This repository provides definitions and examples for different categories of embedded computing boards. The goal is to establish clear terminology for discussing and categorising various types of development boards, single-board computers, and embedded platforms.
Last Updated: 2026-07-02
- Categories
- Classification Guide
- Common Form Factors
- Consumer Electronics Integration
- Contributing
- Notes
A complete computer system implemented on a single board, capable of running a full operating system like Linux or Windows.
Key characteristics:
- Runs full multi-tasking operating systems (requires MMU-equipped processor)
- Has standard display output (HDMI, DisplayPort, DVI) backed by a 3D-capable GPU for desktop graphics
- Generally includes standard PC interfaces (USB, network, etc.)
- May be lower power than traditional PCs
- Complete system on a single board
- Typically credit card sized (85.6 × 53.98 mm) or up to Mini-ITX (170 × 170 mm)
Examples:
- Raspberry Pi 2/3/4/5
- BeagleBone Black
- Radxa Rock 4 SE
- Orange Pi 4 Pro
- Banana Pi BPi-M4
- Libre Computer Le Potato
- Milk-V Mars
- Pine64 ROCK64
- SpacemiT Muse Pi Pro
The core computing system without fixed I/O, designed for integration into custom carrier boards. The terms "Compute Module" and "System-on-Module" are used interchangeably in the industry with no meaningful technical distinction.
Key characteristics:
- Contains the essential computing elements (CPU, RAM, storage, often power management)
- Requires a carrier board for I/O and power delivery
- Highly flexible in terms of final implementation
- Used in both prototyping and production deployments
- Same software capabilities as full SBCs
- May include additional peripherals (WiFi, Bluetooth, etc.)
- Organised by connector/form factor standards
CM4/CM5 Connector Standard (emerging de facto standard):
- CM4 format: 2× 100-pin B2B connectors, 55 × 40 mm
- CM5 format: 55 × 40 mm, largely backward compatible with CM4 carriers for basic I/O. Connector count varies by vendor — Raspberry Pi CM5 keeps the CM4 arrangement of 2× 100-pin B2B connectors, while some third-party CM5-class modules add a third 100-pin connector for extra interfaces (SATA, UFS, DisplayPort, additional PCIe)
- Broad ecosystem support with many compatible carrier boards available
- Examples: Raspberry Pi CM4/CM5, Radxa CM3/CM4/CM5, Orange Pi CM5, Milk-V Mars CM, Pine64 SOQuartz
SODIMM Form Factor (laptop memory edge connector — three generations are in use, and they are not interchangeable):
- 200-pin (DDR2 socket, JEDEC MO-224): Raspberry Pi CM1/CM3/CM3+/CM4S, ~67.6 × 31 mm
- 204-pin (DDR3 socket, JEDEC MO-268, 0.6 mm pitch): Pine64 SOPine, ~67.9 × 31 mm; Toradex Colibri
- 260-pin (DDR4 socket, JEDEC MO-310, 0.5 mm pitch): the largest and most active group — see below
- Pin count, pitch and key-notch position all differ between generations, so a module of one generation cannot be seated in a socket of another
260-pin SODIMM (the "Jetson-style" module footprint):
- Established as a de facto footprint by NVIDIA's Jetson modules, and now adopted by other vendors targeting the same carriers
- Typical module size is 69.6 × 45 mm — taller than the 69.6 × 30 mm JEDEC memory-module outline the socket was designed for
- The socket itself is an off-the-shelf DDR4 SODIMM part; the pinout carried across it is vendor-defined, with no standards body behind it
- Examples: NVIDIA Jetson Nano, Jetson TX2 NX, Jetson Xavier NX, Jetson Orin Nano, Jetson Orin NX, Luckfox Core3576, SpacemiT K3 CoM260 ("CoM260" is SpacemiT's own product name — CoM for Core Module, 260 for the pin count — not a published standard), youyeetoo K1 (Intel N100; 82 × 71 mm module, so it shares the connector but not the 69.6 × 45 mm outline)
Custom Connector Implementations:
- Proprietary or manufacturer-specific connector designs
- Examples: NVIDIA Jetson AGX Xavier/AGX Orin (699-pin board-to-board, 100 × 87 mm) and Jetson TX1/TX2 (400-pin board-to-board, 87 × 50 mm), Toradex Apalis (MXM3 edge connector), Variscite VAR-SOM, Phytec phyCORE, Boundary Devices Nitrogen, Debix SOM, Intel NUC compute elements
