NVIDIA Jetson Thor
Powering the Future of Physical AI
NVIDIA’s Jetson Thor is the latest addition to the NVIDIA Jetson family. NVIDIA Jetson Thor is not a replacement for NVIDIA Jetson Orin. Instead, it is the highest-end member of the Jetson Family that comes to the market to fulfill the hardware needs of applications with high demand for resources, like Generative AI and Humanoids, to mention a few. Its cutting-edge system-on-module (SoM) combines the latest Blackwell GPU architecture and Arm Neoverse V3AE CPUs with 128GB of LPDDR5X memory, designed for high-performance edge AI applications.
This article provides a brief introduction to the new SoC, and we hope it will give you the basis to understand what is new on Thor and how RidgeRun can help you integrate it into your system.
Key Specifications
In this section, you will find the key specifications of Thor. It is important to highlight that Thor is not pin-compatible with NVIDIA Jetson Orin. In the past, the Orin modules were made to be compatible, pin to pin. But this is not the case if you want to upgrade your system with Thor. Also, Thor is a 120W device while Orin is 60W. The Thor datasheet can be found in the Jetson Download Center.
Another important aspect to mention is that Thor implements a key change for the capture side. NVIDIA is targeting two approaches:
- MIPI CSI: If you need to capture from MIPI, the SoC includes the fifth generation of the NVIDIA Camera Solution (NVCSI 2.0, VI 6.0, and ISP 2.x). It includes Virtual Channel ID support and 16 MIPI CSI Lanes. However, note that the devkit sold by NVIDIA is not going to include MIPI CSI Ports exposed. You will need to buy a devkit from a hardware partner. Something interesting is that in Orin, the glass-to-glass latency (camera-display) was around 130ms when using MIPI CSI.
- Holoscan Sensor Bridge: This is a new approach adopted by NVIDIA for customers using Jetson Orin and Jetson Thor. This will be officially supported by the NVIDIA Thor Devkit. You will receive frames (raw – not encoded) via Ethernet. You can see more details of the hardware setup here. This Ethernet link will be used for video frames, metadata, and configuration commands that are later forwarded by the lattice FPGA to the cameras via I2C. The glass to glass latency here is 40ms. RidgeRun can help you create your camera driver for the Holoscan Sensor Bridge.
Other specs include:
- GPU: Blackwell GPU, 2560 CUDA cores, 96 5th-gen Tensor Cores, supports MIG, up to 2070 TFLOPs FP4 (sparse), 1035 TFLOPs FP8 (dense)
- CPU: 14× Arm Neoverse V3AE cores, 64KB+64KB L1, 1MB L2 per core, 16MB shared L3, up to 2.6 GHz
- Memory: 128GB LPDDR5X @ 4266 MHz, 256-bit bus, 273 GB/s bandwidth
- Video Decode/Encode: Supports H.265/H.264/AV1, up to 4× 8Kp30 decode and 6× 4Kp60 encode
- Display: 4× HDMI 2.1 / DP 1.4a, up to 8K at 30 Hz
- AI/Vision Accelerators: PVA 3.0, Always-On DSP (dual Cadence HiFi 5)
- I/O: 4× 25GbE, 16× CSI-2 lanes, PCIe Gen5, USB 3.2 (3×), CAN, I2C, UART, PWM, SPI, GPIO
- Power: Configurable 75W – 120W, Max 130W. Note this important detail. Orin is a 60W device while Thor is 120W.
- Form Factor: Approx. 87×100 mm, 699-pin B2B connector
Advanced Multimedia Capabilities
- Supports decoding of up to 4× 8Kp30 or 10× 4Kp60 video streams in parallel.
- Includes dual NVDEC and NVENC engines for high-performance video decoding and encoding. Note, this is different to NVIDIA Jetson Orin. Where only 1 hardware encoder/decoder was available.
- Can encode up to 6× 4Kp60 video streams simultaneously in H.265 or H.264.
- Compatible with modern codecs like H.265, H.264, AV1, VP9, VP8, MPEG-2/-4.
- Drives up to 4 independent displays via HDMI 2.1 and DisplayPort 1.4a.
- Supports 8K output (7680×4320 @ 30Hz) for ultra-high-resolution displays.
- Ideal for multi-camera systems and real-time video analytics.
The Most Powerful NVIDIA Jetson
Jetson Thor is the most powerful module in the NVIDIA Jetson family, built to handle the toughest AI and robotics challenges at the edge. It packs more performance, memory, and connectivity than any Jetson before it, making it the go-to choice for developers building the next generation of intelligent machines.
