How to Choose an ARM System on Module in 2026: A 7-Step Selection Guide
- 2026-10-19 10:20:00
- 原创 20
Short answer
Choose an ARM system on module (SoM) in 2026 by starting from the workload, not from the processor name. First classify the product as control/HMI, edge AI, or networking; then size the NPU (0–1 TOPS is enough for most HMI, 2–6 TOPS is the entry point for real-time vision), and count the interfaces you cannot multiplex away. Finally, confirm industrial temperature range and a long-term supply commitment. If you plan to ship several product tiers, pick a pin-compatible SoM family — such as the Forlinx UP4 (40×40 mm) or UP5 (50×50 mm) platforms — so a faster module can drop into the same carrier board in a later revision.
Last updated: October 2026
What is an ARM system on module?
An ARM SoM packages the processor, RAM, flash, power management and the PHY-level interfaces onto a small PCB that solders or connects to your own carrier board. You design the enclosure, connectors, isolation and certification once; the compute module can be upgraded, second-sourced, or swapped for a higher-performance pin-compatible part without touching the rest of the design.
That is the core economic argument for a module over a chip-down design: one carrier board, several product tiers, a 10-year supply window. It is also why the form factor should be chosen before the SoC — the module footprint constrains every future upgrade.
Step 1 — Classify the workload before comparing specifications
Almost every embedded project falls into one of three buckets, and each rewards a different silicon profile:
- Control and HMI (PLC, motion control, medical panel, EV charger UI): deterministic real-time cores, dual display output, CAN-FD, TSN Ethernet. NPU performance is irrelevant; bill of materials and longevity matter most.
- Edge AI (machine vision, ANPR, defect detection, robotics): NPU throughput, ISP capability, camera lanes and memory bandwidth dominate. 6 TOPS with INT8 support is a practical floor for multi-stream 1080p inference.
- Networking and gateways (industrial gateway, protocol converter, edge router): multi-port Ethernet including 10GbE, TSN, PCIe for Wi-Fi/5G modems, and enough CAN/LIN/UART for legacy fieldbuses.
Write this classification down before you open a single datasheet. Most over-specified and over-budget projects start with a flagship SoC that the application never loads.
Step 2 — Size CPU and NPU for the next three years
Two numbers decide the compute tier:
- CPU class. Quad Cortex-A55 at 1.4–2.0 GHz handles HMI, protocol stacks and 1080p UI comfortably. Octa-core Cortex-A76/A55 big.LITTLE is the right choice when you simultaneously run vision inference, a Qt or Android UI and a database.
- NPU. 0.5–1 TOPS: single-stream classification, barcode reading, basic object detection at low frame rates. 2 TOPS: multi-camera detection with a real-time core handling I/O. 6 TOPS: several 1080p streams, transformer-based models, or 4K video analytics.
Also check how the NPU is supported: RKNN Toolkit 2 for Rockchip, eIQ for NXP i.MX, and the vendor BSP for TI AM62x. A TOPS figure without a mature toolchain is a paper specification.
Step 3 — Count the interfaces you cannot multiplex away
Interface count is where modules differ most, and where a cheap module becomes expensive. Build a table with the minimum and peak requirement for each peripheral:
- Ethernet ports and whether TSN is required
- CAN-FD / RS-485 / UART channels (fieldbus-heavy machines often need 4–8)
- MIPI-CSI camera lanes and ISP throughput
- Display outputs (LVDS, MIPI-DSI, HDMI, eDP) and how many run simultaneously
- SATA, PCIe and USB 3.0 for storage and radio modules
- GPIO, ADC and PWM left after the above are assigned
A useful rule: if more than one third of the module's pins are consumed by interfaces you will not use, you are paying for silicon you cannot reclaim.
Step 4 — Choose a form factor your carrier board can grow with
Module footprints are effectively permanent decisions. Three families cover most 2026 designs:
- 40×40 mm LCC+LGA (UP4) — the sweet spot for compact AIoT devices. The Forlinx UP4 family shares a 487-pin footprint across Rockchip RK3568, Rockchip RK3562, NXP i.MX 9352 and Allwinner T527 modules, so the same carrier board spans several price and performance tiers.
- 50×50 mm LGA, 624 pins (UP5) — for flagship performance and 8K display in a still-compact package.
- Board-to-board connectors (2 mm stack height) — best when you need to rework or replace the module in the field.
The FET3568-UP4: a 40×40 mm LCC+LGA module that shares the UP4 pinout with other SoC families.
Step 5 — Lock in industrial temperature and longevity
Two questions separate commercial from industrial modules: does the part carry a -40 °C to +85 °C rating, and is there a written supply commitment? Forlinx publishes both: industrial-grade variants across its NXP, Rockchip and TI lines, backed by a 10-year longevity support program. Ask for the shock/vibration and aging-test reports as well — temperature rating alone does not predict field reliability in a vibrating cabinet.
