CtrlK
Template:Infobox computer
NVIDIA Jetson Thor is a high-performance edge AI computing module series for robotics and embedded applications, developed by NVIDIA. Introduced in 2025 as the successor to the Jetson AGX Orin platform, it offers significantly higher AI performance for "Physical AI" systems—general-purpose autonomous robots. The Jetson Thor system-on-module (SoM) features an NVIDIA Blackwell GPU (2560 CUDA cores with 96 Tensor Cores) and a 14-core Arm Neoverse-V3AE CPU, along with 128 GB of LPDDR5X memory in its flagship configuration.[1][2] This hardware delivers up to 2,070 FP4 TFLOPS of AI compute within a 130 W power envelope—about 7.5× the AI performance and 3.5× the energy efficiency of the previous Jetson Orin generation.[1][3] NVIDIA markets Jetson Thor as "the ultimate platform for physical AI and robotics," designed to run multiple advanced AI models (such as vision-language and large language models) in real time at the edge.[2][4]
The Jetson Thor series is part of NVIDIA's Jetson family of embedded AI computing platforms, which are widely used in autonomous machines and robots. Jetson Thor was unveiled amid NVIDIA's push toward more general-purpose humanoid robotics and "physical AI"—moving beyond single-purpose robots to adaptable robots capable of high-level reasoning and diverse tasks.[2] At GTC 2024, NVIDIA CEO Jensen Huang introduced the Isaac GR00T robotics platform and previewed Jetson Thor as the hardware backbone for next-generation generalist robots.[5] Jetson Thor was officially launched as a developer kit (Jetson AGX Thor Developer Kit) in late August 2025, with general availability announced at a price of US$3,499 for the kit.[6] The platform is not intended to replace Jetson Orin outright, but rather to sit above it as the highest-end Jetson offering, targeting applications requiring substantially more compute (e.g., generative AI and advanced humanoid robots).[7]
Jetson Thor is built around NVIDIA's Blackwell GPU architecture, which introduces support for Multi-Instance GPU (MIG) virtualization and a new Transformer Engine for 4-bit precision computing. The Jetson Thor GPU supports native FP4 data types, dynamically switching between 4-bit and 8-bit modes to optimize performance for transformer models (e.g., large neural networks).[2] Key GPU specifications include:
2,560 CUDA cores at 1.57 GHz
96 fifth-generation Tensor Cores
Multi-Instance GPU (MIG) capability with 10 TPCs (Texture Processing Clusters)
Transformer Engine with dynamic FP4/FP8 precision switching
The SoM includes a 14-core Arm Neoverse-V3AE 64-bit CPU (with 1 MB L2 cache per core and 16 MB shared L3), providing strong general-purpose and real-time processing capabilities alongside the GPU.[8]
The module also integrates specialized accelerators:
Third-generation Programmable Vision Accelerator (PVA 3.0) for computer vision tasks
Dual hardware video encoders (NVENC) and decoders (NVDEC)
Optical flow accelerator
Always-on DSP[2][7]
With its Blackwell GPU and advanced accelerators, Jetson Thor can run multiple high-end AI models simultaneously. Video processing capabilities include:
Decode: Up to four 8K@30fps or ten 4K@60fps video streams in parallel
Encode: Up to six 4K@60fps streams
Codecs: H.265 (HEVC), H.264 (AVC), AV1 (decode), VP9/VP8
Display: Up to four independent displays via HDMI 2.1 and DisplayPort 1.4a at resolutions up to 8K (7680×4320 @30 Hz)[7]
Jetson Thor offers significantly expanded I/O and networking compared to its predecessors:
Networking:
4× 25 GbE high-speed Ethernet interfaces (aggregated via QSFP28 connector)
1× 5 GbE RJ45 Ethernet port
Wi-Fi 6E and Bluetooth (developer kit)
Storage and Expansion:
PCIe Gen5 lanes (configurable up to x8 + x4 + x2)
NVMe M.2 support
Multiple USB 3.2 Gen2 ports
Industrial Interfaces:
CAN bus
Multiple UARTs
SPI
I²C
GPIO[7]
Holoscan Sensor Bridge technology is supported for time-synchronized sensor streaming over Ethernet (enabling camera data over 10GbE links with low latency), which is a new approach for high-bandwidth sensor input on Jetson platforms.[7] Jetson Thor modules have a 699-pin board-to-board connector (87 × 100 mm module size) but are not pin-compatible with Jetson Orin modules due to the new interface changes and higher power requirements (Thor modules draw up to ~120–130 W, whereas AGX Orin modules were limited to ~60 W).[7]
The Jetson Thor series consists of two module variants (SoMs) and an associated developer kit:
| Model (SoM) | GPU (architecture) | AI Performance**(FP4 sparse) | CPU (cores) | Memory (LPDDR5X) | Power | | --- | --- | --- | --- | --- | --- | | Jetson T5000** | 2560-core Blackwell GPU**(96 Tensor Cores, MIG with 10 TPCs) | 2070 TFLOPS | 14-core Arm Neoverse-V3AE | 128 GB @ 273 GB/s | 40–130 W | | Jetson T4000*** | 1536-core Blackwell GPU**(64 Tensor Cores, MIG with 6 TPCs) | 1200 TFLOPS | 12-core Arm Neoverse-V3AE | 64 GB @ 273 GB/s | 40–70 W | | *Specifications for Jetson T4000 are preliminary (under development). |
