DGX Spark Shared Storage with SupremeRAID™

Shared Storage for Deskside AI: Storage Review Tests SupremeRAID™ with Eight DGX Spark Systems

Deskside AI systems built on the NVIDIA GB10 platform put 128GB of unified memory and a petaflop of compute right next to the people using them. Storage is where they come up short. Each DGX Spark and its OEM variants ships with a single short M.2 slot, which caps local capacity around 4TB and leaves every model and dataset on that machine outside the governance, backup, and access controls that live in the data center. Across a lab or a department, that adds up to dozens of isolated copies of the same data.

In a new report, Storage Review addressed that problem by moving the data back to the data center. The team built a shared storage host from a Dell PowerEdge R770AP, eight Solidigm D5-P5336 61.44TB QLC SSDs protected by SupremeRAID™ in RAID 5, two Broadcom 400GbE NICs, and Tuxera Fusion NFS serving NFS over RDMA. Eight GB10 systems from Dell, GIGABYTE, HP, Acer, and ASUS connected over their existing 100GbE ports and mounted the same pool with a one-line command on stock Ubuntu.

How SupremeRAID™ Fits

With eight NVMe drives each rated up to 7,000MB/s sequential read, Storage Review noted that a conventional hardware RAID card becomes the bottleneck. SupremeRAID™ moves parity calculations to a GPU, in this case a compact NVIDIA RTX A2000, while keeping the data path on the host PCIe lanes. The result was roughly 430TB of usable, parity-protected capacity with single-drive fault tolerance.

What the Testing Showed

• With all eight systems reading at once, the array delivered 24.9GiB/s of aggregate sequential read while using about 27% of the host’s 800Gb/s network capacity.
• A single system reading from the shared array reached 10.8GiB/s, matching the fastest internal M.2 drive Storage Review has measured in a GB10 system.
• A 78GB Qwen3.5-122B model loaded from the shared array within 13% of its load time from local NVMe. When all eight systems loaded the same single copy at once, five finished faster than the single-system NFS run.

Why It Matters

Oregon State University shared its thinking with Storage Review as it explores a large-scale DGX Spark lab for AI learning, where students would work with research datasets measured in tens of terabytes. A central, governed array lets any system be added, reimaged, or reassigned without affecting anyone else’s data, and keeps one copy of each model and dataset under existing data center controls. The same approach applies to enterprises deploying desktop AI systems as developer workstations.

Read the full Storage Review report: https://www.storagereview.com/review/dgx-spark-shared-storage-solidigm-qlc-100gbe

Ready to explore GPU Accelerated Storage for your AI infrastructure? Contact the Graid Technology team.

‍