NVIDIA Jetson AGX Thor Developer Kit
NVIDIA Jetson AGX Thor Developer Kit — 2070 FP4 TFLOPS — 128 GB LPDDR5X — Blackwell GPU The NVIDIA® Jetson AGX Thor™ Developer Kit is the most powerful Jetson platform...
Read MoreSKU:NVD-016
- Regular price
- Rs. 394,299.99
- Sale price
- Rs. 394,299.99
- Regular price
-
Express Shipping
Shipping Info
Fast delivery options for your convenience
Pick-Up in Store
Free
Select this option to pick up your order directly from our store. Fast and convenient!
NVIDIA Jetson AGX Thor Developer Kit — 2070 FP4 TFLOPS — 128 GB LPDDR5X — Blackwell GPU
The NVIDIA® Jetson AGX Thor™ Developer Kit is the most powerful Jetson platform ever built, delivering 2070 FP4 TFLOPS of AI compute and 128 GB of unified memory within a 130 W power envelope. Powered by the NVIDIA Blackwell GPU architecture, it achieves up to 7.5× the AI performance and 3.5× the energy efficiency of Jetson AGX Orin — the definitive platform for humanoid robotics, physical AI, and real-time generative AI at the edge.
Key Highlights
- 2070 FP4 TFLOPS AI Performance — Delivers 7.5× the AI compute of Jetson AGX Orin, enabling on-device execution of large generative AI models — including Vision-Language-Action models, LLMs, and diffusion networks — without cloud dependency.
- Blackwell GPU with Multi-Instance GPU (MIG) — The 2560-core Blackwell GPU with 5th-gen Tensor Cores and MIG technology partitions compute resources across parallel inference pipelines, guaranteeing deterministic latency for simultaneous robotics workloads.
- 128 GB Unified LPDDR5X Memory at 273 GB/s — The largest memory pool in any Jetson platform eliminates bottlenecks for multi-modal AI models, giving perception, planning, and control workloads room to coexist on a single device.
- 14-Core Arm® Neoverse®-V3AE CPU at 2.6 GHz — A high-performance CPU cluster handles real-time control loops, sensor preprocessing, and OS tasks in parallel with GPU inference — no external compute nodes required.
- 3rd-Gen Programmable Vision Accelerator (PVA v3) — Dedicated vision hardware offloads stereo depth estimation, optical flow, and feature detection from the GPU, freeing Blackwell cores entirely for AI inference while maintaining sub-millisecond perception latency.
- 100 Gbps QSFP28 Networking — The onboard 4× 25GbE QSFP28 port delivers up to 100 Gbps aggregate bandwidth for high-speed sensor fusion, multi-camera streaming, and tethered robot communication — no external network card required.
- PCIe Gen 5 Expansion — M.2 Key M (×4 PCIe Gen5) and M.2 Key E (×1 PCIe Gen5) slots provide next-generation storage and wireless throughput, future-proofing the platform for high-bandwidth peripherals.
- Scalable 40–130 W Power Envelope — Configurable TDP from 40 W to 130 W lets you balance performance and power for mobile, tethered, or lab deployments — all from a single developer kit form factor.
- Humanoid Robot-Ready Design — Engineered for seamless integration with existing humanoid robot platforms, featuring CAN headers, JTAG, automation headers, and a compact 243 × 112 × 57 mm chassis for rapid tethered prototyping.
- Complete NVIDIA AI Software Stack — Ships with JetPack 7, NVIDIA Isaac for robotics, Holoscan for sensor processing, and Metropolis for visual AI agents — the entire ecosystem for physical AI development in one kit.
Technical Specifications
| Specification | Details |
| AI Performance | 2070 TFLOPS (FP4 Sparse) / 1035 TFLOPS (FP8) |
| GPU | 2560-core NVIDIA Blackwell architecture GPU with 5th-gen Tensor Cores, Multi-Instance GPU (MIG) with 10 TPCs |
| GPU Max Frequency | 1.57 GHz |
| CPU | 14-core Arm® Neoverse®-V3AE 64-bit CPU, 1 MB L2 cache per core, 16 MB shared L3 cache |
| CPU Max Frequency | 2.6 GHz |
| Vision Accelerator | 1× PVA v3 (3rd-Generation Programmable Vision Accelerator) |
| Memory | 128 GB 256-bit LPDDR5X at 273 GB/s |
| Storage | 1 TB NVMe (M.2 Key M slot, PCIe Gen5 ×4) |
| Video Encode | 2× NVENC |
| Video Decode | 2× NVDEC |
| Camera | High-speed camera via QSFP slot, USB camera support |
| PCIe | M.2 Key M (×4 PCIe Gen5), M.2 Key E (×1 PCIe Gen5) |
| USB | 2× USB-A (3.2 Gen2), 2× USB-C (3.1) |
| Networking | 1× 5GbE RJ45, 1× QSFP28 (4× 25GbE, up to 100 Gbps aggregate) |
| Display | 1× HDMI 2.0b, 1× DisplayPort 1.4a |
| Other I/O | QSFP connector, M.2 Key E & M slots, CAN headers, automation headers, LED, JTAG, fan connector, audio header, power jack, RTC connector |
| Power | 40 W – 130 W (configurable TDP modes) |
| Mechanical | 243.19 mm × 112.40 mm × 56.88 mm |
| Thermal | Thermal Transfer Plate (TTP), optional fan or heatsink |
Common Applications & Use Cases
- Humanoid Robotics — The 128 GB memory pool and 2070 TFLOPS enable on-device execution of Vision-Language-Action (VLA) models, giving humanoid robots real-time perception, language-driven reasoning, and dexterous manipulation capabilities without cloud latency.