A shared connector is a mechanical fact, not an electrical guarantee:
- Supply voltages differ between modules that fit the same socket. Among 260-pin modules alone: Jetson Nano and Jetson Xavier NX are rated 5 V on VDD_IN, Jetson Orin NX/Orin Nano accept 5–20 V, Luckfox Core3576 is specified 5 V, and the youyeetoo K1 takes 12 V
- NVIDIA added a MODULE_ID strap (pin 217) specifically so a carrier can identify a 5 V-only module before applying a higher rail — vendor acknowledgement that the hazard is real
- Pinouts are vendor-defined and can change between generations of one vendor's own range: NVIDIA states Jetson Orin NX/Orin Nano modules are not pin-compatible with Jetson Xavier NX modules
- Raspberry Pi make the same point about their own modules: the CM1/CM3/CM3+/CM4S fit a standard DDR2 SODIMM socket, but "the pinout of the Compute Module is not the same as a DDR2 SODIMM module; they are not electrically compatible"
- Always check the module datasheet against the carrier specification. A module that fits is not necessarily a module that is safe to power.
Similar to an SBC but designed for headless or embedded display operation. The key differentiator is the lack of GPU capabilities for general computing.
Key characteristics:
- Runs full multi-tasking operating systems (requires MMU-equipped processor)
- Either no display output, only embedded display interfaces (MIPI DSI, LVDS), or display output driven by a VPU or 2D hardware engine — but no 3D-capable GPU
- A VPU or hardware 2D engine driving HDMI for HMI applications does not qualify as a GPU for SBC classification
- May have ISP (Image Signal Processor) for camera processing but not general-purpose graphics
- Accessed via SSH, UART, or similar for headless variants
- Often used in embedded applications (IoT gateways, industrial controllers, camera systems)
- May have reduced I/O compared to full SBCs
- Often in gumstick (51 × 21 mm, extending to ~70 mm), cracker (43 × 43 mm to 50 × 50 mm) or stamp (~25 × 25 mm) form factors
Examples:
- Pine64 Ox64
- Milk-V Duo S
- Luckfox Pico Pro
- Luckfox Lyra (RV3506G2 — HDMI via VPU for HMI, no 3D GPU)
- FriendlyElec NanoPi Zero2 (RK3528A — headless, no display output of any kind)
- Many industrial Linux boards (headless variants without GPU)
General-purpose boards built around an MCU (microcontroller) or simple SOC without MMU, designed for embedded applications.
Key characteristics:
- Based on microcontroller or simple system-on-chip (typically without MMU)
- Cannot run full multi-tasking operating systems like Linux
- Typically runs bare-metal firmware, Arduino, MicroPython, or RTOS (FreeRTOS, Zephyr, etc.)
- May have real-time capabilities due to simpler architecture
- Designed for both development and deployment
- Part of a broader development ecosystem
- Lower power consumption than SBC/eSBC counterparts
Examples:
- Raspberry Pi Pico 2
- Arduino boards (Uno, Mega, etc.)
- ESP32 development boards
- Teensy boards
- Adafruit Feather boards
Boards specifically designed to demonstrate and evaluate the capabilities of a particular chip.
Key characteristics:
- Focused on showcasing specific chip features
- Primary purpose is prototyping and evaluation
- Often includes additional debugging features
- May expose all chip pins for testing
- Usually comes with manufacturer-specific tools and documentation
Examples:
- STM32 Nucleo boards
- Nordic nRF52 Development Kit
- Texas Instruments LaunchPad
- ESP32-P4-Function-EV-Board
- Microchip PIC development boards
- NXP evaluation boards
FPGA Development Boards (special case — see Notes):
- BeagleBone BeagleV-Fire
- Xilinx/AMD development boards (Zynq, Artix, Kintex series)
- Intel/Altera DE-series boards
- Lattice iCE40 boards
- Digilent Arty, Basys, Nexys boards
- Terasic DE10-Nano
Specialised boards with dedicated AI/ML acceleration hardware, designed for inference or training workloads.
