Raspberry Pi, Jetson or a microcontroller? Choosing the host computer for a USB radiometric thermal module
Once a module works on the desk, the next question is which computer will sit next to it in the product: a Raspberry Pi, a Jetson, an x86 mini PC, or even a microcontroller board. The answer comes from how much data the module sends and what you do with it. Here is the arithmetic, with MINI 640-2 Radiometric as the example. Data rates, transfer times and file sizes below are calculated, not measured.
How much data comes out
A 640 × 512 frame has 327,680 pixels. At two bytes per pixel, one frame of temperature data is about 0.66 MB, and at 30 frames per second that is about 20 MB/s for the temperature data alone; any image data sent with it adds to that.
USB 2.0 high speed is rated at 480 Mbit/s. A single isochronous stream (the reserved-bandwidth mode many USB cameras use) is capped by the protocol at about 24.6 MB/s. Bulk transfers, which other USB cameras use, reach about 35–40 MB/s in practice, but only get the bandwidth other devices leave free. The module uses a large share of one USB 2.0 link, so:
- Give the module a USB link with nothing else busy on it. On some boards several ports share one link (on the Pi 4, all USB 2.0 devices share one, whichever port they use), so a USB 2.0 camera, disk or network adapter elsewhere can still cost you frames.
- Keep the cable within the USB 2.0 limit of 5 m, and avoid cheap hubs.
- On Raspberry Pi boards up to the 3B+, Ethernet itself runs over the board’s single USB 2.0 link, so sending data out over Ethernet competes with the module. The Pi 4 and Pi 5 do not share it.
Why a typical microcontroller is not enough
Most microcontroller boards have no USB host at all, or a full-speed one at 12 Mbit/s, about 1 MB/s in practice. That is a small fraction of the stream, and one 0.66 MB frame is also larger than the RAM of most microcontrollers, so a typical Arduino-class board cannot take the USB video or temperature frames.
The serial port is no way round it: at 115,200 baud a UART moves about 11.5 KB/s, so one 640 × 512 frame of 16-bit values would take about 57 seconds. The core’s UART and I²C are control ports. If your design needs only a few numbers on a microcontroller, ask the supplier exactly what the serial protocol provides before you commit.
A microcontroller with a parallel camera port is a different route: it may be able to capture the DVP version of the module, if its camera port handles the module’s pixel clock and logic level (the DVP port defaults to 1.8 V, switchable to 3.3 V) and it has memory for a full frame.
Raspberry Pi, Jetson or an x86 mini PC
All three run Linux and have high-speed USB host ports, so the choice depends on what else the computer has to do:
- Raspberry Pi: low cost and low power, common in monitoring units and data loggers. Check the CPU headroom for your own processing on the model you pick, and give the Pi its own power supply.
- Jetson: when you will run neural networks on the thermal image, or combine it with other cameras.
- An x86 mini PC with Ubuntu: when the product already has one, or the analysis is heavy and power is not tight.
TSDK, our SDK for MINI 640-2 Radiometric, has builds for Linux (Ubuntu, Raspberry Pi and Jetson), which covers all three. It has been tested on our modules, comes at no extra charge and is provided after payment.
Power
The module draws under 0.7 W at 5 V (typical, USB output, 25°C), about 140 mA (calculated), well within the 500 mA a USB 2.0 port supplies. The host is usually the bigger load, so size the power supply for the host.
Storage: decide what you keep
Full-frame temperature adds up fast. At two bytes per pixel (calculated):
- Every frame at 30 Hz: about 20 MB/s, about 71 GB per hour
- One frame per second: about 2.4 GB per hour, about 57 GB per day
- 32-bit floating-point values double these figures
Most monitoring products keep region statistics (maximum, minimum, average) all the time and save full frames only on an event or on a schedule.
Time stamps
If you combine thermal data with other sensors or cameras, stamp every frame on arrival with one host clock, and measure once how long each stream takes from capture to arrival so you can subtract it. Keep clocks synchronised (NTP) when several computers are involved. Lining data up afterwards is much harder.
Checklist
- A high-speed USB port the module does not share with other busy devices
- A Linux build of the SDK for your board
- CPU headroom for your own processing
- A storage plan: what you keep, and how often
- Clock synchronisation if you combine sensors
- A power supply sized for the host, not only the module
Related guides
Choosing the host for your product?
Send the board, the operating system and what the product does with the data. We reply within 24 hours on working days.