- Autonomous Mobile Robots (AMR) — MIG-isolated inference pipelines and 100 Gbps QSFP28 networking allow AMRs to simultaneously fuse LiDAR, radar, and multi-camera data at production throughput, enabling reliable navigation in complex, dynamic environments.
- Surgical & Medical Robotics — Deterministic low-latency inference on the Blackwell GPU combined with real-time CAN bus control enables surgical robots to respond to force feedback and visual cues within milliseconds of each other.
- Industrial Inspection & Quality Control — PVA v3 and dual NVENC/NVDEC engines enable parallel high-resolution visual inspection across production lines simultaneously, detecting sub-millimeter defects at line speed without sacrificing throughput.
- Edge Generative AI (LLMs & VLMs) — 2070 FP4 TFLOPS and 128 GB of unified memory make it feasible to run 70B-parameter language and vision models entirely on-device, powering embodied AI agents that reason and plan without cloud round-trips.
- Drone & UAV Perception — The 40 W minimum power mode and compact 243 × 112 × 57 mm chassis allow power-constrained aerial platforms to run full AI inference on multi-camera payloads without prohibitive weight or battery penalties.
- Real-Time Video Analytics — Dual NVENC and NVDEC hardware engines enable simultaneous encoding, decoding, and AI analysis of multiple 4K video streams for smart city surveillance, retail analytics, and live sports applications.
- Smart Manufacturing & Collaborative Robots — NVIDIA Holoscan sensor processing pipelines and real-time CAN control allow cobots to dynamically adapt to unstructured environments alongside human workers, enabling safe and productive collaboration.
- Digital Twins & Multi-Sensor Fusion — The QSFP28 and 5GbE networking, CAN headers, and 128 GB memory allow a single Jetson AGX Thor to aggregate and synchronise data from dozens of heterogeneous sensors in real time, powering accurate digital twin pipelines.
- Research & Sim-to-Real Transfer — PCIe Gen 5 expansion and high-bandwidth networking allow researchers to connect GPU clusters or high-speed storage arrays, making the kit ideal for dataset collection, model fine-tuning, and validating simulated AI behaviours on physical hardware.
What's in the Box
- NVIDIA Jetson T500 Module
- Reference Carrier Board
- Heatsink and Fan
- 1 TB NVMe SSD (pre-installed in M.2 Key M slot)
- 140 W DC Power Supply
- 802.11ax (Wi-Fi 6) Wireless NIC (pre-installed in M.2 Key E slot)
- Quick Start Guide
Note: accessories such as QSFP cables, USB peripherals, displays, cameras, and cases are sold separately and not included unless stated above.
Frequently Asked Questions
What operating systems and software frameworks are compatible with the Jetson AGX Thor?
The Jetson AGX Thor runs Ubuntu 22.04 LTS via JetPack 7, NVIDIA's official SDK for the Jetson platform. It is fully compatible with the NVIDIA AI software stack, including NVIDIA Isaac for robotics, Holoscan for sensor processing, and Metropolis for visual AI. Popular deep learning frameworks including PyTorch, TensorFlow, ONNX Runtime, and TensorRT are all fully supported on the Aarch64 platform. Containerised workflows via Docker and the NVIDIA NGC catalogue are natively supported, enabling rapid deployment of pre-optimised AI models without manual dependency management.
What are the power requirements for the Jetson AGX Thor Developer Kit?
The developer kit is powered by the included 140 W DC power supply via a barrel jack connector on the reference carrier board. The module itself operates across a configurable TDP range of 40 W to 130 W, selectable through software power modes within JetPack. The 40 W mode suits power-constrained or mobile deployments, while 130 W mode unlocks the full 2070 FP4 TFLOPS of AI compute for maximum throughput. Always use the supplied or a compatible 140 W+ adapter — underpowering under heavy inference loads can cause thermal throttling or unexpected shutdowns.
Which OS and firmware version does the Jetson AGX Thor ship with?
