0
Log In
011-40727769

Custom-Built Jetson AGX Thor and T4000 Edge Computer from ProX PC

Pro Sentinel Thor from ProX PC puts NVIDIA’s Jetson T5000 or T4000 in a rugged wall-mount box. Specs, ports, and how to pick the right module.

HomeBlogsCustom-Built Jetson AGX Thor and T4000 Edge Computer from ProX PC
Custom-Built Jetson AGX Thor and T4000 Edge Computer from ProX PC

A Jetson Thor dev kit makes a great demo. The questions start when someone asks where it mounts, what powers it, and how it behaves in a shed in May. Pro Sentinel Thor is ProX PC’s answer: a custom-built edge computer around NVIDIA’s Jetson T5000 (the module sold as Jetson AGX Thor) or the lighter T4000, designed to be bolted to a wall and left alone.

Blog image

What is Pro Sentinel Thor?

It is a robotic controller and edge AI computer built on an NVIDIA Jetson Thor module. A module is the compute card itself, CPU, GPU and memory together, and it needs a carrier board, a chassis and cooling around it before it does any work. ProX PC builds all of that: a finned aluminium body, a fan on the top heatsink, and mounting brackets for the wall.

NVIDIA made the T4000 and T5000 the same size and pin compatible, so a carrier board can be designed once and take either module. That is how one Pro Sentinel Thor design covers both.

What can the Jetson T5000 and T4000 do?

Both run on Blackwell, the GPU architecture behind the RTX 50 series, paired with Arm Neoverse-V3AE CPU cores.

Specification Jetson T4000 Jetson T5000 (AGX Thor)
AI compute (FP4, sparse) 1,200 TFLOPS 2,070 TFLOPS
GPU cores 1,536 2,560
CPU 12-core 14-core
Memory 64 GB LPDDR5X 128 GB LPDDR5X
Memory bandwidth 273 GB/s 273 GB/s
Module power range 40 to 70 W 40 to 130 W

Those TFLOPS figures are measured in FP4, a 4-bit number format. Smaller numbers mean less data to move around, so the chip completes more operations per second. Rate the same T5000 in FP8, an 8-bit format, and it shows 1,035 TFLOPS, half the headline number. So, when you compare two products, check which format sits beside the figure. A chip quoted in FP4 will look faster than one quoted in FP8 even when they perform alike.

Against the previous generation, NVIDIA puts the T5000 at 7.5 times the AI compute of AGX Orin with 3.5 times the efficiency. For a site that already runs Orin and keeps hitting its ceiling, that gap is the reason to look at Thor.

How is a custom-built Pro Sentinel Thor different from the dev kit?

A dev kit is built for a desk. A factory brings dust, vibration, an unsteady 24 V supply and nobody around to restart it at 2 a.m.

Pro Sentinel Thor is built with site conditions in mind. Here is what that looks like:

  • Power: 19 to 36 VDC through a 3-pin terminal block, a range that takes a 24 V industrial supply in stride. Plan for up to 90 W on the 64 GB build and 130 W on the 128 GB build. The default kit includes a 24 V, 9.1 A adapter.
  • Temperature: -20 to 60 °C operating, -40 to 85 °C storage, 10 to 95% humidity non-condensing.
  • Machine connections: 2 × CAN-FD, a faster version of the CAN bus that machines and vehicles use to talk to each other.
  • Network: 2 × 5 GbE LAN, plus a dual SIM socket and an M.2 slot for 4G/5G, with another M.2 slot for Wi-Fi and Bluetooth.
  • Everything else: 4 × USB 3.2, HDMI 2.0 out, an M.2 NVMe slot, TPM 2.0 security chip, built-in IMU and GNSS (satellite positioning, the family GPS belongs to).
  • Size and certification: 200 × 145 × 97 mm, 2.8 kg, wall mount, CE and FCC.

CAN was designed by Bosch in the 1980s to cut the wiring inside cars. Forty years later it still moves data between motors and controllers, which is why a robot or vehicle controller without it gets awkward fast.

T4000 or T5000: which Pro Sentinel Thor should you pick?

Memory decides which models load at all. Compute decides how fast they run and how many camera streams you can feed. Choosing too little memory is the edge AI version of buying a GPU with too little VRAM and regretting it a month later.

The T4000 gives up half the memory and roughly 40% of the AI compute, and keeps the same 273 GB/s memory bandwidth as the T5000. It also draws less, which suits sites with tighter power and heat budgets.

A rule of thumb:

  • T4000: multi-camera video analytics, robots running a few models together, deployments where power and heat are the limit.
  • T5000: large vision and language models running on-site across many streams, several models at once, and room to grow after the first rollout.

If you are unsure, send ProX PC your model and stream count. Sizing it on paper costs less than swapping a module later.

Who is Pro Sentinel Thor built for?

It targets the heaviest on-site AI work: large vision and language models across many streams, with data staying inside the building. Robotics, factory lines, vehicles and mobile setups fit well, because CAN-FD, GNSS, SIM slots and a wide DC input are already on the box.

Pro Sentinel Thor is the first custom-made Jetson Thor in India. ProX PC has deployed edge AI at scale with government PSUs, and its network of 7,500+ technicians handles deployment and maintenance across the country, so the hardware arrives with a service plan. You can view the full specs and order it directly on the Pro Sentinel Thor page.

Still prototyping on a desk? ProX PC’s DGX Spark (GB10) performance test against the RTX 5090 covers the DGX Spark across LLM, image and video generation.

Frequently Asked Questions

Is Pro Sentinel Thor the same as the Jetson AGX Thor dev kit?

It uses the same T5000 module (or the T4000), built into an industrial chassis with its own carrier board and I/O. The dev kit suits testing and development, and Pro Sentinel Thor suits real deployments.

Can it run on a 24 V supply?

Yes. Input is 19 to 36 VDC.

Does it support 4G or 5G?

Yes, through a dual SIM socket and an M.2 B-key slot for 4G/5G.

How do I order one?

Pick the 64 GB (T4000) or 128 GB (T5000) configuration on the Pro Sentinel Thor page and order it directly, or call 011-40727769 or write to sales@proxpc.com.

Share this article:

Chat with us