data center storage
GP Spark AI Storage Companion
Data center-grade NVMe-oF storage node designed specifically for NVIDIA DGX Spark, hardware offload engine enables end-to-end data acceleration, plug-and-play, desktop-grade form factor.
corecharacteristic
Hardware offload engine
RDMA (RoCEv2) and NVMe-oF protocol stack are solidified in dedicated chips, and single-port 100GbE provides 12GB/s throughput.
plug and play
DAC/AOC cables connect directly to DGX Spark, automatically discovering devices and allowing GDS access without formatting.
Desktop-grade design
150×150×36mm, 0.81L volume, aluminum-magnesium alloy shell, silent fan, power consumption <100W.
Flexible expansion
A single machine has 4×M.2 NVMe SSDs (up to 4TB each); a single host can mount 2 units, with 12GB/s throughput.
Product details
Product overview
LUISUAN Technology GP Spark is a data center-level NVMe-oF storage node specially designed for NVIDIA DGX Spark. It solidifies the full link of RDMA (RoCEv2) and NVMe-oF protocol stack into a dedicated chip through a hardware offload engine to achieve end-to-end data acceleration; single-port 100GbE can provide 12GB/s Throughput vs. 20μs End-to-end latency.
The whole machine is 150mm×150mm×36mm（0.81L) supports 4 M.2 SSD bays within the volume, and consumes only 100W; Using a one-piece aluminum-magnesium alloy shell as an efficient temperature-uniform heat sink, the silent fan can provide a quiet office environment even at 100% load, providing a plug-and-play storage acceleration solution that can be placed directly on the desktop for data-intensive workstations such as AI training, real-time special effects, and financial quantification.
Product structure
GP Spark adopts an integrated compact structure design. The one-piece aluminum-magnesium alloy shell also undertakes the function of uniform temperature dissipation. The internal 4 M.2 bays, hardware offload engine and power module are compact and reasonable in layout, taking into account desktop-level volume and data center-level performance.
Product features
Hardware offload engine
The full link of RDMA (RoCEv2) and NVMe-oF protocol stack is solidified into a dedicated chip to achieve end-to-end data acceleration; single-port 100GbE provides 12GB/s throughput and 20μs end-to-end latency.
plug and play
DGX Spark and GP Spark can be connected using a DAC/AOC cable. The device is automatically discovered on the host side and can be directly accessed through GDS without formatting.
Desktop-grade design
150×150×36mm, 0.81L volume, one-piece aluminum-magnesium alloy casing, silent fan, power consumption <100W, can be placed directly on the desk.
Flexible expansion
A single machine supports 4 M.2 NVMe SSDs (up to 4TB each); a single host can concurrently mount 2 GP Sparks, with an aggregate bandwidth of 200Gbps.
Core advantages
| Contrast Dimensions | Peregrine GP Spark | USB hard drive | NAS storage | Thunderbolt 5 |
|---|---|---|---|---|
| Delay | 4μS | 5-10ms | 2-5ms | 100μS |
| Bidirectional bandwidth | 200Gbps | 5.0Gbps | 5.0Gbps | 120Gbps |
| Spark compatibility | Native support | Need to drive | Requires NFS mount | Not supported |
| Multi-machine sharing | 200Gbps cluster | Not supported | Support (slow) | Not supported |
| Host CPU resource usage | Nearly zero | Higher | Higher | Higher |
Application Scenarios
AI large model + RAG knowledge base construction
It solves the pain point of insufficient native DGX Spark storage (only 1TB/4TB), quickly expands storage space, and supports massive knowledge base data caching and training.
High-speed data backup and migration
Through RDMA high-speed network interconnection, extremely fast data transmission between GP Spark and DGX Spark is realized, and rapid synchronization of multi-device clusters is supported.
KV Cache externalization acceleration
Provides up to 100GbE bandwidth connection, realizes high-speed KV Cache offloading, releases DGX Spark computing power, and improves overall training and inference efficiency.
technologySpecification
| parameter | GP Spark |
|---|---|
| maximum support | 4 × NVMe SSD |
| SSD specifications | M.2 2280 / 22110 |
| maximum bandwidth | Bidirectional 200Gbps |
| Maximum throughput | 12GB/s |
| Access latency | <4μs (to SSD controller) |
| Machine power consumption | < 100W |
| Internet method | 100GbE high-speed network |
| protocol | RDMA (RoCE v2), NVMe-oF, with GPUDirect Storage (GDS) support |
| Compatible platforms | NVIDIA DGX Spark, DGX Station, graphics workstation, computing server |
| Dimensions | 150*150*36mm |
| Advanced options | Supports RAID protection (optional) |