The Jetson AGX Thor ships with JetPack 7, NVIDIA's latest SDK built specifically for the Blackwell architecture and generative AI workloads. JetPack 7 bundles the Linux kernel, board support package (BSP), CUDA, cuDNN, TensorRT, and the full suite of NVIDIA developer libraries in a single install. Firmware and OS updates are distributed via NVIDIA SDK Manager (a free tool) or directly through the Jetson APT software repository. It is strongly recommended to flash the latest stable JetPack 7 release from the NVIDIA developer portal before beginning development to ensure up-to-date drivers, security patches, and library versions.
What storage options are available, and can I upgrade the pre-installed drive?
The developer kit ships with a 1 TB NVMe SSD pre-installed in the M.2 Key M slot (PCIe Gen5 ×4), providing fast sequential throughput for large model files and datasets. The slot accepts standard 2280 or 2242 form-factor NVMe drives, so upgrading to a larger capacity is straightforward. There is no soldered eMMC on this module — the NVMe SSD is both the boot and primary storage device. External USB storage via the USB-A and USB-C ports is also supported for dataset staging, backup, or additional working space.
What additional accessories do I need to get started with the Jetson AGX Thor?
To begin developing you will need a monitor with HDMI or DisplayPort, a USB keyboard, and a USB mouse — none of which are included. A QSFP28 cable or transceiver is required to use the 100 Gbps networking port, and a standard Ethernet cable connects the 5GbE RJ45 port. The included 802.11ax Wi-Fi NIC means wireless internet is available out of the box for basic connectivity. For initial flashing, a Linux host PC with NVIDIA SDK Manager (free download) is required to write the JetPack image to the NVMe drive over USB-C.
How does the Jetson AGX Thor compare to the Jetson AGX Orin?
The Jetson AGX Thor is a generational leap over the Jetson AGX Orin, delivering up to 7.5× higher AI compute (2070 FP4 TFLOPS vs ~275 TOPS) and 3.5× better energy efficiency. Memory doubles from a maximum of 64 GB on Orin to 128 GB on Thor, while memory bandwidth increases from 204.8 GB/s to 273 GB/s. The GPU architecture upgrades from Ampere to Blackwell — adding 5th-gen Tensor Cores and full MIG support — while the CPU moves from Cortex-A78AE to the higher-performance Arm Neoverse-V3AE. Connectivity also advances with PCIe Gen 5 (up from Gen 4) and the addition of a QSFP28 100 Gbps network port absent on Orin.
What GPIO, CAN, and I/O interfaces are available for robotics integration?
The reference carrier board exposes a broad set of robotics-focused interfaces including CAN bus headers for industrial motor controllers and actuators, automation headers for digital I/O integration, and a JTAG port for hardware-level debugging and bring-up. Additional I/O includes an audio header, LED connector, fan connector for active cooling, and an RTC connector for real-time clock battery backup. USB connectivity is provided via 2× USB-A 3.2 Gen2 and 2× USB-C 3.1 ports, with simultaneous HDMI 2.0b and DisplayPort 1.4a display outputs also available.
Is the Jetson AGX Thor suitable for beginners, or is it aimed at advanced developers?
The Jetson AGX Thor is primarily designed for advanced robotics engineers and AI researchers who require maximum on-device compute for generative AI and physical AI applications. That said, JetPack 7 provides a familiar Ubuntu-based environment, and NVIDIA's extensive tutorials, sample code, and pre-built Docker containers significantly lower the entry barrier. Developers with solid Python, Linux, and machine learning fundamentals can get productive quickly using Isaac ROS and the NVIDIA Hello AI World guide. For those entirely new to embedded AI, starting on a Jetson Orin Nano or Jetson AGX Orin to build foundational skills before stepping up to the Thor is a practical approach.
What is the most common mistake developers make when first setting up the Jetson AGX Thor?
The most common mistake is running the kit at full 130 W performance mode without adequate airflow or without the fan installed — this causes thermal throttling within minutes of a sustained heavy inference workload. Always verify the heatsink is properly seated and the fan connector is plugged in before enabling high-power modes. A second frequent error is installing CUDA, cuDNN, or TensorRT from upstream x86 binaries — Aarch64 Jetson builds must come from the JetPack APT repository or NVIDIA SDK Manager, as generic x86 packages will not run on the Arm CPU. Always use the JetPack-native library sources to avoid subtle incompatibilities that can be difficult to diagnose.
Where can I find documentation, community support, and firmware updates for the Jetson AGX Thor?
Official documentation, JetPack SDK downloads, and software release notes are available on the NVIDIA Developer website at developer.nvidia.com/embedded. The NVIDIA Developer Forums (forums.developer.nvidia.com) host an active Jetson community with dedicated AGX Thor threads where NVIDIA engineers regularly respond to technical questions. The NVIDIA NGC catalogue (ngc.nvidia.com) provides pre-optimised AI model containers, reference applications, and ready-to-deploy pipelines for JetPack 7. For robotics-specific resources, the Isaac ROS GitHub repository and NVIDIA Isaac documentation site offer tutorials, ROS 2 packages, and worked examples for the most common physical AI use cases.
