NVIDIA Jetson AGX Thor (T5000 Blackwell) Architecture Analysis: 2,070 FP4 TFLOPS & Hardware MIG for Humanoid VLA
Architectural analysis of the $3,499 NVIDIA Jetson AGX Thor Developer Kit (T5000 module) based on official documentation—examining 2,070 FP4 TFLOPS Blackwell compute, 128GB LPDDR5X memory budgeting, and hardware Multi-Instance GPU (MIG) isolation.
⚡ Key Engineering & Supply-Chain Takeaways
- Powered by a 2,560-core Blackwell GPU and 14-core Arm Neoverse-V3AE CPU with 128GB LPDDR5X (273 GB/s), the Jetson T5000 is rated at up to 2,070 Sparse FP4 TFLOPS across a 40W–130W power envelope.
- Hardware Multi-Instance GPU (MIG) partitioning isolates deterministic whole-body MPC/RL balance controllers from memory-intensive multi-camera Isaac GR00T VLA inference.
- Integrated 4× 25GbE QSFP28 Holoscan sensor bridge supports GPUDirect RDMA to stream multi-camera and 3D LiDAR packets directly into GPU memory without CPU interrupt overhead.
🔬 Technical & Supply-Chain Specification Matrix
| Compute & Architecture | NVIDIA Jetson T5000 (Blackwell 2,560 CUDA + 96 5th-Gen Tensor Cores, 2,070 FP4 TFLOPS) |
|---|---|
| Memory & CPU Complex | 128GB LPDDR5X (273 GB/s) + 14-Core Arm Neoverse-V3AE CPU ($3,499 DevKit) |
| Reference Memory Budget | 50.9 GB Active Humanoid VLA/SLAM/Planner Stack + 77.1 GB Free Headroom (of 128 GB) |
| MIG Real-Time Isolation | Up to 2 Independent Hardware GPU Partitions (Control Slice + Foundation Model Slice) |
Single-Module Humanoid Supercomputing With Jetson T5000
Our technical architecture review of the $3,499 NVIDIA Jetson AGX Thor Developer Kit (T5000 module)—compiled from official NVIDIA Jetson Thor documentation—examines how Blackwell's 5th-generation NVFP4 Tensor Cores and hardware Multi-Instance GPU (MIG) partitioning consolidate multi-board torso compute in bipedal humanoids:
- Hardware MIG Partitioning: By isolating a dedicated real-time MIG slice for the 1,000 Hz whole-body locomotion controller while assigning the primary GPU partition to Isaac GR00T N1.5 / OpenVLA-7B / Qwen3-32B, the humanoid's control loop is protected from VLM prompt-prefill memory bandwidth spikes.
- 128GB Unified Memory Budget (50.9 GB Active / 77.1 GB Headroom): Co-hosting Isaac GR00T N1.5 (9.2 GB), a 4-bit 32B reasoning planner (19.5 GB), cuVSLAM + 3D Gaussian map buffer (14.8 GB), ASR/TTS audio (3.4 GB), and the MIG control slice (4.0 GB) totals 50.9 GB, leaving 77.1 GB of free LPDDR5X headroom.
🔗 Related AI Hardware in RECIWO Database
All 93 Devices →NVIDIA Jetson AGX Thor Developer Kit (T5000 Blackwell)
Flagship physical AI and humanoid robotics supercomputer powered by the NVIDIA Jetson T5000 module (Blackwell GPU + 14-core Arm Neoverse-V3AE CPU + 128GB LPDDR5X), delivering 2,070 FP4 TFLOPS and 4× 25GbE Holoscan sensor bridging.
NVIDIA Jetson Orin NX Super Edge AI DevKit
Palm-sized generative AI supercomputer delivering 157 TOPS (Super Mode) and 102 GB/s memory bandwidth to run 8B multimodal LLMs and VLA robotics policies locally.
NVIDIA DGX Spark (GB10 Grace Blackwell Superchip)
Palm-sized personal AI supercomputer powered by the NVIDIA GB10 Grace Blackwell Superchip with 128GB coherent unified LPDDR5X memory and ConnectX-7 200GbE clustering.
AgiBot Lingxi X2 Multimodal Embodied Humanoid
1.31-meter, 35kg agile humanoid robot powered by AgiBot's Genie Operator-1 (GO-1) Vision-Language-Latent-Action (ViLLA) foundation model, featuring 28 compliant degrees of freedom (including a 3-DoF waist) and expressive emotional interaction.
📚 Related Deep-Dive Engineering Teardowns
Browse All 113 Articles →NVIDIA Jetson AGX Thor (T5000 Blackwell) Technical Analysis: 2,070 FP4 TFLOPS & 128GB LPDDR5X Architecture for Humanoid VLA Models
An architectural and specification analysis of the $3,499 NVIDIA Jetson AGX Thor Developer Kit (T5000 module, 2,560-core Blackwell GPU, 14-core Arm Neoverse-V3AE CPU, 128GB LPDDR5X) based on official NVIDIA documentation—examining hardware Multi-Instance GPU (MIG) partitioning, Holoscan sensor bridging, and memory budgeting for Isaac GR00T VLA workloads.
AgiBot Lingxi X2 vs EngineAI PM01 vs Unitree R1 Benchmark: ViLLA Foundation Models, 320-Degree Waist Kinematics & Straight-Knee Gait RL
RECIWO Labs evaluated the three defining mid-size bipedal humanoids of late 2026: AgiBot's 35kg Lingxi X2 powered by the Genie Operator-1 (GO-1) ViLLA foundation model, EngineAI's 40kg open-source PM01 with its 320-degree rotating waist and 2.0 m/s straight-knee human gait, and Unitree's $5,900 25kg R1.
Quantizing Vision-Language-Action (VLA) Models on NVIDIA Jetson Orin NX Super: FP8 vs. INT4 Control Loop Jitter
RECIWO Labs benchmarked OpenVLA-7B, Pi-0-Fast, and ACT policies on the 157 TOPS Jetson Orin NX Super, measuring action-chunk inference latency and grasp precision across FP16, FP8, and AWQ INT4.
