DGX Spark + Jetson Thor Custom Mini Rack — Compact AI Infrastructure
Published:
Custom-designed compact AI infrastructure combining 2x NVIDIA DGX Spark, 2x NVIDIA Jetson Thor, and a QNAP QSW-M7308R-4 optical switch — concentrating real AI compute power in the smallest possible footprint.
System Configuration
| Component | Specs | Role |
|---|---|---|
| 2x DGX Spark | NVIDIA Grace Blackwell | Workstation-class training & inference |
| 2x Jetson Thor | 800 TFLOPS FP8, 50W | Real-time edge robotics inference |
| QNAP QSW-M7308R-4 | High-speed optical | Low-latency interconnect |
Why This Matters
This setup represents an emerging trend in AI development:
- Distributed AI — Seamlessly split workloads between edge and compute tiers
- Edge + Local Compute — Prototype on Jetson Thor, scale on DGX Spark, deploy everywhere
- Compact but Powerful — Full Physical AI development stack in a single mini rack
- End-to-End Pipeline — Train in simulation on DGX Spark, deploy on Jetson Thor, iterate in real-time
This architecture enables a single developer or small team to run the complete Physical AI pipeline — from simulation and training to real-time robot inference — without cloud dependencies.