NVIDIA DGX Spark

A deskside training box with 128 GB of coherent unified memory. In a robot-learning programme this is the machine that sits beside the capture rig: you post-train or fine-tune a policy on the demonstrations you just collected, then push the result to the Jetson on the robot.

Specifications

• Up to 1 petaFLOP of FP4 AI performance
• NVIDIA GB10 Grace Blackwell superchip
• 20-core Arm CPU (10× Cortex-X925, 10× Cortex-A725)
• 128 GB coherent unified LPDDR5x at 273 GB/s
• 4 TB self-encrypting NVMe storage
• ConnectX-7 200 Gbps networking for two-unit clustering
• 240 W power supply, 140 W chip TDP
• Includes a 90-day NVIDIA AI Enterprise licence

Where it fits

Layer 7 of the stack — see data & policy for how collected episodes become a trained policy.

See also teleoperation stations and the robotics hub.