Luisuantech GP Spark Review: Nearly 10GB/s of Plug-and-Play Storage for the DGX Spark

Luisuantech's GP Spark enclosure adds up to 16TB of NVMe storage to the NVIDIA DGX Spark via a 100GbE connection, addressing the DGX Spark's lack of full-size M.2 slots. The 0.58-liter, four-bay box supports M.2 2280/22110 drives, presents them as native NVMe devices over NVMe-oF RDMA, and achieves nearly 10GB/s throughput in testing. The prototype unit, labeled GP-Spark-1000, is rated under 100W and uses passive cooling.

When we reviewed the NVIDIA DGX Spark https://www.storagereview.com/review/nvidia-dgx-spark-review-the-ai-appliance-bringing-datacenter-capabilities-to-desktops , storage was the platform’s clearest design flaw, and it is a form-factor problem before it is anything else. The Spark’s internal slots take short M.2 drives, the 2230 and 2242 class, where packaging wins, and capacity loses. The high-capacity end of the client SSD market lives in full-size 2280 drives, where 8TB models ship today, and the Spark simply has nowhere to put one. That leaves a machine built for serious AI work with a storage ceiling better suited to a thin-and-light laptop, and no internal path around it. The Luisuantech GP Spark is a solution for exactly that problem: a 0.58-liter, four-bay box for full-size M.2 drives that cables to the Spark’s 100GbE port, shows up as native NVMe devices with no drivers or formatting, and serves GPU Direct Storage traffic at close to line rate. The pitch is simplicity with client-drive economics. The GP Spark’s four bays take ordinary M.2 2280 or 22110 NVMe SSDs, the form factors the Spark itself locks out, and presents them over NVMe-oF RDMA through a hardware offload engine on a dedicated chip. Luisuantech’s spec sheet validates drives up to 4TB today, 16TB per enclosure, though these are the same slots where 8TB client drives already ship, so the practical ceiling is a validation question rather than a mechanical one. There is no enterprise array here, no licensing, and no storage OS to learn. Plug a DAC or AOC cable between the GP Spark and the DGX Spark, run modprobe nvme-rdma on the Spark’s Ubuntu base, and the drives appear as /dev/nvme devices ready for GDS access. Design and Build The GP Spark is a 150mm x 150mm x 26mm box, a smaller footprint than the Spark itself, wrapped in a perforated metal chassis with a single power button that doubles as a status light: green for normal, red for fault. Power comes over USB-C PD from a 20V/5.4A external adapter, with the whole unit rated under 100W, including drives. The rear panel carries exactly three connectors: the USB-C power input, a USB-C factory debug port, and the QSFP28 100GbE data port, which accepts copper DACs or optical modules. Inside, a dedicated data processor and coprocessor handle the NVMe-oF offload, and the four M.2 bays sit under the top cover. The rear panel is all business: USB-C power and debug ports on the left, the QSFP28 cage on the right, and a row of copper fin stacks visible through the vents between them. Cooling is entirely passive. Pop the top cover, and the four M.2 bays sit in a row, here populated with our KIOXIA XG8 test drives. The lid itself is the drive cooler: blue thermal pads on its underside couple each SSD to the finned heatsink that forms the top of the chassis, a clean passive solution for client drives that never see sustained enterprise duty cycles. A disclosure before the numbers: our unit is a prototype. The bottom label reads GP-Spark-1000, marks the device Prototype, Not for Resale, and carries a February 2026 build date under the Swingsoon brand Luisuantech uses on hardware. Production units may differ in fit and finish, though the platform behavior we tested is what Luisuantech is shipping to reviewers today. Two further notes on the out-of-box experience. Our unit shipped with two printed manuals entirely in Chinese, and initial setup appears to route through Wi-Fi onboarding. Neither is a blocker for the audience this box targets, but a Western launch will need English documentation. Setup and Architecture There is no RAID controller and no storage abstraction onboard: the GP Spark is a JBOD in the literal sense, exposing each installed SSD as its own NVMe-oF namespace. In our configuration, four drives appeared as four /dev/nvme devices on the host. Redundancy or striping is the host’s job. The vendor spec sheet lists a single 100GbE port at 10GB/s and 2.7M IOPS; the product report separately references 2x100GbE configurations and up to 24GB/s, a figure Luisuantech confirmed is aggregate read plus write. Our unit and testing used the single-port configuration. Luisuantech GP Spark Specifications Specification | Luisuantech GP Spark | |---|---| Platform Overview | | Drive Bays | 4 x M.2 NVMe 2280 / 22110 Mixed capacities supported, up to 4TB per drive | Network | QSFP28 100GbE DAC or optical