How to Connect DGX Station GB300 to Enterprise NAS and High-Speed Storage

For most offices, connect the DGX Station GB300's 10GbE port to the NAS over a dedicated storage network, mount the shares with NFS, and copy whatever the GPU is actively using onto local NVMe. Move to 25GbE or 100GbE on the Station's ConnectX-8 ports only when large model and checkpoint transfers start to dominate the working day. NVIDIA hasn't published which switches, optics or breakout cables it validates for those ports, so confirm the exact parts with the switch and OEM vendor. GPUDirect Storage, NVIDIA's direct path from storage to GPU memory, works on the Station's local NVMe in a published review. It is not documented for NFS shares on the Station, it doesn't cover SMB, and NVIDIA's own docs flag a conflict with Secure Boot, which the Station enables by default. SourceIT has not run these integrations on a Station. The steps below follow NVIDIA, Red Hat and StorageReview documentation, with our calculations labelled CALC.

How long does a model transfer take at each link speed?

Source: our calculation (CALC) at line rate, with no protocol overhead and no storage limit; real throughput depends on the NAS. Port list from NVIDIA's DGX Station product page (2x QSFP112 at 400 Gb/s, 1x 10GbE RJ45, 1x 1GbE BMC). Local NVMe reference: StorageReview MSI XpertStation WS300 review, 24 Aug 2026 (13.1 GiB/s, about 14.1 GB/s). Facts as of 9 Nov 2026.

Which port should the NAS connect to?

The Station has three network interfaces: a 10GbE RJ45 port, a 1GbE port reserved for the BMC management controller, and two QSFP112 ports on the ConnectX-8 SuperNIC at 400 Gb/s each. The 10GbE port is the practical first connection to an office NAS and switch. Keep the BMC port on a separate management network; our security hardening checklist covers why.

The QSFP112 ports are built for 400G and for linking Stations. NVIDIA's generic ConnectX-8 documentation says its Ethernet side complies with standards from 10GbE to 400GbE and supports port splitting, but which lower speeds, breakouts and transceivers are validated on the Station is not published. Before buying a 25G or 100G switch for those ports, get the exact cable and optic part numbers confirmed. Our ConnectX-8 networking post explains the ports in more depth.

NFS or SMB?

Use NFS for data the GPU reads. NVIDIA's GPUDirect Storage release notes list NFS among supported file systems, along with BeeGFS, EXT4, XFS, VAST, WekaFS and others, and do not list SMB or CIFS. SMB data therefore goes through the CPU and its page cache (our inference), which is fine for documents and user files but not the fastest path for large model loads. Most business NAS units serve both protocols, so keep SMB for people and NFS for the Station.

Get identity right before you mount anything. Files on the NAS belong to user and group IDs, and the Station's Linux accounts must map to the same IDs, either through a directory service or NFSv4 identity mapping configured the same on both ends. If they don't match, jobs fail with permission errors or, worse, write files that other users can read. NVIDIA publishes no Station guidance for LDAP or single sign-on, so follow your NAS vendor's Linux client documentation.

Should you use NFS over RDMA?

Only if your NAS supports it and you need the extra speed. NFS over RDMA needs a running NFS service, an InfiniBand or RoCE device and an IP address on it. Red Hat's documentation gives the default port as 20049 and shows a client mount like this:

mount -o rdma server.example.com:/nfs/projects/ /mnt/

Afterwards, check that the mount output shows proto=rdma. These steps are documented for Red Hat Enterprise Linux. The Station ships Ubuntu 24.04, and Ubuntu-specific steps with ConnectX-8 are not covered by that source (HP says its ZGX Fury is certified for RHEL). RoCE also needs a lossless network design on the switch, and NVIDIA's two-Station playbook uses jumbo frames at MTU 9000. Whatever MTU you choose, set it the same on the Station, the switch and the NAS, or large transfers will stall.

Rule of thumb: copy what the GPU reads every hour to local NVMe, mount what it reads every day from the NAS, and only pay for faster networking when the copying itself becomes the bottleneck.

How should data flow between the NAS and the Station?

Treat the NAS as the source of truth and local NVMe as a working copy. A simple pattern:

• Before a job: copy the model and the active dataset from the NAS to local NVMe, for example with rsync, so the GPU never waits on the network.
• During training: write checkpoints locally, then copy each finished checkpoint to the NAS with a scheduled job, and delete old local copies under your retention rule.
• For RAG: mount document shares read-only and index from them, as described in our RAG guide.
• For serving: keep the model library on the NAS and stage the models you serve; our vLLM guide covers caching locations.

Put storage traffic on its own VLAN so large copies don't slow the office network, and size the NAS with our storage planning worksheet. For why the network can matter more than the GPU in a company AI project, see our networking post.

Does GPUDirect Storage work on the Station?

Partly, with caveats. StorageReview ran NVIDIA's gdsio tool on the MSI WS300 and measured 13.1 GiB/s reads from one local Gen5 NVMe drive. NVIDIA's GDS 1.20 release notes say Grace-based platforms are supported but don't name the GB300 Station, and say that on Linux kernel 6.17 or newer, DOCA 3.3.0 or newer is required. The Station ships kernel 6.17 with DOCA-OFED pre-installed, so check the DOCA version. NVIDIA's verification tool is /usr/local/cuda-<x>.<y>/gds/tools/gdscheck.py -p, and its sample output shows NFS in compatibility mode until NFS over RDMA is set up.

Two more cautions. NVIDIA's GPU Operator documentation says GDS "is not supported with secure boot", and Secure Boot is on by default on the Station, so you may have to choose. And the DGX OS documentation's extra driver packages and rebuild steps for GDS apply to the older DGX Station A100 and A800, not the GB300, so don't follow them on a GB300 Station.

Buying in Singapore

For a 10GbE start (listed 9 Nov 2026): the Synology RackStation RS822RP+ with redundant power is S$2,086.80, the QNAP TS-873AEU-RP 8-bay rackmount NAS is S$3,010, and the D-Link DXS-1210-10TS 10GbE switch is S$1,680. To step up to 25GbE, the D-Link DXS-1210-28T adds four 25G SFP28 ports for S$4,122, and the Synology E25G30-F2 dual 25GbE card (Synology NAS only) is S$547.20. Prices change; ask for a current quotation. The DGX Station GB300 systems themselves, from ASUS, Dell, HP, MSI, GIGABYTE and Supermicro, are price on request and supplied through each OEM's channel. For a NAS shortlist, see our NAS buying guide, and for the 10GbE decision itself, is 10GbE worth it. 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 NAS connect to the DGX Station's 400G ports? Possibly, at lower speeds, but NVIDIA hasn't published which speeds, breakouts and optics it validates for the Station. Confirm part numbers before buying, or start on the 10GbE port.

Should the Station use NFS or SMB? NFS for data the GPU reads, because NVIDIA's GPUDirect Storage supports NFS and not SMB. SMB is fine for documents and user files.

Is 10GbE fast enough for a DGX Station? For many teams, yes, if hot data is staged locally. A 141 GB model takes about two minutes at 10GbE line rate; teams that move large checkpoints daily benefit from 25GbE or faster.

Does GPUDirect Storage work with my NAS? Only for supported file systems such as NFS over RDMA, and NVIDIA hasn't documented it for the Station. Secure Boot, on by default, may also block it.

What MTU should I use? NVIDIA's two-Station playbook uses 9000. Any MTU works if it is identical on the Station, switch and NAS; a mismatch causes stalled transfers.