Skip to main content

NVIDIA Jetson Thor (Blackwell Edge)

Overview

The NVIDIA Jetson Thor is the next-generation edge AI module released in 2025, based on the Blackwell architecture GPU. With AI compute up to 2,070 FP4 TFLOPS (sparse), it delivers 7.5× the performance of the previous-generation AGX Orin (275 TOPS). Jetson Thor is purpose-built for humanoid robots, autonomous machines, and physical AI, capable of running LLMs, VLMs (vision-language models), and VLAs (vision-language-action) models at the edge.

Core Specifications

SpecJetson AGX Thor Developer KitJetson T5000Jetson T4000
AI Compute (FP4 sparse)2,070 TFLOPS2,070 TFLOPS1,200 TFLOPS
GPU2,560-core Blackwell + 5th Gen Tensor Cores2,560-core Blackwell + 5th Gen Tensor Cores1,536-core Blackwell + 5th Gen Tensor Cores
MIG10 TPC10 TPC6 TPC
CPU14× Arm Neoverse-V3AE 64-bit14× Arm Neoverse-V3AE 64-bit12× Arm Neoverse-V3AE 64-bit
CPU FrequencyUp to 2.6 GHzUp to 2.6 GHzUp to 2.6 GHz
L3 Cache16 MB shared16 MB shared16 MB shared
Memory128 GB LPDDR5X128 GB LPDDR5X64 GB LPDDR5X
Bandwidth273 GB/s273 GB/s273 GB/s
Power40-130 W40-130 W40-70 W
Networking1× 5GbE + 1× QSFP28 (4× 25GbE)4× 25GbE3× 25GbE
PCIeM.2 Key M Gen5 x4Gen5 x8Gen5 x8
Vision Accelerator1× PVA v31× PVA v31× PVA v3
Video Codec2× NVENC + 2× NVDEC2× NVENC + 2× NVDEC1× NVENC + 1× NVDEC

Vendor Information

ItemDetail
ManufacturerNVIDIA Corporation
Websitehttps://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/
ReleaseAugust 2025 (GA)
EcosystemIsaac GR00T N1 (humanoid robot foundation model), VSS (video search and summarization), Holoscan Sensor Bridge

Key Features

  • 7.5× AGX Orin compute, 3.5× energy efficiency
  • Supports LLM / VLM / VLA models running at the edge
  • 128GB memory enables 70B+ model loading
  • 5× 25GbE networking for multi-machine coordination
  • Holoscan Sensor Bridge for real-time multi-sensor fusion

Use Cases

  • Humanoid robots (Isaac GR00T)
  • Autonomous mobile robots (AMR)
  • Edge LLM inference
  • Multimodal AI (video + voice + decision-making)
  • Low-latency real-time inference