Jetson Thor

Practical

Jetson Thor is NVIDIA’s most powerful edge AI computing platform, purpose-built for humanoid robots, autonomous vehicles, and advanced robotics requiring datacenter-class AI at the edge. Announced at GTC 2024, Thor delivers up to 2,070 FP4 TFLOPS with native Transformer Engine support.

Why Thor?

Thor represents a generational leap designed specifically for the foundation model era:

Transformer Engine: Native FP8 support for running large vision-language-action (VLA) models
Unified memory: Up to 128GB shared between CPU and GPU
Multi-modal AI: Run perception, planning, and control models simultaneously
Humanoid-ready: Designed for the compute demands of next-gen robots

Specifications

Spec	T5000	T4000
AI Performance	2,070 FP4 TFLOPS (sparse)	1,200 FP4 TFLOPS
GPU Architecture	NVIDIA Blackwell	NVIDIA Blackwell
GPU Cores	2,560 CUDA, 96 Tensor (5th gen)	1,536 CUDA, 64 Tensor
Transformer Engine	Yes (FP4/FP8)	Yes (FP4/FP8)
CPU	14-core Arm Neoverse V3AE @ 2.6 GHz	12-core Arm Neoverse V3AE
Memory	128GB LPDDR5X unified	64GB LPDDR5X unified
Memory Bandwidth	273 GB/s	273 GB/s
Power	40W - 130W (configurable)	40W - 70W
Process	4nm	4nm

Architecture

┌──────────────────────────────────────────────────────────────────┐
│                       Jetson Thor SoC (T5000)                     │
├──────────────────────────────────────────────────────────────────┤
│  ┌────────────────────┐  ┌────────────────────────────────────┐  │
│  │   Arm CPU          │  │         NVIDIA Blackwell GPU       │  │
│  │   Neoverse V3AE    │  │  ┌──────────┐  ┌───────────────┐  │  │
│  │   14 cores @2.6GHz │  │  │  CUDA    │  │  Transformer  │  │  │
│  │                    │  │  │  Cores   │  │  Engine       │  │  │
│  └────────────────────┘  │  │  2,560   │  │  FP4/FP8      │  │  │
│                          │  └──────────┘  └───────────────┘  │  │
│  ┌────────────────────┐  │  ┌──────────┐  ┌───────────────┐  │  │
│  │  Safety Island     │  │  │  Tensor  │  │  RT Cores     │  │  │
│  │  Lockstep cores    │  │  │  Cores   │  │  Ray tracing  │  │  │
│  │  ASIL-D capable    │  │  │  96 (5th)│  │               │  │  │
│  └────────────────────┘  └──┴──────────┴──┴───────────────┴──┘  │
│  ┌──────────────────────────────────────────────────────────┐   │
│  │           128GB LPDDR5X Unified Memory (273 GB/s)        │   │
│  └──────────────────────────────────────────────────────────┘   │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────────┐  │
│  │  NVDLA v3   │  │  PVA v3     │  │  Video: 8K60 decode     │  │
│  │  (2x)       │  │  (2x)       │  │  4K120 encode           │  │
│  └─────────────┘  └─────────────┘  └─────────────────────────┘  │
└──────────────────────────────────────────────────────────────────┘

Key Innovations

Blackwell GPU: Latest architecture with 5th-gen Tensor Cores
Transformer Engine: Hardware-accelerated FP4/FP8 with dynamic switching for LLM/VLA inference
Safety Island: Dedicated lockstep cores for functional safety (ASIL-D)
Unified Memory: CPU and GPU share up to 128GB pool—no copying overhead

Target Applications

Humanoid Robots

Full-body control, real-time VLA models, multi-camera perception at 800+ TOPS

Autonomous Vehicles

Level 4/5 autonomy with functional safety, sensor fusion, redundant compute

Industrial Manipulation

High-DOF arms, force feedback, real-time path planning with foundation models

Medical Robotics

Surgical assistance, diagnostic AI, safety-critical applications

Software Stack

┌─────────────────────────────────────────────────────────────┐
│                    Your Application                          │
├─────────────────────────────────────────────────────────────┤
│  Isaac Lab │ Isaac ROS 4.0 │ Omniverse │ cuMotion │ OSMO   │
├─────────────────────────────────────────────────────────────┤
│  TensorRT 10.13 │ cuDNN 9.12 │ CUDA 13.0 │ Triton Server   │
├─────────────────────────────────────────────────────────────┤
│                  JetPack 7.1 SDK                             │
├─────────────────────────────────────────────────────────────┤
│        Linux Kernel 6.8 LTS + Ubuntu 24.04 LTS              │
└─────────────────────────────────────────────────────────────┘

