The SoC is the easiest part to get right. The carrier board is where industrial IoT projects go to die.
If you're specifying a Rockchip RK3588 CM4 for a new industrial edge device—and plenty of people are, since the module started shipping in volume around 2023—don't lose sleep over the chip specs. The RK3588 is a proven design: 8nm process, four Cortex-A76 cores, four Cortex-A55 cores, a 6 TOPS NPU, and enough PCIe 3.0, USB 3.0, and display interfaces for most industrial use cases. It runs Linux well, and Rockchip's official documentation covers the hard parts.
The risk lives in everything around the module. The circuit board carrier. The power delivery. The thermal path. The connectors. None of that shows up on a marketing slide, but it's where every field failure I've ever chased traces back to.
Not the sexy stuff. The make-or-break stuff.
Why you should listen to me
I'm a quality and compliance manager at an industrial IoT hardware company. I review every board design before it reaches production—roughly 30 to 40 designs a year, mostly carrier boards and system boards built around high-end ARM compute modules from Rockchip, NXP, and a few others. I rejected about 12% of first-pass designs in 2024. Sometimes more, depending on how you count revisions.
I don't design these boards. I'm the person who has to tell the product team that their “ready to ship” design is going back for a respin and that the launch date is slipping. I've stood in front of a whiteboard explaining why a carrier board's power sequencing would kill the module on the first power-up. That is not a fun conversation to have. This is where this opinion comes from.
The part that fails most often: power
In my experience, the #1 cause of industrial IoT device failures is not the processor, not the memory, and not the software. It's power delivery on the carrier board. The RK3588 can pull 8–10W of sustained load during edge inference. Add a couple of cameras and an NVMe drive and you're looking at 20W+ before you leave the board. The module handles its own core rails, but everything upstream—input protection, DC-DC conversion, power sequencing, brown-out behavior—lives on the carrier.
It's tempting to think that “the module handles the power stuff.” But the module only handles what happens after the carrier board does its job. I've tested supposedly industrial-grade carrier boards that brown out below -10°C because the inrush and peak-current design was validated on a lab bench at room temperature. The most frustrating part of this work: the same power issues keep showing up even when we spell out the requirements in exact detail. You'd think a written spec with current waveforms and voltage curves would prevent it. But interpretation varies wildly.
Heat is where performance goes to hide
Thermal design is the second killer. The RK3588 has a real thermal budget; if the carrier board doesn't provide a heat path, the chip will throttle. There's something sobering about watching two identical RK3588 modules run the same workload on two different carriers, and seeing one hold a steady 2.4GHz while the other drops to around 1.2GHz at 68°C because the PCB acts as a heat trap. Same silicon. Same load. Way bigger performance gap than anyone on the team expected.
What I mean is that the carrier layout—copper pour, thermal vias, the position of the module relative to other heat sources, the enclosure design—determines whether your “industrial” device actually works at 50°C inside a sealed cabinet. If the thermal path is wrong, no heatsink or software tuning will fully save you later.
There's something deeply satisfying about a carrier board that passes first bring-up. After all the review cycles, thermal modeling, and connector spec debates, the board just boots without complaining. That moment doesn't happen by accident. It happens because someone treated the carrier board as the product, not as an accessory to the SoC. That's the payoff for all the tedious review work.
Connectors: the 10,000-cycle problem
Connectors are the unglamorous place where cost gets cut. A $0.80 board-to-board connector rated for a hundred mating cycles vs. a $3.50 connector rated for 10,000. For a device that gets assembled once and deployed, that doesn't matter. But for a device that will be installed, serviced, and reworked in the field—which is most industrial IoT devices—those cycles get consumed fast. I've seen units fail in their third year of service because the Ethernet jack was chosen for price, not durability. Not ideal. Workable at prototype stage. Terrible in the installed base.
Also check the connector's datasheet, not the marketing copy. “Industrial grade” is one of the most overused phrases in this industry. A thermal camera and a careful datasheet review will debunk half of them.
And if you're tempted to postpone the certification question: the carrier board layout decides whether your device passes CE, FCC, and IEC 60068 environmental tests on the first attempt. Poorly routed high-speed interfaces (PCIe, USB 3.0, Gigabit Ethernet) cause EMI failures. Marginal thermal design causes temperature test failures. Both send you back for a respin that costs months and thousands of dollars. That's a budget line, not a technicality.
Software is a hardware decision
If the carrier board is the hardware risk, software support is the timeline risk. The RK3588 is in decent shape: mainline Linux support has improved a lot (as of early 2025, anyway), and you have options—Yocto, Debian, buildroot, or vendor BSPs. But someone has to own that software for five to ten years. That's a budget line, not an afterthought. If your team can't sustain a Linux distribution, choose the module and carrier based on how easy the software is to maintain, not just the spec sheet.
The edge/cloud decision is partly a hardware decision too. If your device can't do local preprocessing, buffering, and inference when the network drops, “edge computing” is just a nicer way to say “network-dependent.” Whether that works depends on the carrier board's storage, memory, and a power design that doesn't crash mid-write. Cloud and edge computing aren't competing architectures—they're a split. The split point is decided by the board.
What “cheap” actually costs
Here's the total cost question: what are you actually comparing?
The $89 module price is the beginning, not the total. At low-to-mid volumes, a serious industrial carrier board—wide-input power conversion, proper protection, industrial-temperature-rated passives, a layout that passed review—will cost two to four times the module. A cheaper carrier looks like a huge saving on paper.
Real example from a few years back (2022, to be exact). We had a vendor quote a $120 carrier board for a pilot run of 400 units. It beat the robust option by $45 per board on paper. Total “savings”: about $18,000. Six months in, we started seeing intermittent Ethernet dropouts on about 60 units. The failure traced back to a marginal magnetics choice on the Ethernet PHY. The eventual cost—RMA logistics, replacement units, service bulletins, customer recovery—blew well past $50,000. The “discount” carrier wasn't cheaper. It was the expensive one with a delay attached.
I don't tell that story to scare you. I tell it because the math usually runs the other way: the better carrier is the lower-cost decision. When I calculate total cost of ownership, I include the unit price, integration cost, certification and compliance burden, rework and field-failure exposure, and the long-term support cost. Most procurement processes stop at the first item. That's the mistake.
When to ignore this advice
Not every project needs an industrial-grade custom carrier. Be honest about the real requirements.
If you're making fewer than 500 units and your deployment environment is controlled—clean rooms, climate-controlled enclosures, minimal vibration—a well-made off-the-shelf single-board computer with solid support may beat a custom carrier on both cost and schedule. Custom carrier design starts to earn its keep at higher volumes, or when you need specific I/O, a specific form factor, or long-term supply guarantees. If you need wide-temperature operation with guaranteed availability, spec the RK3588J industrial variant of the module—it's worth the premium.
If you're already late and need something in the field within weeks, buy a ruggedized commercial board and move on. Don't make your first hardware project a custom carrier on the critical path. Take it from someone who has spent too many hours with a thermal camera.
Honestly, I'm not sure why some vendors consistently nail carrier design while others miss the same basic requirements, project after project. My best guess: it comes down to internal review discipline. The teams that get it right are the ones that review power, thermal, and connector choices before the first board is spun, not after it comes back from the assembly line. Get the carrier right, and the rest of the system tends to get out of its own way.
