---
title: "Introduction"
description: "The Alp SDK is the unification software layer for Alp Lab edge AI modules built on the E1M open-standard form factor. It provides a single C/C++ API — `` — that works across every E1M SoM variant by wrapping each vendor's SDK on top of ARM CMSIS."
source: https://docs.alplab.ai/sdk/introduction
---

# Alp SDK

The **Alp SDK** is the unification software layer for Alp Lab edge AI modules built on the [E1M open-standard form factor](https://docs.alplab.ai/e1m/standard). It provides a single C/C++ API — `<alp/...>` — that works across every E1M SoM variant by wrapping each vendor's SDK on top of ARM CMSIS.

> Write once, run on any E1M module.

## Design principles

- **Vendor-independent** — one API across Alif Ensemble, Renesas RZ/V2N, and any future conformant silicon.
- **CMSIS-compatible** — sits on top of ARM CMSIS so existing CMSIS code and DSP libraries integrate without friction.
- **Heterogeneous-OS** — Zephyr, Yocto Linux, and bare-metal are peers, declared **per-core** in one `board.yaml`. One SoM can run all three at once.
- **Hand-written friendly** — the SDK supports hand-written firmware as a first-class consumer, not just `alp-studio` codegen.
- **Declarative configuration** — a single `board.yaml` at your app root replaces scattered Kconfig fragments, CMake `-D` flags, and `local.conf` edits.

## Supported targets

Per-core OS choice (heterogeneous orchestration):

| Module family                       | A-class cores             | M-class cores              | Heterogeneous? |
|-------------------------------------|---------------------------|----------------------------|----------------|
| [E1M-AEN E3/E4](https://docs.alplab.ai/e1m/e1m-aen)       | —                         | M55-HP + M55-HE (Zephyr)   | No — RTOS-only |
| [E1M-AEN E5..E8](https://docs.alplab.ai/e1m/e1m-aen)      | A32 cluster (Yocto)¹      | M55-HP + M55-HE (Zephyr)   | Yes            |
| [E1M-X V2N](https://docs.alplab.ai/e1m/e1m-x-v2n)         | A55 cluster (Yocto)       | M33-SM (Zephyr)            | Yes            |
| [E1M-X V2N-M1](https://docs.alplab.ai/e1m/e1m-x-v2n-m1)   | A55 cluster (Yocto)       | M33-SM (Zephyr) + DEEPX    | Yes            |

¹ The AEN A32 Yocto MACHINEs are **not buildable yet**. `meta-alp-sdk` carries `e1m-aen701-a32` and `e1m-aen801-a32` only as placeholders, because there is no Scarthgap-compatible Alif Yocto base for them to build on, and `e1m-aen501-a32`, `e1m-aen601-a32` and `e1m-aen803-a32` have no MACHINE conf at all. For any of these the planner emits no `bitbake` command: the slice carries `command: null` and a `yocto-machine-unbuildable` [build-plan](https://docs.alplab.ai/sdk/heterogeneous-builds#build-plan) warning. The M55 Zephyr cores on the same SoMs build normally.

alp-sdk v0.17.0 removed the E1M-NX9101 (NXP i.MX 93) module, which was never produced. Its SoM preset, SoC spec, Yocto MACHINE, vendor directory and the `rpmsg-imx93` example are gone, and a `board.yaml` that names it no longer plans.

## Stack

```mermaid
flowchart TB
  classDef layer fill:#0f1424,stroke:#475569,color:#e2e8f0
  classDef sdk fill:#1e3a8a,stroke:#2f57d6,color:#ffffff
  classDef hw fill:#161b2e,stroke:#7dd3fc,color:#e2e8f0

  APP["Application code"]:::layer
  SDK["Alp SDK &lt;alp/*.h&gt;<br/>Peripherals · Audio · Camera · Inference · IoT · BLE · Security<br/>Multi-proc IPC · HW info · 80 chip drivers"]:::sdk
  OS["OS backend (Zephyr · Yocto · bare-metal)"]:::layer
  HAL["Vendor SDK / HAL (Alif · Renesas · DEEPX · TI)"]:::layer
  HW["E1M / E1M-X SoM hardware"]:::hw

  APP --> SDK --> OS --> HAL --> HW
```

## SDK components

### Public API — `<alp/*.h>`

| Group | Headers |
|---|---|
| Peripherals | `peripheral.h` (GPIO/I²C/SPI/UART), `pwm.h`, `adc.h`, `counter.h`, `i2s.h`, `can.h`, `rtc.h`, `wdt.h`, `usb.h` |
| Audio / camera / display | `audio.h` (PDM in + I²S out), `camera.h`, `gui.h` (LVGL), `display.h` |
| Connectivity & security | `iot.h` (Wi-Fi/MQTT), `ble.h` (BLE 5.4), `security.h` (MbedTLS PSA Crypto) |
| Inference dispatcher | `inference.h` — silicon-determined dispatcher set (TFLM CPU always; Ethos-U / DRP-AI3 / DEEPX DX-M1 per SoM caps); per-handle backend pick at runtime via `alp_inference_open()` |
| Cross-core IPC | `rpc.h` — framed RPC over OpenAMP/RPMsg; `mproc.h` — raw mailbox + shared memory |
| Hardware info | `hw_info.h` — 128-byte EEPROM manifest + BOARD_ID ADC |
| Signal / DSP | `dsp.h` — FIR/IIR/FFT chains over CMSIS-DSP |
| Power / storage | `power.h`, `storage.h`, `tmu.h`, `gpu2d.h` |
| Capability tables | `soc_caps.h` (generated), `e1m_pinout.h` (portable instance IDs) |
| Generated headers | `system_ipc.h` — endpoint IDs + addresses + mailbox channels emitted from `board.yaml`'s `ipc:` block |
| Chip drivers | 80 under `<alp/chips/...>` — opt-in via `board.yaml` |

See the [API overview](https://docs.alplab.ai/sdk/api/overview).