JetPack 7.1
Isaac ROS 4.0

# JetPack 7.1 - Latest Thor SDK (Jetson Linux 38.4)
# Ubuntu 24.04, CUDA 13.0, TensorRT 10.13
# Transformer Engine support included

# Flash Thor developer kit
sudo ./flash.sh jetson-thor-devkit internal

# Install full SDK
sudo apt update
sudo apt install nvidia-jetpack

# Isaac ROS 4.0 generally available for Thor
# Includes GR00T foundation model support

# Add Isaac ROS apt repository
sudo apt-add-repository ppa:nvidia/isaac-ros
sudo apt update

# Install Isaac ROS packages
sudo apt install ros-jazzy-isaac-ros-core
sudo apt install ros-jazzy-isaac-ros-visual-slam
sudo apt install ros-jazzy-isaac-ros-gr00t

Transformer Engine for Robotics

Thor’s Transformer Engine enables running foundation models at the edge:

import torch
import transformer_engine.pytorch as te

# Run GR00T-style VLA model on Thor
class RobotPolicy(torch.nn.Module):
    def __init__(self):
        super().__init__()
        # FP8 automatic mixed precision
        self.vision_encoder = te.Linear(768, 1024)
        self.transformer = te.TransformerLayer(
            hidden_size=1024,
            ffn_hidden_size=4096,
            num_attention_heads=16,
        )
        self.action_head = te.Linear(1024, 32)  # Joint commands

    def forward(self, images, proprioception):
        # Runs in FP8 automatically on Thor
        x = self.vision_encoder(images)
        x = self.transformer(x)
        return self.action_head(x)

Thor vs Orin Comparison

Aspect	Jetson Thor (T5000)	Jetson AGX Orin
AI Performance	2,070 FP4 TFLOPS	275 TOPS
GPU Architecture	Blackwell	Ampere
Transformer Engine	Yes (FP4/FP8)	No
Max Memory	128GB	64GB
Memory Bandwidth	273 GB/s	204 GB/s
Power Range	40-130W	15-60W
Foundation Models	Native support	Limited
Target	Humanoids, L4/5 AV	AMRs, drones, industrial

Development Workflow

Simulate in Isaac Sim: Train and validate with Omniverse digital twin
Develop on DGX Spark: Use desktop supercomputer for model development
Deploy to Thor: Seamless transition with JetPack 7.1 compatibility
Scale with OSMO: Orchestrate fleets across edge and cloud

Getting Started

1. Order Developer Kit

Thor developer kits available through NVIDIA partners. Includes:

Jetson Thor module (128GB)
Developer carrier board
Power supply (200W)
Cooling solution

2. Flash and Setup

# Download JetPack 7.1 from NVIDIA
# Use SDK Manager or command line
sudo ./flash.sh jetson-thor-devkit internal

# After boot, verify
tegrastats
nvidia-smi

3. Run Benchmark

import torch
import time

# Verify Transformer Engine
device = torch.device('cuda')
x = torch.randn(32, 1024, 4096, device=device, dtype=torch.float16)

# Measure FP8 inference
with torch.cuda.amp.autocast(dtype=torch.float8_e4m3fn):
    start = time.time()
    for _ in range(100):
        y = torch.nn.functional.linear(x, torch.randn(4096, 4096, device=device))
    torch.cuda.synchronize()
    print(f"FP8 throughput: {100/(time.time()-start):.1f} iter/s")

Jetson Orin Previous-gen platform (275 TOPS), now in LTS

DGX Spark Desktop AI supercomputer for development

Isaac ROS GPU-accelerated ROS 2 packages

GR00T Foundation model for humanoid robots

Learn More

Sources

NVIDIA Jetson Thor Official Page — Specifications and module variants
NVIDIA Newsroom: Jetson Thor Now Available — GA announcement and pricing
NVIDIA Developer Blog: Introducing Jetson Thor — Architecture details
JetsonHacks: JetPack 7.1 and T4000 — Latest SDK and T4000 availability
CNX Software: Jetson AGX Thor Developer Kit — Developer kit specs verification