RDMA required | Protocols | NVMe-oF RDMA GPU Direct Storage GDS | Performance Vendor-Stated | | Throughput | 10GB/s per 100GbE port Up to 24GB/s aggregate read plus write | IOPS | 2.7M | Access Latency | Under 20 microseconds | Power and Physical | | Power | Under 100W total 20V/5.4A USB-C PD external adapter | Dimensions | 150mm x 150mm x 26mm 0.58L | Operating Temperature | 0 to 40C | Compatibility | NVIDIA DGX Spark DGX Station Workstations and servers with RDMA-capable NICs | Performance Our test configuration paired the GP Spark with a GIGABYTE DGX Spark https://www.storagereview.com/review/gigabyte-ai-top-atom-review over a direct 100GbE connection, with four 1TB KIOXIA XG8 client NVMe SSDs populating the bays. It’s important to keep in mind that the drives you pick will play a significant role in the measured performance. We leveraged client Gen5 SSDs; some models, especially enterprise SSDs, may offer higher sustained write performance. We ran FIO sweeps across 4K and 64K random and 1M sequential workloads, read and write, stepping iodepth and numjobs to map the full envelope. Results reflect the final retest after applying Luisuantech’s MTU guidance, which improved transfer behavior over our initial runs. 4K Random Performance Small-block reads are where the offload engine shows its worth. 4K random reads scaled with queue depth to a peak of 2.43 million IOPS at 9,475 MiB/s, within sight of the vendor’s 2.7M claim and effectively saturating the 100GbE link with 4K transfers. For a passively powered four-bay box feeding a desk-side AI system, that is a remarkable figure. Writes follow the same shape at roughly half the height, peaking at 1.19 million IOPS. The gap between read and write ceilings is consistent across every workload we ran. The performance is directly related to the underlying drives, so results here will vary depending on configuration. Read latency bottoms out at 65.3 microseconds on average at low queue depth. That is higher than the vendor’s sub-20-microsecond claim, but results will vary depending on drive selection and network configuration. The network round trip is also doing work in that number; latency stays flat and predictable until the link saturates. Write latency is the one place the spec sheet claim lands: 20.4 microseconds average at minimal depth, right at the vendor’s under-20-microsecond figure and low enough that the fabric is effectively invisible to the application. 64K Random Performance At 64K, the story becomes purely about bandwidth. Random reads hold 9,503 MiB/s at peak, statistically identical to the 4K and 1M ceilings. Whatever block size the workload brings, the GP Spark delivers the same answer: the full line rate of its 100GbE port. 64K random writes plateau at 4,742 MiB/s, the same ceiling we measured at every other block size. 1M Sequential Performance Large-block sequential reads, the profile of model loading and dataset streaming, reach 9,496 MiB/s and hold there from modest queue depths onward. This is the workload the GP Spark exists for, and it runs at the wire. Sequential writes hold 4,742 MiB/s, roughly half of read throughput, and that ceiling is identical at every block size we tested. We flagged the asymmetry to Luisuantech during testing and worked through a round of tuning with the company, including MTU changes; the figures here represent the best the platform delivers in its current single-port configuration, and Luisuantech confirmed they are consistent with its specifications for the write path. For the read-dominated workloads this box targets, model loading, dataset streaming, and RAG retrieval, it is a footnote; for heavy ingest, size expectations accordingly. Conclusion The GP Spark does one thing and does it cleanly: it gives one or more DGX Spark the storage the platform really needs for heavy lifting. Cable it up, load the kernel module, and nearly 10GB/s of GDS-accessible flash appears without a driver install, a storage OS, or an enterprise invoice. Filling it with client M.2 drives is the point; capacity gets relatively cheap when the box accepts whatever 2280 or 22110 SSDs you have, and the offload engine handles the protocol work the drives never see. Our take is that this is a neat, well-executed add-on rather than a breakthrough. Reads stop at the single link’s line rate; writes stop at roughly half of that. Pricing is the open question: the GP Spark is not yet listed at retail in the US or China, our test unit is a prototype, and Luisuantech has not published pricing. The value argument rests on the box coming in meaningfully below enterprise NVMe-oF alternatives, which its client-drive design should allow. For Spark owners who hit the internal storage wall, and our original review suggests that many of them will, this is an easy path to solve that issue without carving out storage from a large enterprise storage estate.