### Dev tooling

- **`board.yaml` project config** — single source of truth for SoM SKU, board (inline `name:` + `populated:` + `e1m_routes:` or `preset:` shortcut), E1M-pad → board-side macro assignments, per-core app + libraries + peripherals + inference arena + memory + power, IoT toggles, IPC carve-outs, diagnostics + per-module log levels, MCUboot, OTA (mender / hawkbit / mcumgr), filesystem partitions, PSA Crypto + TF-M. Per-core `os:` is optional (topology default). See the [reference](https://docs.alplab.ai/sdk/board-yaml).
- **`scripts/alp_orchestrate/`** — fans out `board.yaml` into per-core build slices; emits `system-manifest.yaml`, `alp/system_ipc.h`, `dts-reservations.dtsi`, `dts-partitions.dtsi`, `sysbuild/tfm/tfm.conf`, per-slice Kconfig fragments / Yocto `local.conf` snippets / CMake `-D` flags.
- **The [`tan` CLI](https://docs.alplab.ai/sdk/tools/cli)** — the SDK's sole user-facing command: `tan build` (fan-out), `tan image` (assemble bundle), `tan flash` (boot-order-aware programming), `tan clean` (idempotent teardown), `tan run` (headless `native_sim`, or build+flash+console with `--flash`). The five `west alp-{build,image,flash,clean,size}` extension commands it replaced were retired in v0.12.0 (ADR-0020 Phase 4), and `alp-renode`/`tan renode` went with Renode itself in v0.16.0 (ADR 0022 Amendment 2). Four extension commands survive: `west alp-lock`, `alp-migrate`, `alp-quality`, `alp-emit`. `tan` versions separately from the SDK; **tan 0.7.0 requires alp-sdk v0.17.0 or newer** (against an older SDK a Zephyr build fails Kconfig configure on the undefined `ALP_SDK_SOM_HW_REV`).
- **`scripts/validate_board_yaml.py`** — customer-side linter; runs JSON Schema + cross-field validator (5 rules + 2 warnings).
- **`scripts/program_eeprom.py`** — packs `board.yaml` + serial + mfg date into the 128-byte EEPROM manifest for production-test programming.
- **[VS Code extension](https://docs.alplab.ai/sdk/tools/vscode)** — schema-aware editing, GUI configurator, west wrappers, per-OS bootstrap. Ships from its own repo at [alplabai/alp-sdk-vscode](https://github.com/alplabai/alp-sdk-vscode) since 2026-05-12.

### Inference back-ends

The `<alp/inference.h>` dispatcher routes TFLite Micro models to silicon-specific NPU back-ends:

- **Arm Ethos-U** — Alif Ensemble (AEN family — every E3..E8 SKU carries U55; E4/E6/E8 additionally carry Transformer-capable U85)
- **Renesas DRP-AI3** — RZ/V2N (V2N family)
- **DEEPX DX-M1** — V2N-M1 (V2M family); ONNX → DXNN compiler, model-family agnostic
- **CPU** — reference-kernel fallback on any target

### Supported on-device model families

| Class                 | Examples                              | Backends                                |
|-----------------------|---------------------------------------|-----------------------------------------|
| Classification        | MobileNet, EfficientNet-Lite          | Ethos-U, DRP-AI3, DEEPX, CPU            |
| Detection             | YOLOv5, YOLOv8, SSD-MobileNet         | **YOLO v5/v8 on DEEPX + DRP-AI3**; SSD on Ethos-U + CPU |
| Segmentation          | DeepLabV3, U-Net                      | Ethos-U, DRP-AI3, DEEPX                 |
| Keyword spotting      | DS-CNN, Conformer-Tiny                | Ethos-U (HE core)                       |
| Pose estimation       | MoveNet, BlazePose                    | DRP-AI3, DEEPX                          |
| Anomaly / forecasting | 1D-CNN, TinyTransformer (U85 only)    | Ethos-U55 (1D-CNN); Ethos-U85 (Transformer) |

Offline training (off-device) stays in TensorFlow / PyTorch.

## Two consumer paths

The SDK supports both equally — pick whichever fits.

### Standalone / hand-written

Write Zephyr, Yocto, or bare-metal app code directly against `<alp/...>` headers. Pick instance IDs from `<alp/e1m_pinout.h>` (`ALP_E1M_I2C0`, `ALP_E1M_PWM3`, …). Your app stays portable across every E1M-conformant SoM.

### alp-studio codegen

The [Studio](https://studio.alplab.ai) reads block manifests, runs the pin allocator against the active SoM, and emits C that calls the same `<alp/...>` API. Pin-allocation correctness comes for free.

The standalone path is **not** a Studio escape hatch — it's a first-class consumer.

## Where next

- [Install the SDK](https://docs.alplab.ai/sdk/installation)
- [Quick start](https://docs.alplab.ai/sdk/quick-start) — build and flash the GPIO example
- [`board.yaml` reference](https://docs.alplab.ai/sdk/board-yaml) — flat schema + the declarative blocks
- [Heterogeneous builds](https://docs.alplab.ai/sdk/heterogeneous-builds) — dual-OS project walkthrough
- [Firmware quickstart](https://docs.alplab.ai/sdk/firmware-quickstart) — per-SoM patterns
- [Architecture](https://docs.alplab.ai/sdk/architecture) — wrapper layers, codegen split, orchestrator
- [API reference](https://docs.alplab.ai/sdk/api/overview)