Features Comparison Table
The following table compares NVIDIA Jetson Thor to the Jetson Orin-based modules (Jetson AGX Orin, Jetson Orin NX, and Jetson Orin Nano) across key hardware features and performance metrics:
| Feature | Jetson Thor | Jetson AGX Orin | Jetson Orin NX | Jetson Orin Nano |
|---|---|---|---|---|
| GPU Architecture | Blackwell, 2560 CUDA, 96 Tensor Cores, MIG support | Ampere, up to 2048 CUDA, 64 Tensor Cores | Ampere, 1024 CUDA, 32 Tensor Cores | Ampere, 512 (4GB) / 1024 (8GB) CUDA, 16 Tensor Cores |
| AI Performance | Up to 2070 TFLOPs FP4 (sparse) / 1035 TFLOPs FP8 (dense), 8.064 FP32 TFLOPs | Up to 275 TOPS | Up to 100 TOPS | Up to 40 TOPS |
| CPU | 14× Neoverse V3AE, 64KB+64KB L1, 1MB L2/core, 16MB L3, up to 2.6 GHz | 12× Cortex-A78AE, 3MB L2, 6MB L3 | 8× Cortex-A78AE, 2MB L2, 4MB L3 | 6× Cortex-A78AE, 1.5MB L2, 4MB L3 |
| RAM | 128 GB LPDDR5X, 256-bit, 4266 MHz, 273 GB/s | 32–64 GB LPDDR5, 204.8 GB/s | 8/16 GB LPDDR5, ~102 GB/s | 4/8 GB LPDDR5, ~68 GB/s |
| Storage | >64 MB NOR, NVMe (PCIe Gen5, x4), SSD via USB 3.2 | 64 GB eMMC, NVMe over PCIe Gen4 | External eMMC supported | External eMMC supported |
| Networking | 4× 25GbE MACs (100 Gbps total) | 1× 10GbE + 1× 1GbE | 1× 1GbE | 1× 1GbE |
| PCIe | Gen5 (x8+x4+x2), Root & Endpoint | Gen4 (up to 22 lanes) | Gen4 (up to 16 lanes) | Gen4 (up to 8 lanes) |
| Camera Interfaces | 16× CSI-2 lanes, up to 6 cameras, 32 virtual channels | 16× CSI-2 lanes | 8× CSI-2 lanes | 8× CSI-2 lanes |
| Display Outputs | 4× HDMI 2.1 / DP 1.4a, up to 8K @30Hz | 3–4× outputs, up to 8K | 2× 4K | 1× 4K |
| Video Decode | Dual NVDEC @ 1.56 GHz, 10× 4Kp60 or 4× 8Kp30 (H.265/H.264), AV1, VP9 | 3x4K@60
11x1080p@60 |
2x4K@60
9x1080p@60 |
1x4K@60
5x1080p@60 |
| Video Encode | Dual NVENC @ 1.56 GHz, 6× 4Kp60, H.265/H.264 | 1x4K@60
7x1080p@60 |
1x4K@60
6x1080p@60 |
No HW Encoder |
| Other I/O | 3× USB 3.2, 4× USB 2.0, 13× I2C, 4× CAN, 4× UART, 5× I2S, SPI, PWM, GPIO | Similar | Reduced | Minimal |
| Power | 75W, 95W, 120W modes, Max 130W TMP | 15W–60W | 10W–40W | 7W–20W |
| Form Factor | 87×100 mm, 699-pin B2B, integrated heatplate | 100×87 mm AGX | 70×45 mm SO-DIMM | 70×45 mm SO-DIMM |
| Target Use Cases | Advanced robotics, autonomous machines, generative AI at edge | High-end robotics, vision servers | Smart cameras, drones | IoT, kiosks, entry-level AI |
Potential Applications of NVIDIA Jetson Thor
NVIDIA Jetson Thor is targeted at applications that demand server-class AI performance and extensive I/O at the edge. Its capabilities enable a range of advanced use cases.
|
🦾 Next-Generation Robotics & Autonomous Machines
🚗 Autonomous Vehicles & Intelligent Machines
🚀 Unified Edge AI Power
|
🧠 Edge AI Inference Servers
🏭 Industrial Automation & Vision Systems
🏥 Healthcare & Advanced Imaging
|
RidgeRun Services for Jetson Thor
RidgeRun, an embedded software engineering company and official NVIDIA Jetson ecosystem partner, provides a comprehensive set of services tailored for Jetson Thor developers. These services help unlock the full potential of the module and accelerate time-to-market.
Custom Driver Development (Camera and Sensors)
- Experts in Linux device driver development for Jetson platforms.
- Holoscan Sensor Bridge Drivers and Pipeline customization.
- Support for V4L2 camera drivers, GMSL/FPD-Link, and advanced image sensors.
- Experience with 50+ camera integrations (Sony, Omnivision, OnSemi).
- Services include driver bring-up, device tree config, and multi-camera sync.
- Extendable to IMUs, LiDARs, radars, and FPGAs over PCIe.
- Ideal for multi-sensor robotics and vision systems.
GStreamer and Multimedia Acceleration
- Deep GStreamer expertise for NVIDIA Jetson platforms.
- Design of efficient video pipelines using NVENC, NVDEC, VIC, and zero-copy.
- Real-time 4K video ingestion, AI inference, annotation, and streaming.
- Applications: public safety, AR/VR, drone video, IP camera systems.
- Performance tuning with NVIDIA Nsight and RidgeRun tools.
