On-Robot Compute: Jetson Thor vs Orin Power and Safety

Q: Should I choose Jetson Thor or Jetson Orin for my robot in 2026?

Choose by the sustained compute your workload needs inside the production enclosure at the power mode your battery allows. Many AMRs running 2D lidar navigation and classic perception fit an Orin NX at 10 W to 25 W with headroom. Thor makes sense when you run large vision-language-action or multimodal models on the robot and have the power and cooling for 40 W to 130 W.

Q: Is Jetson Thor really 7.5 times faster than Orin?

That is NVIDIA's figure comparing Thor to AGX Orin on its own workloads. For your workload the ratio depends on whether your model runs in FP4, whether it has the sparsity pattern the hardware accelerates, and whether it is limited by memory bandwidth. Orin is quoted in TOPS and Thor in sparse FP4 TFLOPS, so convert and measure your own model before believing any multiplier.

Q: What are the Jetson T3000 and T2000 modules?

They are smaller Thor-architecture modules NVIDIA announced on 15 July 2026 for mass-market robots and edge AI. The T3000 offers 865 FP4 teraflops with 32 GB of memory at roughly half the size and power of the T5000, and the T2000 offers 400 FP4 teraflops with 16 GB. NVIDIA lists both as scheduled for availability in Q1 2027, with an IGX T3000 variant adding functional safety.

Q: Why does my robot run slower in the field than on the developer kit?

Usually heat. A sealed, fanless enclosure in a warm site raises the module temperature until it reduces clocks to protect itself, and sustained performance falls below the desk figure. Lower power modes chosen for battery life also cut speed. Run a thermal soak of at least 30 minutes at your hottest contracted ambient with the real mixed workload to find the true compute budget.

Q: Can the emergency stop run on the Jetson as a ROS 2 node?

It should not. The AI computer runs Linux, GPU drivers and models that cannot be proven correct, so a crash or overload would remove the safety function. Put the e-stop chain, speed and zone limits and a heartbeat watchdog on a separate safety controller running a certified real-time operating system, and let the Jetson only propose trajectories that the controller clamps.

Q: How much does on-robot compute affect battery runtime?

More than teams expect. In an illustrative mobile robot with a 1.44 kWh pack and 250 W of drive and sensor load, cutting the AI computer from 100 W to 40 W extends runtime from about 4.1 to about 5 hours, roughly 21 percent. On small robots compute can be the second-largest electrical load after the drive motors, so set the power mode from the shift budget.

Q: When should robot AI run off-board instead of on the robot?

When the task tolerates a second or more of latency and the site network is reliable, slow reasoning such as language-model task planning can run on an on-site server or in the cloud. The robot must hold a safe default when the link drops. Control loops, obstacle perception and every safety function stay on the robot, because they cannot wait for a network round trip.

Q: How much does Jetson Thor cost?

At launch NVIDIA priced the Jetson AGX Thor developer kit from $3,499 and the T5000 module from $2,999 at 1,000 units, according to its own newsroom release. NVIDIA's July 2026 post mentions high memory prices, so check current distributor quotes before setting a bill of materials. An earlier AppScale article said Thor targeted a $1,000 price point; that figure was wrong.

Satyam Kumar

Retour au blog

architecture

On-Robot Compute Is Three Budgets, Not One TOPS Number: Watts, Heat and Safety in 2026

By Satyam KumarOctober 27, 202613 min read

on-robot compute jetson thor jetson agx orin jetson t3000 jetson t5000 robot power budget thermal throttling robot thermal design functional safety partition safety controller mixed-criticality robotics autonomous mobile robots humanoid robot compute fp4 sparse tflops ros 2 real-time physical ai hardware

On-Robot Compute Is Three Budgets, Not One TOPS Number: Watts, Heat and Safety in 2026

Frequently Asked Questions

Partager l'article

Twitter LinkedIn WhatsApp

Satyam Kumar

Founder & AI Architect, AppScale LLP

Architecte AI & Cloud. J'aide les équipes à construire des systèmes qui scalent pour des millions.

LinkedIn GitHub

On-Robot Compute Is Three Budgets, Not One TOPS Number: Watts, Heat and Safety in 2026

Frequently Asked Questions

Partager l'article

Comments

Leave a comment