How to Choose an ARM-Based Single Board Computer: 3 Scenarios, 3 Answers

There's no single "best" ARM-based single board computer. That's the part that cost me time and money before I figured it out.

I've been selecting hardware for industrial IoT and edge computing deployments since 2017. In my first year, I made the classic specification error: I assumed "powerful" meant "better" and "cheaper" meant "smarter." I've personally made—and documented—14 significant mistakes since then, totaling roughly $26,000 in wasted budget. Now I maintain our team's pre-purchase checklist so we don't repeat them.

If you're trying to decide between a mobile edge computing 5G node, a low power single board computer, or an industrial control board, the right answer depends on which scenario you're actually in. Let me walk you through the three I keep seeing, and the mistakes that taught me each one.

The right single board computer isn't the most powerful one, or the cheapest one. It's the one whose failure mode you can live with.

Scenario 1: Mobile Edge Computing with 5G

This is where people ask me about the Rockchip RK3588 price before they ask me anything else. I get it—the chip looks great on paper: 8 cores, an NPU with up to 6 TOPS, PCIe 3.0 for a 5G module. It's become the default answer for mobile edge computing 5G workloads, and honestly, it's often a good one.

But here's what I learned the hard way: the RK3588 price is the smallest line item in your budget. In 2022, I chose a cheaper, lower-power board for a smart-city camera project. It had enough CPU for the image pipeline but no NPU—or rather, no NPU that could run our model without significant quantization losses. The AI inference ended up on the CPU. By the time we added an external NPU module, a bigger heatsink, and a custom carrier, we'd spent about 40% more than the RK3588-based industrial single board computer we'd originally dismissed.

What I should have calculated was TCO—total cost of ownership. Not just the unit price. TCO includes:

• Development time: with the RK3588 board, our team got the 5G module and the vision pipeline running in about two weeks less.
• Thermal management: more compute means more heat, but a board that throttles under load costs you more than the heatsink ever will.
• Integration costs: carrier boards, module certification, power sequencing.
• Ongoing support: how active is the BSP, how fast do security patches come out.

If your project involves video inference, on-device AI, or high-bandwidth 5G uploads, buy the high-performance board. But don't buy it because it's "powerful." Buy it because it's often the cheapest path to a working system—which is the only price that actually matters.

Scenario 2: Low-Power and Battery-Powered Deployments

This is where I made the opposite mistake—buying too much board.

In 2023, I spec'd an RK3588 board for a remote environmental monitoring device. Battery-powered, solar charging, sending a small telemetry packet every 15 minutes. The device worked, but the board drew around 3.2W at idle. The battery bank we'd planned for lasted 11 hours instead of two days. We had to redesign the power system, reschedule the pilot, and eat the cost of 20 battery packs already cut to size.

A low power single board computer would've been the right call. A board like the Rockchip RK3308 or an STM32MP1-based SBC, drawing under 1W at idle, was more than enough for telemetry. That project didn't need 6 TOPS. It needed low idle drain and reliable sleep modes.

Here's the TCO math: the RK3588 board I bought cost $180. The low-power board that would've worked cost about $90. But the real cost wasn't the $90 difference—it was the power redesign, the battery replacement, and a two-week slip. That ran us about $2,300. After that, I started putting "idle power consumption" at the top of our spec sheet, right above "cores."

So ask yourself: will the device sit idle most of the time? Is it line-powered or battery-powered? If the answer is battery, the low power single board computer wins almost every time—regardless of how impressive the high-end board's benchmarks look. At least, that's been my experience across about 30 deployment projects.

Scenario 3: Industrial Control and Furnace Control Systems

This one's less glamorous, but it's where I've seen the most money disappear.

An industrial single board computer for furnace control—or what we call a furnace control board carrier in the field—has different priorities than an edge AI box. And before I go further: I'm not an EE, so I can't speak to the electrical design of a carrier board. What I can tell you from a systems integration perspective is where real-world requirements diverge from the spec sheet. It needs:

• Deterministic behavior: no unexpected reboots, no thermal throttling mid-cycle.
• Industrial interfaces: RS-485, Modbus, CAN, and multiple analog/digital I/O.
• Wide operating temperature: often -20°C to 70°C or worse, including heat inside a control enclosure.
• Long-term availability: you want the same board, same pinout, same firmware behavior in 5 years—not a new revision every 18 months.

My mistake in 2021: I used a consumer-grade development board for a furnace controller prototype. It worked perfectly in the lab at 21°C. On the factory floor, in late-July ambient heat, it started resetting. The thermal design was fine for a desk—not for an enclosure sitting a few feet from a furnace.

Per FTC guidelines (ftc.gov), performance claims need to be substantiated with evidence. When a vendor says "industrial temperature range" or "wide voltage input," that's a claim you should verify—not just take at face value. I now ask for thermal test reports and EMC certification results before we commit. If a vendor can't produce them, that's a red flag, no matter how clean the spec sheet looks.

That 2021 prototype failure cost us about $890 per unit in field service calls and replacement downtime. Four units, two service visits each: we were out roughly $7,120—on a "cost-saving" board that was $80 cheaper than the industrial option. The expensive board was the cheap one all along.

For furnace control and similar industrial applications, my advice is deliberately boring: choose a board that's been on the market for at least a year, has documented firmware updates, and comes from a supplier that commits to multi-year availability. The excitement of a new chip isn't worth a shutdown.

How to Tell Which Scenario You're Actually In

The key question isn't "which board is best." It's "which failure mode hurts me most?"

• Scenario 1—mobile edge computing 5G: if your device needs AI inference, video processing, or high-bandwidth 5G uploads, and has a reasonable power budget, go with a high-performance ARM-based single board computer like the RK3588. Verify 5G module support and check BSP maturity before you buy.
• Scenario 2—low power: if your device runs on battery, spends most of its life asleep, and only wakes to send small data, pick a low power single board computer. Don't let the high-end spec sheet distract you.
• Scenario 3—industrial control: if your device controls machinery, heating, or any process where failure means a shutdown, buy the industrial single board computer with proper temperature ratings, long lifecycle, and a furnace control board carrier that's proven in the field.

I keep a checklist now, after the third time I ordered the wrong board. Before any purchase, I verify:

1. Idle and load power consumption—not just "typical" from the datasheet.
2. Working temperature range and thermal design under maximum load.
3. OS and driver support maturity: can I upgrade without breaking things?
4. Supply reliability: will this chip and board still be available in 3–5 years?
5. Actual TCO: board + carrier + power over 5 years + integration time + expected failure cost.

In the interest of timing: the prices and availability I've mentioned reflect what I saw through Q1 2025. The SBC market moves fast, so verify current board revisions and distributor pricing before you finalize a budget.

The Rockchip RK3588 price is easy to quote. The low-power board's price is easy to quote. But easy numbers are rarely the ones that matter. I've wasted enough budget on the gap between "cheaper" and "actually cheaper." Calculate the full cost, verify the claims, and pick based on your scenario—not based on what looks impressive in a benchmark chart.