• Quick Start: docs.nvidia.com
• Tips

Introduction

Meet NVIDIA Jetson Thor—the first “physical-AI” edge computer built for humanoid robots, AMRs, industrial vision gateways and medical instruments. A 2070-TFLOPS Blackwell GPU, 128 GB LPDDR5X and 14-core Arm Neoverse-V3AE deliver up to 7.5× the AI performance of Jetson AGX Orin while drawing only 40-130 W. Run multi-modal transformer pipelines locally, cut cloud costs and guarantee <40 ms closed-loop control. JetPack 7, TensorRT 10 and pre-trained vision-language-action models get you coding in hours, not weeks.

Features

[Ultra-Performance AI] 2070-TFLOPS FP4 Blackwell GPU + 96 Tensor Cores let Jetson Thor run LLM, VLM & VLA models on-device; latency <40 ms, no cloud fees.
[Expandable Memory] 128 GB LPDDR5X 273 GB/s and 1 TB NVMe give engineers room for 70-B-parameter models, 4-K video buffers and ROS 2 stacks.
[Scalable Power] 40-130 W dynamic budget—run fanless on AMRs or unlock 130 W for dual-arm cobot control; carrier board included.
[Rich I/O] PCIe Gen4 x8, 4×25 GbE, USB-C, MIPI CSI/DSI, QSFP, CAN—hook lidar, stereo cams, 5G or GMSL in minutes.
[Production Ready] Same T5000 module in dev kit; seamless path from lab to field, supported by NVIDIA JetPack 7, Metropolis & Holoscan SDKs.

Specification

GPU: 2560 CUDA, 96 Tensor, Blackwell | CPU: 14-core Arm Neoverse-V3AE @ 2.2 GHz | RAM: 128 GB LPDDR5X 273 GB/s | Storage: 1 TB NVMe (expandable) | Power: 40-130 W, 12-54 V DC | Dimensions: 243×112×57 mm | I/O: PCIe Gen4 x8, 4×25 GbE, USB-C, MIPI CSI/DSI, QSFP, CAN, UART, I2C, SPI | OS: JetPack 7 (Linux 6.8, CUDA 13, TensorRT 10)