NVIDIA Jetson Orin vs Thor: Which to Choose for Robotics
Choose Jetson Thor for humanoids, vision-language-action (VLA) models and advanced multi-sensor robots that need server-class compute on-device. Choose Jetson Orin (Nano or AGX) for prototypes, vision AI and most current production robots, where its compute is sufficient at lower cost and power. The IGX variants add functional safety for industrial and medical use.
Key takeaways
- Thor is the 2026 flagship: Blackwell GPU, 128 GB, for VLA and humanoids.
- Orin (Nano / AGX) remains the practical, lower-cost choice for most robots.
- Pick by model size, sensor count, power budget and safety needs.
- IGX Thor / IGX Orin add ISO 26262 and IEC 61508 functional safety.
- Orin devkits ship from stock; Thor is on allocation and quoted per order.
Spec comparison
The headline gap is compute and memory: Jetson AGX Thor delivers roughly 7.5× the AI performance of AGX Orin and about 3.5× better energy efficiency, per NVIDIA, with twice the memory. Figures below are approximate and vary by SKU and power mode.
| Spec | Jetson Orin Nano | Jetson AGX Orin | Jetson AGX Thor |
|---|---|---|---|
| GPU architecture | Ampere | Ampere | Blackwell (MIG) |
| Peak AI compute | ~67 TOPS (INT8) | ~275 TOPS (INT8) | ~2,070 TFLOPS FP4 / ~1,035 FP8 |
| CUDA / Tensor cores | 1,024 / 32 | 2,048 / 64 | 2,560 / 96 (5th-gen) |
| CPU | 6× Arm A78AE | 12× Arm A78AE | 14× Arm Neoverse V3AE |
| Memory | 8 GB | 32–64 GB | 128 GB LPDDR5X |
| Mem. bandwidth | ~102 GB/s | ~204 GB/s | ~273 GB/s |
| Power envelope | 7–25 W | 15–60 W | 40–130 W |
| Typical role | Prototypes, smart cameras | Production robots, AMRs | Humanoids, on-device VLA |
The practical reading: Orin counts AI throughput in TOPS for INT8 perception; Thor counts it in TFLOPS at FP4/FP8 because it is built to run transformer-scale vision-language-action models on the robot. The 128 GB of memory is the other deciding factor — large VLA models simply do not fit on an 8–64 GB Orin.
Which to choose, by scenario
The decision is rarely "the fastest chip" — it is the cheapest, lowest-power module that runs your model within the control loop. Map your workload to the table below.
| Your build | Recommended module | Why |
|---|---|---|
| Proof of concept / smart camera | Jetson Orin Nano | Cheapest, lowest power; runs classic detection/segmentation |
| AMR / AGV, industrial inspection | Jetson AGX Orin | Multi-camera perception + control at 15–60 W; from stock |
| On-device VLA / reasoning VLM | Jetson AGX Thor | FP4/FP8 compute and 128 GB fit transformer-scale models |
| Humanoid / advanced autonomy | Jetson AGX Thor | Heavy multi-sensor fusion in real time |
| Safety-critical industrial / medical | IGX Thor or IGX Orin | Functional safety (ISO 26262, IEC 61508) + long lifecycle |
When to choose Orin
Orin is the right call when the model fits and power or budget is tight. Orin Nano covers prototypes and proofs of concept; AGX Orin covers the large installed base of production robots running classic perception plus a control policy — AMRs, pick-and-place cells, inspection rigs. It is cheaper, runs at lower power, and is widely available from stock, which is why it remains the practical choice for most robots shipping today. If your model is a CNN or a small policy network and you are not running a large language or VLA model on-board, Orin is almost always enough.
When to choose Thor
Thor earns its premium in one situation: when the "decide" step of the robot is a large model. That means an on-device vision-language-action (VLA) model, a reasoning VLM such as NVIDIA Cosmos Reason, or real-time fusion of many high-bandwidth sensors (several cameras plus LiDAR and radar). These need both the FP4/FP8 throughput and the 128 GB of memory that Orin cannot provide. Humanoids and advanced autonomous machines are the typical case — see humanoid robots in industry and what VLA models are. The trade-off is power (up to ~130 W) and cost, so do not reach for Thor unless the model genuinely requires it.
IGX: when you need functional safety
The IGX line is a separate axis from raw speed. Choose it for industrial and medical deployments that must meet functional-safety standards (ISO 26262, IEC 61508) and carry long, supported lifecycles — IGX Thor for new high-compute builds, IGX Orin where Orin-class compute is enough. If a person can be hurt by the machine or the deployment must be certified, the safety platform matters more than the benchmark.
Migrating from Orin to Thor
Thor is a different architecture (Blackwell, Arm Neoverse), so a model tuned for Orin is not automatically optimal on Thor. Plan for re-profiling and re-optimizing the runtime (TensorRT), validating power and thermals at the higher envelope, and re-checking the carrier board and I/O. The upside is headroom: a robot that is memory-bound on AGX Orin today can move to a larger VLA model on Thor without redesigning the rest of the system.
Availability and cost
Orin is the budget and lead-time winner; Thor is the capability winner on allocation.
| Module | Indicative price | Availability |
|---|---|---|
| Jetson Orin Nano devkit | from ~$249 | Typically from stock |
| Jetson AGX Orin devkit | from ~$1,999 | Frequently from stock |
| Jetson AGX Thor devkit | Quoted per order | On allocation |
Production module volumes are quoted against NVIDIA supply with realistic lead times. Haink configures and supplies all of these as edge inference nodes alongside sensors and controllers — see robotics & physical-AI hardware and edge AI inference.
Frequently asked questions
Is Jetson Thor better than Orin?
Thor has far more compute and memory and is better for VLA and humanoid workloads, but Orin is cheaper, lower-power and sufficient for most current robots. The best choice depends on model size, sensors, power and budget.
Should I buy Jetson Thor or Orin?
Buy Thor for humanoids, VLA models or heavy multi-sensor fusion; buy Orin for prototypes, vision AI and most production robots where its compute is enough.
What is the difference between IGX and Jetson?
Jetson modules are general robotics edge compute; IGX platforms add functional safety (ISO 26262, IEC 61508) and long lifecycle support for industrial and medical deployments.
Is Jetson Thor available to buy?
Thor is supplied on allocation and quoted per order with realistic lead times, whereas Orin Nano and AGX Orin devkits typically ship from stock.
How much faster is Jetson Thor than Orin?
NVIDIA cites roughly 7.5× the AI compute and about 3.5× better energy efficiency for AGX Thor versus AGX Orin, with twice the memory (128 GB vs up to 64 GB).
Can I run a VLA model on Jetson Orin?
Small or heavily optimized models can run on AGX Orin, but large vision-language-action models are usually memory- and compute-bound on Orin; AGX Thor's FP4/FP8 throughput and 128 GB are designed for them.
How much does a Jetson devkit cost?
Approximately: Orin Nano devkit from around $249, AGX Orin devkit from around $1,999; Thor is quoted per order.