AI Integration and Optimization (DeepStream, TensorRT)
- Port and optimize AI models (ONNX, TensorFlow, PyTorch) to TensorRT.
- Develop DeepStream-based AI vision analytics pipelines.
- Support for multi-stream, multi-model concurrency (FP8, INT8).
- RidgeRun.ai services include quantization, pruning, and DNN pipeline deployment.
- Enable efficient inference on Jetson Thor’s GPU and PVA.
Camera Frameworks and Vision SDKs
- Integration with LibArgus, OpenCV, and custom CUDA vision modules.
- Support for HDR imaging, triggered capture, and multi-camera coordination.
- Development of advanced computer vision features: image stitching, stabilization, birds eye view.
- Algorithms optimized for Jetson Thor’s GPU for minimal latency.
Platform Security and BSP Integration
- Secure Boot enablement and firmware signing.
- Disk encryption and OP-TEE integration for secure execution.
- OTA update mechanisms compatible with Jetson security stack.
- BSP services: device tree, kernel, and power profile customization for custom carrier boards.
- Ensures full hardware bring-up and production readiness.
Full-Stack Support for Jetson Thor Deployments
From low-level driver development and multimedia acceleration to AI deployment and platform hardening, RidgeRun offers end-to-end support. Our experience with NVIDIA Jetson platforms makes us the ideal partner for projects requiring multi-camera setups, real-time AI, and robust deployment on Jetson Thor.
FAQ (Frequently Asked Questions)
Q1. What is NVIDIA Jetson Thor, and why is it significant?
- Jetson Thor is NVIDIA’s most powerful Jetson module to date, delivering over 1000 TOPS of AI performance, 128 GB of high-speed memory, and advanced I/O for edge computing. It brings server-class capabilities to embedded AI applications like robotics, automation, and real-time video analytics.
Q2. How does Jetson Thor compare to Jetson Orin modules like AGX, NX, or Nano?
- Thor significantly outperforms all Orin modules with 3–4× the AI performance of AGX Orin, double the memory capacity, and newer interfaces like PCIe Gen5 and 25GbE networking. It’s built for the most demanding edge applications.
Q3. What are the ideal applications for Jetson Thor?
- Jetson Thor is optimized for advanced robotics, autonomous vehicles, AI inference servers, industrial inspection systems, medical imaging devices, and generative AI use cases deployed at the edge.
Q4. What kind of multimedia performance can I expect from Jetson Thor?
- Thor supports decoding of up to 10× 4Kp60 or 4× 8Kp30 video streams and encoding up to 6× 4Kp60. It supports modern codecs (H.265, AV1, VP9, etc.) and drives up to 4 independent 8K displays, making it ideal for high-resolution, multi-camera systems.
Q5. How can RidgeRun help me integrate cameras and sensors with Jetson Thor?
- RidgeRun offers deep expertise in V4L2 driver development and has built drivers for over 50 camera models. They support multi-camera setups (CSI, GMSL, FPD-Link) and handle synchronization, and sensor fusion for advanced robotics and vision systems.
Q6. Does RidgeRun support GStreamer and multimedia pipeline development on Thor?
- Yes. RidgeRun builds high-performance GStreamer pipelines that leverage Thor’s NVENC/NVDEC engines and other hardware blocks. We also optimize pipelines for real-time streaming, inference, and video analytics. RidgeRun has been working with GStreamer for more than 15 years and for 8 years delivering NVIDIA Jetson GStreamer-based solutions.
Q7. Can RidgeRun help with AI model deployment and performance tuning on Thor?
- Absolutely. RidgeRun supports model conversion to TensorRT, integration with DeepStream, multi-model concurrency, and performance tuning using FP8/INT8 formats to fully exploit Jetson Thor’s AI throughput. RidgeRun's AI division ridgerun.ai will speed up your time to market with a full catalog of AI services.
Q8. What options are available for camera frameworks and custom vision development?
- RidgeRun supports LibArgus, OpenCV, and custom CUDA-based vision workflows. They develop specialized features like stereo imaging, HDR pipelines, and 360° video stitching—all optimized for GPU acceleration on Thor. RidgeRun also supports custom PVA development for NVIDIA Jetson Thor.
Q9. What security features does RidgeRun help implement on Jetson Thor?
- We provide secure boot configuration, OP-TEE integration, disk encryption, and over-the-air (OTA) update systems. RidgeRun ensures your platform is protected and production-ready, especially for safety-critical systems.
Q10. Can RidgeRun support my custom hardware (carrier board) and bring-up?
- Yes. RidgeRun offers BSP (Board Support Package) services, including device tree customization, kernel tuning, and power profile configuration for custom carrier boards, ensuring complete hardware/software integration for Jetson Thor.
For direct inquiries, please refer to the contact information available on our Contact page. Alternatively, you may complete and submit the form provided at the same link. We will respond to your request at our earliest opportunity.
Links to RidgeRun Resources and RidgeRun Artificial Intelligence Solutions can be found in the footer below.
ARTIFICIAL INTELLIGENCE
