NVIDIA IGX Thor
NVIDIA IGX Thor is the second-generation industrial edge AI platform in the IGX family, unveiled on October 28, 2025 at GTC Washington DC. The successor to NVIDIA IGX Orin (1st gen) — based on the NVIDIA Blackwell architecture, it delivers up to 8× higher AI compute on iGPU, 2.5× higher on dGPU, and 2× better connectivity. Dedicated to physical AI applications in industrial and medical environments requiring low latency, deterministic response, functional safety and long-term support.
Dual GPU architecture
A signature feature of IGX Thor is its dual NVIDIA Blackwell GPU: an integrated GPU (iGPU) on the compute module — efficient for sensing and control tasks — and a discrete GPU (dGPU) for the demanding compute of running LLMs, VLMs and other large generative models directly at the network edge. Combined: 5,581 FP4 teraflops of AI compute with 400 GbE connectivity.
Production variants
NVIDIA IGX T5000 — production module, the same variant used among others in the industrial humanoid AGIBOT G2 as the local inference engine for the GO-1 + GE-1 stack (2,070 FP4 TFLOPS in that specific deployment).
NVIDIA IGX T7000 — production-ready board kit (carrier + dGPU + interfaces) for system integrators. Available from December 2025.
NVIDIA IGX Thor Developer Kit Mini — developer variant for integrators and researchers (small desktop form factor).
Software stack
The platform runs the full NVIDIA AI Enterprise stack: NVIDIA NIM microservices (accelerated physical AI deployments from cloud to edge), NVIDIA Isaac (robotics), NVIDIA Metropolis (visual AI), NVIDIA Holoscan (sensor processing). It integrates elements of NVIDIA Halos — a full-stack safety system — for functional safety in medical and industrial robotics, leveraging both onboard and infrastructure sensors to provide traditional and outside-in safety.
Early adopters
Industrial, robotics, and medical leaders already adopting IGX Thor: Diligent Robotics (Moxi 2.0), EndoQuest Robotics, Hitachi Rail (predictive maintenance and autonomous inspection of rail networks), Joby Aviation (autonomous flight), Maven Robotics (general-purpose industrial robots), SETI Institute. CMR Surgical is evaluating IGX Thor for its robotic surgical systems. OEM ecosystem: Advantech, ADLINK, ASRock Rack, Barco, Connect Tech, Curtiss-Wright, Dedicated Computing, EIZO, Inventec, NexCOBOT, Onyx, WOLF Advanced Technology, YUAN — providing carrier boards, edge servers and accessories for IGX Thor.
Classification
Industrial edge AI platform · serves as: AI Inference, Compute, High-level compute, Computer vision processing. · compatible with: Humanoid, Manipulator, Industrial Robot.
Which group NVIDIA IGX Thor belongs to and how it is built
Compute Modules is a subcategory of hardware components that provide processing power for robotic systems. It encompasses onboard computers, single-board computers (SBCs), AI accelerators, embedded processors, GPU/NPU compute modules, and other units responsible for processing sensor data and executing control logic. These modules form the foundation of modern autonomous, humanoid, and perception-capable robots.
A component type covering industrial-class edge AI platforms — unlike single developer SoMs (e.g. Jetson AGX Thor) or data-center accelerator cards, an industrial platform combines: (1) a high-performance compute module (typically ARM + iGPU), (2) an optional discrete GPU for added compute, (3) high-bandwidth sensor interfaces and 100+ GbE networking, (4) functional safety certifications (e.g. IEC 61508, ISO 13849), (5) long-term hardware and software support (typically 7-10 years) for production deployments. Classical applications: surgical robots, medical imaging devices, intelligent industrial systems, autonomous mobile machines.
A design class covering edge AI platforms with a specific hybrid architecture: simultaneous use of an integrated GPU (iGPU, part of the SoC/SoM for efficient sensing and control tasks) and a discrete GPU (dGPU, for demanding compute — running LLMs, VLMs and large vision models at the network edge). The carrier board integrates high-level interfaces: 100+ GbE, PCIe, high-bandwidth sensor inputs, and external functional-safety signals (E-stop, watchdog, safety MCU). Certified against IEC 61508 (SIL), ISO 13849 (PL) or ISO 26262 (ASIL) depending on the application domain. The first representative of this class: NVIDIA IGX (Orin, Thor).
Relations
Other hardware parts related to NVIDIA IGX Thor
NVIDIA flagship compute module for humanoids and physical AI - 2070 TFLOPS FP4, Blackwell GPU, 128 GB LPDDR5X.
NVIDIA FPGA platform for low-latency sensor data streaming to GPU over Ethernet. Bridges cameras and sensors to IGX/AGX via GPUDirect RDMA. Open-source Apache 2.0, latest release 2.6 (June 2026).
Software
IGX Thor runs the full NVIDIA AI Enterprise stack: NVIDIA NIM microservices (inference microservices), NVIDIA Isaac (robotics, including Isaac GR00T N — VLA models for humanoids), NVIDIA Cosmos Reason (LLM/VLM), Metropolis (visual AI), Holoscan (sensor processing), Halos (functional-safety AI agents). Sim-to-real workflow: train models in Isaac Lab/Isaac Sim, deploy on production IGX Thor. 10-year LTS support for medically and industrially certified deployments.