Key characteristics:
- Contains specialised AI/ML accelerators (NPU, TPU, or AI coprocessors)
- Optimised for neural network inference or training
- May be standalone boards or companion accelerators
- Often includes a host processor (ARM, x86) alongside accelerator
- Typically runs Linux with AI frameworks (TensorFlow, PyTorch, etc.)
- Focus on performance-per-watt for AI workloads
- May support specific model formats (ONNX, TensorFlow Lite, etc.)
Examples:
- Google Coral Dev Board (Edge TPU)
- NVIDIA Jetson Nano/Xavier/Orin (GPU + tensor core AI acceleration)
- BeagleBone AI-64 (TDA4VM with C7x DSP + MMA for AI)
- Canaan CanMV-K230 (RISC-V with KPU neural network accelerator)
This section helps you determine which category a board belongs to, especially for edge cases.
Follow this flowchart to classify a board:
-
Is AI/ML acceleration the primary purpose?
- Yes → AI/ML Accelerator Board
- No → Continue
-
Is it designed primarily to evaluate a specific chip?
- Yes → Development/Evaluation Board (includes FPGA development boards — see Notes)
- No → Continue
-
Does it have an MMU-capable processor?
- No → Embedded System Board (ESB)
- Yes → Continue
-
Does it require a carrier/base board for basic I/O?
- Yes → Compute Module / System-on-Module (CM/SoM)
- No → Continue
-
Does it have a 3D-capable GPU for desktop graphics?
- Yes → Single Board Computer (SBC)
- No → Embedded Single Board Computer (eSBC)
Step-by-step classification of the Milk-V Duo S (SG2000 SoC, Linux-capable, MIPI CSI, no HDMI):
- AI/ML acceleration primary purpose? No — it is a general Linux board. Continue.
- Chip evaluation / demonstration? No — it is a production-ready eSBC. Continue.
- MMU-capable processor? Yes — the SG2000 runs Linux with full MMU. Continue.
- Requires carrier board? No — it is standalone. Continue.
- 3D-capable GPU? No — the SG2000 has no 3D GPU; display is via MIPI DSI only, driven by a 2D engine. → eSBC
Result: Embedded Single Board Computer (eSBC) — cracker form factor (43 × 43 mm).
Key technical aspects that define categories:
| Feature | SBC | CM/SoM | eSBC | ESB | Dev/Eval | AI/ML |
|---|---|---|---|---|---|---|
| MMU | ✓ | ✓ | ✓ | ✗ | Varies | ✓ |
| 3D GPU | ✓ | ✓ | ✗ | ✗ | Varies | Optional |
| Full OS | ✓ | ✓ | ✓ | ✗ | Varies | ✓ |
| Display Out | ✓ | Via carrier | ✗/Embedded | ✗ | Varies | Optional |
| Standalone | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ |
| Production Ready | ✓ | ✓ | ✓ | ✓ | ✗ | ✓ |
Some boards span multiple categories or present classification challenges:
Compute Module/SoM on Carrier Board:
- The module itself is classified as CM/SoM
- However, the complete system (module + carrier) can be described as:
- SBC if the carrier provides GPU-equipped display outputs (e.g., CM4 on official I/O board with HDMI)
- eSBC if the carrier is headless or only has embedded display outputs
- Classification depends on the complete system configuration
Boards with Multiple Purposes:
- NVIDIA Jetson series can be classified as:
- CM/SoM (when referring to the module alone)
- AI/ML Accelerator (when emphasis is on GPU/AI capabilities of the complete system)
- SBC (when the complete system is used as a general-purpose computer)
- Use the context and primary intended use case for classification
Development vs. Production:
- Many ESBs (like Arduino or Pico) are used in both development and production
- DevBoards are primarily for evaluation, not deployment
- The distinction is about intent and features, not actual usage
SBC vs. eSBC Ambiguity:
- The key differentiator is a 3D-capable GPU, not merely the presence of a display connector
- Boards with both HDMI and MIPI DSI outputs → Classify by GPU presence
- If it has a 3D-capable GPU and can render a composited desktop GUI → SBC
