ArmSoM
← All stories

Choosing an SBC with Onboard AI: What Raspberry Pi Users Should Check

An onboard NPU can remove the need for a separate accelerator, provided the model works with its tools. For ArmSoM’s Rockchip platforms, model conversion and runtime support belong in the board-selection process, before a TOPS figure influences the purchase.

ArmSoM Sige5

ArmSoM Sige5

Try the model before committing the hardware

Check the model’s operators, input shapes and supported conversion path. Run representative inputs and inspect the outputs against the original model. A successful conversion is useful, but accuracy and complete response time determine whether the application can use it. Keep CPU preprocessing and postprocessing in the timing measurements.

Count every stage of inference

Measure capture, resize, normalization, data transfer, inference and result handling. An accelerator can shorten one stage while leaving a CPU-bound preprocessing stage unchanged. Include the memory used by buffers and simultaneous services. Compare an add-on accelerator with an integrated NPU at the same model precision and output quality, and retain the CPU implementation as a correctness reference.

ArmSoM Sige7s

ArmSoM Sige7s

Match the choice to your requirements

Requirement or situation

What to evaluate

Decision detail

Unsupported operator

Review toolkit support

A CPU fallback may change latency.

Quantization error

Evaluate representative samples

Do not use only an easy demo image.

High inference speed, slow application

Measure non-NPU stages

End-to-end response time is the product metric.

Example design

Qualify the model before selecting the board

Take a small representative dataset and save the original model’s outputs. Convert the model through the intended Rockchip toolchain and compare the outputs against the application’s acceptable error. If conversion requires changing preprocessing or replacing an operator, record it. Only then does an inference timing comparison become useful. This avoids choosing an NPU board on a headline capability and discovering later that the application depends on an unsupported model path.

Hardware and documentation

Sige7s · RK3588S

RK3588S, 6 TOPS NPU, Gigabit Ethernet and removable eMMC. Dimensions: 85 × 56 × 17 mm. Disconnect power before replacing eMMC.

Hardware documentation · Product details

Sige5 · RK3576

RK3576, four Cortex-A72 and four Cortex-A53 cores, 6 TOPS NPU, dual Gigabit Ethernet and M.2 PCIe 2.1 x1.

Hardware documentation · Product details

Frequently asked questions

Are all 6 TOPS boards equally fast?

No. Software, model structure, memory traffic and pipeline design affect real results.

Product information and support

Explore Sige7s

Contact ArmSoM

Sources

Written by ArmSoM using the documentation below. Example designs are proposals, not measured customer installations.

Related reading