NVIDIA Jetson AGX Thor
Blackwell-powered robotics supercomputer delivering up to 2070 FP4 TFLOPS for physical AI
14-core Neoverse-V3AE | 2560-core Blackwell GPU | 128GB LPDDR5X | 40-130W
View on NVIDIA.comOverview
The NVIDIA Jetson AGX Thor is the flagship Jetson module, delivering 2070 FP4 TFLOPS — 7.5x the AI performance of AGX Orin at 3.5x better efficiency. Built on the Blackwell GPU architecture with a 14-core Arm Neoverse-V3AE CPU, it includes 128GB LPDDR5X memory, PCIe Gen5, 16x CSI-2 camera lanes, and 4x25GbE networking. It is the definitive platform for humanoid robotics, autonomous systems, and next-generation physical AI.
Specifications
| Specification | Value | Notes |
|---|---|---|
| CPU | 14-core Arm Neoverse-V3AE | |
| GPU | 2560-core Blackwell, 96 Tensor Cores | |
| AI Performance | 2070 FP4 TFLOPS / 1035 FP8 TFLOPS | |
| Memory | 128GB LPDDR5X (273 GB/s) | |
| Storage | PCIe Gen5 NVMe | 1TB in dev kit |
| Video Decode | Up to 10x 4Kp60 or 4x 8Kp30 | H.265, H.264, AV1, VP9 |
| Video Encode | Up to 6x 4Kp60 | H.265, H.264 |
| Camera | 16x CSI-2 lanes, 20+ cameras | |
| Networking | QSFP 4x25GbE + RJ45 Multi-GbE | |
| Display | Dual HDMI 2.1 + DP 1.4a | |
| USB | 3x USB 3.2, USB-C | |
| PCIe | Up to 8 lanes Gen5 | |
| Other I/O | CAN, I2C, UART, PWM, SPI, GPIO | |
| Module Size | ~87 x 100 mm (699-pin B2B) | |
| Power | 40-130W configurable |
Use Cases
- Humanoid robotics and general-purpose robotic platforms requiring real-time reasoning
- Autonomous vehicles and mobile machines with multi-sensor fusion
- Edge AI servers for large-model inference (LLMs, VLMs) in industrial and medical settings