The Jetson T5000** is the flagship module, incorporating the full 2560-core Blackwell GPU and 128 GB memory to achieve the maximum performance. The lower-tier Jetson T4000 is a cost-reduced variant with a smaller GPU (1536 cores, 64 Tensor Cores), 64 GB of memory, and roughly 60% of the AI compute throughput of the T5000.[8] Both modules use the same form-factor and include the new high-speed interfaces, but the T4000 targets a lower power range (up to ~70 W) for applications that don't require the absolute highest performance.[8]
NVIDIA provides the Jetson AGX Thor Developer Kit, which includes a Jetson T5000 module pre-mounted on a reference carrier board with additional peripherals:
Kit Contents:
Jetson T5000 module
Reference carrier board
1 TB NVMe SSD (M.2 slot)
Wi-Fi 6E + Bluetooth module (M.2 Key E)
Active cooling solution (heatsink + fan)
140W power adapter
I/O Ports:
Video: 1× HDMI 2.0b, 1× DisplayPort 1.4a
USB: Multiple USB 3.2 and USB 2.0 ports
Networking: 5 GbE and 4×25 GbE ports (via QSFP28)
Industrial: CAN bus, UART, I²S audio headers
Debug: JTAG connectors
The entire dev kit assembly, including the module and cooling solution, measures approximately 243 × 112 × 57 mm in size.[8] The Jetson AGX Thor developer kit is designed to operate as a "robot brain" out of the box, enabling researchers and engineers to evaluate performance on real workloads without designing a custom board.[8] The dev kit began shipping to customers in Q3 2025 following its announcement.[8][4]
Jetson Thor runs on the same software stack as other NVIDIA Jetson platforms, including:
JetPack 7.0 based on:
Ubuntu 24.04 LTS
Linux kernel 6.8
CUDA 13.0
TensorRT 10.13
cuDNN 9.12
VPI (Vision Programming Interface) 4.0[2]
The platform supports deployment of generative AI frameworks including:
PyTorch
ONNX
Models from Hugging Face, Meta, OpenAI, Gemini, Qwen, and DeepSeek[2]
NVIDIA Isaac robotics software includes:
Isaac Sim: Physics-accurate simulation built on NVIDIA Omniverse
Isaac Lab: Framework for reinforcement and imitation learning
Isaac GR00T: Foundation models for humanoid robots
Isaac ROS: Hardware-accelerated ROS 2 packages
NVIDIA Metropolis provides vision AI and smart city applications, while NVIDIA Holoscan enables real-time sensor and medical imaging pipelines.[2][4]
Jetson Thor is built to handle emerging generative AI workloads at the edge. Performance metrics include:
| Metric | Performance |
|---|---|
| Time to First Token (TTFT) | < 200ms |
| Time Per Output Token (TPOT) | < 50ms |
| Speedup vs Orin (generative) | 5× |
| Speculative Decoding | 2× additional speedup |
These capabilities enable real-time inference for models like large language models (LLMs) and vision-language-action models.[2] The platform's massive 128 GB memory allows deployment of large AI models and handling large sensor data in real time.[4]
Jetson Thor is targeted at advanced robotics and autonomous machines that require server-class AI capability at the edge:
Humanoid robots: General-purpose robots for unstructured environments
Autonomous mobile robots (AMRs): Warehouse and logistics automation
Drones: Advanced aerial robotics
Industrial automation: Cobots and smart manufacturing
Autonomous vehicles: DRIVE AGX Thor variant for automotive
Edge AI servers: High-performance local inference
The high compute density and wide I/O of Thor enable:
Multi-modal sensor fusion (camera, lidar, radar, IMU)
Autonomous navigation and manipulation
On-device generative models for human-robot interaction[4]
Several major companies announced adoption of Jetson Thor:
Robotics Companies:
Agility Robotics – Digit humanoid robot
Amazon Robotics – Warehouse automation
Boston Dynamics – Legged robots
Figure – Humanoid robotics
1X Technologies – EVE and NEO robots
Industrial Partners:
Caterpillar – Autonomous industrial vehicles
John Deere – Agricultural automation (evaluating)
Medtronic – Medical robotics
Technology Companies:
Meta – AI research for physical AI
OpenAI – Robotics research[4]
According to NVIDIA:
Over 7,000 customers deployed Jetson Orin-based hardware by 2025
150+ hardware partners provide production-ready solutions
2+ million developers active on NVIDIA robotics platforms[4]
| Category | Specification |
|---|---|
| AI Performance | 2,070 TFLOPS (FP4 sparse) |
| 1,035 TFLOPS (FP8 dense) | |
| 517 TFLOPS (FP16 sparse) | |
| GPU | 2,560-core Blackwell GPU |
| 96 Tensor Cores (5th gen) | |
| 10 TPCs for MIG | |
| CPU | 14-core Arm Neoverse-V3AE |
| Up to 2.6 GHz | |
| 1 MB L2 per core, 16 MB L3 shared | |
| Memory | 128 GB LPDDR5X |
| 256-bit bus | |
| 273 GB/s bandwidth | |
| Video Encode | Up to 6× 4K@60fps |
| H.265/H.264 | |
| Video Decode | Up to 4× 8K@30fps |
| Up to 10× 4K@60fps | |
| H.265/H.264/AV1 | |
| Networking | 4× 25 GbE via QSFP28 |
| 1× 5 GbE RJ45 | |
| Wi-Fi 6E (dev kit) | |
| Storage | PCIe Gen5 support |
| NVMe M.2 | |
| USB 3.2 | |
| Power | 40W – 130W configurable |
| Form Factor | 100mm × 87mm |
| 699-pin connector | |
| Price | $3,499 (developer kit) |
| $2,999 (module, 1K units) |
NVIDIA Jetson – Family of Jetson modules and kits
Jetson AGX Orin – Predecessor high-end Jetson platform (2022)
NVIDIA Isaac – NVIDIA's robotics software platform
NVIDIA Blackwell – GPU architecture
Humanoid robot
Physical AI
Edge computing and Edge AI
Multi-Instance GPU