Hello everyone,
We are developing a custom ARM-based SBC.
Instead of using an off-the-shelf SBC, we plan to select a suitable bare SoC and design the complete supporting circuitry around it, including memory, storage, power, interfaces and other required circuits.

Our main requirements
Quad-core ARM Cortex-A, Minimum 1.4 GHz, 64-bit, 4 GB DDR3 RAM, 32 GB eMMC, eMMC boot required, Minimum 11 GPIO, 1 × Ethernet, 1 × HDMI, 2 × USB 2.0, 2 × USB 3.0, Minimum 2 × UART, Minimum 2 × I²C, Minimum 2 × SPI, Minimum 2 × CAN, RTC, Linux support, Linux BSP, U-Boot, Device Tree, Linux kernel source, SDK and cross compiler, Ubuntu 18.04.6 LTS support

We need expert suggestions on

1. Which ARM SoC would you recommend for these requirements, and why?
2. What important parameters should we consider while selecting the SoC?
3. Can all these interfaces realistically be implemented simultaneously on a custom board?
4. What external ICs/circuits will be required around the SoC?
5. What should we consider for DDR3, eMMC and PMIC design?
6. What are the major challenges with USB 3.0, HDMI and Ethernet?
7. What PCB/SI/PI considerations should we consider from the beginning?
8. What Linux BSP/U-Boot/Device Tree support should we verify before selecting the SoC?
9. What are the major risks for a first custom ARM SoC board?
10. Can you recommend any experienced engineers or companies who have designed and brought up similar ARM Linux boards?

We are particularly interested in responses from engineers who have personally designed and brought up custom ARM Cortex-A/application-processor boards with DDR, eMMC and Linux.

We are currently at the SoC selection and architecture stage, so we would like to get expert feedback before starting schematic and PCB design.

Thank you.

 

Your requirements are feasible, but the SoC choice should follow a feasibility matrix rather than start with core count and clock speed. From a manufacturing-review perspective, I would freeze these items before selecting a part:

1. Build a pin-mux table for every required interface, including the exact USB 3.0, CAN, SPI, UART and I2C instances. On many application processors the requested interfaces cannot all be routed simultaneously because pins and high-speed lanes are shared.
2. Obtain the vendor reference schematic, DDR layout rules, PMIC/power-sequencing guide, clocking requirements and a known-good BSP for the exact package and memory topology. If any of these are unavailable under the vendor's normal support channel, treat that as a selection risk.
3. Treat 4 GB DDR3, USB 3.0, HDMI and Ethernet as board-architecture items, not late PCB tasks. Confirm stack-up, controlled impedance, reference-plane continuity, return paths, length/skew rules, placement zones and escape routing before committing to the SoC package. The required pin count may drive HDI, layer count and test access.
4. Verify boot, recovery and production-test paths early: eMMC boot straps, serial console, secure/recovery mode, programming access, rail test points and a fixture strategy. These are much harder to add after the first layout.
5. Be cautious with Ubuntu 18.04.6 as a hard requirement: confirm the exact kernel, bootloader, device-tree and userspace support for the chosen SoC, not just that Linux is listed on a marketing page. Also verify the vendor's lifecycle, BSP maintenance commitment and availability of a reference board close to your intended design.

I would create a short weighted matrix covering interface pin-mux compatibility, DDR reference design availability, BSP maturity, lifecycle/sourcing, package escape/PCB cost, power architecture and lab bring-up support. That normally narrows the candidate list more reliably than choosing a CPU first. For a first custom application-processor board, an experienced SI/PI and Linux bring-up review before layout is usually money well spent.

Disclosure: I participate here from PCBA Partner, an electronics-manufacturing team. This is a general manufacturability and handoff perspective, not a claim of having validated a particular SoC for this project.

 

In the early Raspberry Pie days, we wanted a "computer board" and the wish list described the Pie. I lobbied for the Pie but was turned down because it was "single sourced", too much money, and did not look professional. In 6 months, we made a board that was 90% of a Pie for 2.5x the price, because of low volume. The board looked much like this one. I don't like re-inventing the wheel.