Published · Phase 3
DGX Spark vs Jetson Thor
DGX Spark is the desktop AI-development appliance; Jetson Thor is the edge robotics platform. Choose Spark for model development and broad local serving, Thor for robots, sensors, real-time physical AI, and a 40–130 W deployment envelope.
Editorial review: complete · Updated 2026-08-30
Quick verdict
Quick verdict guidance: DGX Spark is the desktop AI-development appliance; Jetson Thor is the edge robotics platform. Choose Spark for model development and broad local serving, Thor for robots, sensors, real-time physical AI, and a 40–130 W deployment envelope.
Comparison table
| Configuration | Evidence state | Metric | Source |
|---|---|---|---|
| Qwen3.6-35B-A3B-NVFP4 · NVIDIA DGX Spark · vLLM | Source-reported · Grade C | 171.64 tokens/second | NVIDIA Developer Forums Qwen3.6 NVFP4 cross-platform report |
| Qwen3.6-35B-A3B-NVFP4 · NVIDIA Jetson AGX Thor Developer Kit · vLLM | Source-reported · Grade C | 239.06 tokens/second | NVIDIA Developer Forums Qwen3.6 NVFP4 cross-platform report |
Memory / capacity
Memory / capacity guidance: Both advertise 128 GB shared memory and 273 GB/s bandwidth, but the surrounding compute, I/O, thermal envelope, and deployment goals differ. Model fit alone should not decide between them.
Observed LLM performance
Observed LLM performance guidance: A source-reported Thor server result can show strong aggregate output, but it does not make Thor a universal faster desktop. Robotics pipelines share compute with perception and control.
Prefill vs decode
Prefill vs decode guidance: Long-context language inference is easier to dedicate on Spark; Thor often needs to share memory and compute with vision and sensor models.
Power
Power guidance: Thor's configurable 40–130 W range is designed for edge deployment. Spark's higher system envelope supports a broader workstation workload without the same robotic constraints.
Current market cost
Current market cost guidance: Include carrier boards, storage, networking, cameras, and integration for Thor. Spark is closer to a complete desktop product.
Which models fit
Which models fit guidance: Both advertise 128 GB shared memory and 273 GB/s bandwidth, but the surrounding compute, I/O, thermal envelope, and deployment goals differ. Model fit alone should not decide between them.
Who each option suits
Who each option suits guidance: Spark suits developers at a desk; Thor suits autonomous machines, industrial edge, and power-constrained multimodal inference near sensors.
What stands out
Choose Thor when physical I/O, edge power, rugged deployment, or Jetson software is a requirement. Otherwise choose Spark for the more general local-AI workstation experience.
A source-reported Thor server result can show strong aggregate output, but it does not make Thor a universal faster desktop. Robotics pipelines share compute with perception and control.
RTX 5090 is the speed-first workstation; small Jetson modules cover lighter edge workloads; an x86 industrial PC plus GPU may suit software that cannot target Jetson.
Evidence limitations
- A language-model benchmark does not represent a complete perception-planning-control workload.
- Confirm module, developer kit, and production carrier specifications separately.
Planner CTA
Compare this recommendation against your model, context, concurrency, latency, and budget in the ComputeSage Planner.
Sources / methodology
Recommendations combine official specifications, public model/runtime documentation, adjacent benchmark observations, and clearly labeled engineering estimates. Review the ComputeSage evidence methodology.