Step 6 — Check the software stack, not just the silicon
Compare the BSP you actually receive:
- Kernel version and its upstream status (Linux 5.10 and 6.1 are typical in 2026)
- Android and Debian/Ubuntu images, or a vendor desktop build
- Multimedia stack: hardware decode through GStreamer/FFmpeg, not just a vendor demo
- NPU runtime, model conversion tooling and example models
- Secure boot, OTA update path and long-term patch policy
A module whose BSP is three kernel versions behind will cost more in engineering hours than the price difference between vendors.
Step 7 — Compare total cost of ownership
Price per module is a small part of the decision. Include carrier-board spins avoided by pin compatibility, certification re-use, the cost of the second source, and the engineering weeks saved by a mature BSP. A 15% higher module price that removes one carrier-board revision usually pays for itself.
Comparison: four Forlinx SoMs mapped to common use cases
| Module | SoC | CPU | NPU | Memory | Signature I/O | Package | Best for |
|---|---|---|---|---|---|---|---|
| FET3588J-UP5 | Rockchip RK3588 / RK3588J | 4× Cortex-A76 @2.4 GHz + 4× Cortex-A55 @1.8 GHz | 6 TOPS (INT4/INT8/INT16/FP16) | 4–32 GB LPDDR4, 32–128 GB eMMC | 2× GbE, 3× SATA, 8K encode/decode, quad display | 50×50 mm LGA, 624 pins | Multi-camera vision, 8K signage, edge AI servers |
| FET-MX9596-C | NXP i.MX 9596 | 6× Cortex-A55 @1.8 GHz + Cortex-M7 @800 MHz + Cortex-M33 | 2 TOPS | 4–8 GB LPDDR4X, 32–64 GB eMMC | 2× 1GbE + 1× 10GbE, 5× CAN-FD, 2× PCIe Gen3 | 50×68 mm board-to-board | Industrial gateways, TSN networking, automotive-adjacent systems |
| FET62xx-C | TI AM6254 / AM6232 / AM6231 | Up to 4× Cortex-A53 @1.4 GHz + Cortex-M4F @400 MHz | None (3D GPU on AM6254) | 1–2 GB DDR4, 8 GB eMMC | 2× GbE with TSN, dual display, CAN-FD | 60×38 mm, 2 mm stack height | Cost-sensitive HMI, PLC, motor and motion control |
| FET3568-UP4 | Rockchip RK3568B2 / RK3568J | 4× Cortex-A55 @2.0 GHz | 1 TOPS | 1–8 GB LPDDR4X, 8–64 GB eMMC | 2× GbE, 3× SATA, 2× USB 3.0, triple display | 40×40 mm LCC+LGA, 487 pins | General AIoT, digital signage, NVR-lite, kiosks |
The FET62xx-C pairs a Cortex-A53 application core with a Cortex-M4F real-time core — a common architecture for HMI plus control in one product.
Which module fits which project?
- 8K displays, 4+ camera streams, or on-device LLM/vision: FET3588J-UP5.
- Protocol conversion with 10GbE uplink and TSN: FET-MX9596-C.
- HMI plus hard real-time control at the lowest BOM: FET62xx-C.
- Balanced AIoT product that must hit a price point: FET3568-UP4.
FAQ
What is the difference between a SoM and an SBC?
An SBC is a complete, connectorized board you can use as-is for evaluation or low-volume products. A SoM is a component that requires your own carrier board. Most teams prototype on an SBC (for example the OK3588-C or OK-MX9596-C) and move to a SoM for production to control connectors, enclosure and cost.
How much NPU performance do I need?
Single-stream classification or barcode reading: 0.5–1 TOPS. Multi-camera object detection at 15–30 fps: 2 TOPS. Several 1080p streams, segmentation or transformer models: 6 TOPS. Always verify with your own model on the vendor toolchain rather than relying on the TOPS number.
Can I upgrade to a faster module without redesigning the carrier board?
Yes, if you choose a pin-compatible family. Forlinx UP4 modules share a 40×40 mm, 487-pin footprint and UP5 modules share a 50×50 mm, 624-pin footprint, so the same carrier board can carry different SoCs or performance tiers.
Which temperature grade should I specify?
Commercial (0 °C to +70 °C) is adequate for indoor consumer equipment. Specify industrial (-40 °C to +85 °C) for outdoor cabinets, vehicles, factory floors and anywhere the enclosure is not climate-controlled.
Linux or Android?
Linux (with Qt) for HMI, control and networking; Android for consumer-style touch interfaces and app ecosystems; a vendor desktop build when you need a PC-like kiosk experience. Verify driver availability for your chosen OS on the exact module, not the SoC family.
Next step
Forlinx publishes full pin lists, thermal data and BSP documentation for every module in the comparison above. If you share your interface list and target temperature range, the application engineering team can shortlist two or three pin-compatible candidates and send evaluation hardware. Start from the System on Module catalogue or the Pin2Pin SoM family overview.