- If display output is driven by a VPU, 2D hardware engine, or framebuffer without 3D acceleration → eSBC
- Example: a board with HDMI driven by a VPU for HMI applications (e.g. Luckfox Lyra with RV3506G2) → eSBC, not SBC
SBC + AI/ML Accelerator Combos (e.g., Raspberry Pi 5 AI Kit):
- A standard SBC fitted with an M.2 or HAT-based AI/ML accelerator (e.g., Hailo-8L) becomes a dual-category system
- Primary classification of the base board remains SBC
- The complete system (board + accelerator) may also be described as an AI/ML Accelerator when the inference workload is the primary use case
- Classify by the component being referenced, not the complete system
RTOS Capabilities:
- Modern RTOSes (Zephyr, NuttX) are increasingly capable
- The key differentiator is MMU support, not OS sophistication
- If it lacks MMU → ESB, regardless of RTOS features
Hybrid Boards:
- STM32MP1 boards (MCU + MPU in one chip) → Classify by primary processor
- RP2040 with Linux (via emulation/PIO tricks) → Still ESB (no true MMU)
- Boards with AI accelerators (BeagleBone AI-64, RK3588 boards) → SBC if they have MMU/GPU, can also be classified as AI/ML Accelerator depending on emphasis
Different board categories typically align with certain form factors:
- Credit Card: 85.6 × 53.98 mm (ISO/IEC 7810 ID-1 standard)
- Mini-ITX: 170 × 170 mm (PC industry standard)
- SODIMM 200/204-pin: ~67.6 × 31 mm (DDR2/DDR3 laptop memory module standard)
- SODIMM 260-pin: 69.6 × 45 mm typical for compute modules (DDR4 socket; the JEDEC memory-module outline is 69.6 × 30 mm)
- CM4 / CM5: 55 × 40 mm (Raspberry Pi Compute Module standard)
- Gumstick: 51 × 21 mm base size, may extend to ~70 mm with Ethernet
- Stamp: ~25 × 25 mm (ultra-compact modules)
- Cracker: 43 × 43 mm to 50 × 50 mm (emerging square format; e.g., Milk-V Duo S at 43 × 43 mm, FriendlyElec NanoPi Zero2 at 45 × 45 mm, Luckfox Pico Ultra at 50 × 50 mm (PoE HAT compatibility untested))
The "cracker" form factor emerged around 2019 (based on market research and Google Trends data) with ultra-compact eSBCs in a square format, starting at 43 × 43 mm. This became more standardised in 2024 with the Milk-V Duo S, which also established a common PoE HAT header pin positioning. A larger variant at 50 × 50 mm appeared with the Luckfox Pico Ultra (PoE HAT compatibility untested).
These boards combine the capabilities of traditional eSBCs with an extremely compact footprint, positioning themselves between stamp-sized and traditional gumstick formats. This trend represents a push towards more compact yet fully-featured Linux-capable boards, particularly in the RISC-V space where new chips are enabling smaller form factors without sacrificing capabilities.
While any category can theoretically use any form factor, certain patterns are common:
SBCs typically use:
- Credit card size (85.6 × 53.98 mm) - Most common
- Mini-ITX (170 × 170 mm) - For high-performance variants
- Custom sizes based on PC standards
Compute Modules / System-on-Modules have emerging standards:
- CM4/CM5 (55 × 40 mm) - Emerging de facto standard with broad compatibility
- SODIMM (67.6 × 30 mm) - Established industrial standard
- Custom implementations - Vary widely with proprietary connectors (LGA, castellated edges, B2B)
- Connector standardisation is more important than physical size
- Size depends on feature set and target application
eSBCs trend towards compact formats:
- Gumstick (51 × 21 mm, extending to ~70 mm with Ethernet) - Very common
- Cracker (43 × 43 mm to 50 × 50 mm) - Emerging standard
- Stamp (~25 × 25 mm) - Ultra-compact
- Custom compact sizes
ESBs have the widest variety:
- Arduino form factors (various standardised sizes)
- Feather/PICO formats (~51 × 21 mm)
- Custom sizes based on ecosystem (Teensy, etc.)
