, the fine-tuned adapters and the backups. A Synology NAS fills that gap. The DGX Spark does the thinking on its fast internal drive; the NAS holds the shared data, keeps point-in-time snapshots and backs up the Spark itself. Connected over a fast wired network, the two make a complete local AI setup for well under the cost of a single GPU server.
What does each device do?
| NVIDIA DGX Spark | Synology NAS | |
|---|---|---|
| Job | Run and fine-tune AI models | Store, share and protect data |
| Strength | GPU compute and 128GB unified memory | Capacity, redundancy, snapshots, backup |
| Storage | Fast internal NVMe, limited capacity | Many terabytes on protected drives |
| Holds | Models in use, active index, working data | Model library, datasets, documents, backups, archives |
| If it fails | AI service stops; data is safe if it was on the NAS | A drive can fail without data loss; a second copy covers the rest |
The NAS does not run the model and will not speed up inference. We set out the division of labour in why your AI server needs a NAS.
How should the two be connected?
By cable, through a switch, on the same local network. The DGX Spark has a 10 Gigabit Ethernet port, so the link is limited by the NAS and the switch. Three practical options:
| Setup | Practical speed | Suits |
|---|---|---|
| NAS and switch at 1GbE | about 110 MB/s | Documents only; painful for model files |
| NAS at 2.5GbE with a multi-gigabit switch | about 280 MB/s | One Spark, occasional model changes |
| NAS with a 10GbE card and a 10GbE switch | several hundred MB/s, limited by the drives | Regular dataset work, more than one AI system |
CALC. Copying a 60GB model file from NAS to Spark: at 110 MB/s, 60,000 / 110 = 545 seconds, about 9 minutes. At 280 MB/s, 214 seconds. At 600 MB/s, 100 seconds. Speeds are typical working figures. Wi-Fi is not a sensible link for this job. See 2.5GbE as the SME sweet spot and 10GbE in a small office.
How do you set it up, step by step?
- Give both devices fixed addresses. Reserve an IP address for the NAS and for the Spark so the connection survives a reboot.
- Create a storage pool with redundancy. Use a RAID level that tolerates a drive failure, on NAS-rated drives.
- Create shared folders by purpose. For example: models, datasets, documents, outputs, backups. Separate folders allow separate permissions and snapshot schedules.
- Enable NFS for the Spark. The Spark runs a Linux-based system, and NFS is the natural protocol. Restrict each NFS rule to the Spark's address. Keep SMB enabled for staff PCs that add documents.
-
Mount the shares on the Spark. An example, with your own address and share name:
sudo mount -t nfs 192.168.10.20:/volume1/models /mnt/nas/models. Add the mount to the system's startup configuration so it returns after a reboot. - Copy, don't run, from the NAS. Keep every model on the NAS and copy the ones in use to the Spark's internal drive. Point the serving software at the local copy.
- Turn on snapshots. Frequent on documents and outputs, less frequent on the model library. Protect snapshots from deletion where the model supports it.
- Back up the Spark to the NAS. A scheduled job that copies configuration, containers, the vector index, adapters and chat data.
- Back up the NAS somewhere else. A second NAS, a backup appliance or an off-site target.
- Test a restore. Recover one folder and one model before you rely on any of it.
Menu names and steps vary by software version; follow the current Synology and NVIDIA documentation for the exact screens and commands.
What goes where?
| Data | On the Spark | On the NAS |
|---|---|---|
| Models being served | Yes | Master copy |
| Models not in use | No | Yes |
| Company documents for RAG | No | Yes, with permissions |
| Vector index | Yes, for speed | Backup copy |
| Training dataset for the current run | Yes, while training | Master copy |
| Checkpoints and adapters | Latest | All retained versions |
| Logs and evaluation results | Recent | Archive |
The reasoning is in NAS vs local SSD for AI models, and the Spark's own capacity choices in 1TB vs 2TB vs 4TB GB10 storage.
The Spark is the engine. The NAS is the fuel tank, the logbook and the spare key. A setup with only the engine runs well until the first thing goes wrong.
What can go wrong?
- Serving models straight from the NAS. Loading is slow and the AI service now depends on the network.
- One big shared folder. It makes permissions and snapshots all-or-nothing.
- Treating the NAS as the backup of itself. RAID covers a failed drive, not fire, theft or ransomware.
- No power protection. A power cut during a write can damage data on either device. A small UPS for both is inexpensive insurance.
- Open access. The NAS holds the company's documents. Give the Spark and each user only the folders they need.
Buying in Singapore
AI compute: the NVIDIA DGX Spark 4TB (S$9,450), or the same GB10 platform as the ASUS Ascent GX10 1TB (S$10,350) or HP ZGX Nano G1n 4TB (S$11,499). Storage: the Synology DS925+ 4-bay (S$1,260) for one AI system, or the Synology DS1825+ 8-bay (S$2,148) for a team. For 10GbE on NAS models with a compatible slot, the Synology E10G18-T1 card (S$202.80); confirm compatibility for your model. Drives: Seagate IronWolf Pro from S$390. Switch: the HPE Networking Instant On 1960 12-port 10GBASE-T (S$2,444.40). Second copy: the Synology ActiveProtect DP340 (S$6,526.80) (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 a Synology NAS be used with an NVIDIA DGX Spark? Yes. The NAS shares folders over the network, and the DGX Spark mounts them, typically using NFS. The NAS holds datasets, documents, the model library and backups while the Spark runs the models.
Should AI models run directly from the NAS? It is better to keep the master copy on the NAS and copy the models in use to the Spark's internal NVMe drive, so loading is fast and the service does not depend on the network.
What network speed is needed between the NAS and the DGX Spark? 2.5GbE is workable for one system with occasional model changes. For regular dataset transfers or several AI systems, 10GbE on the NAS, switch and Spark saves significant time.
Does the NAS replace a backup? No. RAID and snapshots on the NAS protect against drive failure and mistakes, but a second copy on another device or off site is needed to cover loss of the NAS itself.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