- From tiny (< 20 × 20 mm) to large (> 100 × 70 mm)
Dev/Eval Boards are rarely standardised:
- Sized to expose all chip functionality
- Often larger to include debugging features
- May include expansion headers and test points
AI/ML Accelerators follow SBC patterns:
- Standalone: Credit card to Mini-ITX sizes
- Accelerator modules: PCIe cards, M.2, or USB dongles
Many modern consumer electronics devices contain boards that align with our taxonomy categories, even though they're not marketed as such:
- Smart TV mainboards are often SBCs or eSBCs, containing powerful SOCs to handle video processing, apps, and networking
- Set-top boxes and streaming devices typically use eSBC designs
- Digital signage systems often use Compute Modules or eSBCs
- Modern appliances may contain ESBs for control and connectivity
- Gaming consoles use custom-designed SBCs
Consumer electronics boards can often be repurposed. When identifying a board from a consumer device:
- If it runs full Linux (has MMU) and has a 3D-capable GPU with standard display outputs (HDMI/VGA/DP/DVI) → SBC
- If it runs full Linux (has MMU) but only has embedded display interfaces (MIPI DSI/LVDS), or display output is VPU/2D-only without a 3D GPU → eSBC
- If it only runs a basic RTOS or firmware (no MMU) → ESB
- If it has specialised AI/ML hardware and that's the primary function → AI/ML Accelerator
This classification can help in understanding a board's capabilities and potential uses when repurposing hardware.
Example Classifications:
- Smart TV mainboard with HDMI and Android → SBC
- IP camera board with MIPI DSI and Linux → eSBC
- Washing machine control board → ESB
- Security camera with NPU for object detection → AI/ML Accelerator or eSBC (depending on emphasis)
This taxonomy is a living document. We welcome contributions! If you have suggestions for:
- Additional board categories
- New example boards
- Clarifications to definitions
- Edge cases that don't fit well
- Form factor updates
- Classification corrections
Please see our Contributing Guidelines for detailed information on how to contribute. All contributors are expected to abide by our Code of Conduct.
Quick links:
-
Overlapping Categories: Some boards may blur the lines between categories or have features that span multiple categories. See the Edge Cases and Overlapping Categories section for guidance. When in doubt, classify by the primary intended use case.
-
Technical Differentiators: The key technical factors for classification are:
- MMU presence (can it run full Linux?)
- GPU capabilities (can it render desktop GUI?)
- Standalone vs. modular (does it need a carrier board?)
- Primary purpose (evaluation, production, development, acceleration?)
-
Form Factor Flexibility: The presence or absence of specific features (like wireless connectivity or storage) is not a primary factor in categorisation, as these can appear across all categories. Form factors are guidelines, not strict rules.
-
Compute Modules/SoMs and Carrier Boards: CM/SoM modules bridge multiple categories depending on their carrier board implementation - the complete system can function as an SBC, eSBC, or specialised device based on the carrier board design.
-
Living Document: This taxonomy evolves with the industry. Check the CHANGELOG.md for updates and revisions to definitions.
-
FPGA Boards: Field-Programmable Gate Arrays are reconfigurable hardware — the silicon has no fixed function until programmed with an HDL design. FPGA-based development boards are therefore classified at the hardware level as a specialised type of Development/Evaluation Board. The functional characteristics of any specific FPGA design loaded onto such a board (which may emulate an SBC, eSBC, ESB, or any other category) are outside the scope of this taxonomy. Classify the board, not the loaded design.
-
Licence: This taxonomy is available under the MIT License, making it free to use, modify, and distribute.