# nsx-npu

<!-- nsx:facts -->
Atomiq110 glue for the Arm Ethos-U85 NPU: power-domain sequencing, IRQ wiring, and performance-mode selection on top of the pristine upstream driver (nsx-ethos-u-driver) so TFLM runtimes can dispatch Vela-compiled command streams to the NPU.

- Type: Backend specific
- Version: 5.2.26
- Project: nsx-ambiq-sdk
- Revision: v5.2.26
- Source: https://github.com/AmbiqAI/nsx-ambiq-sdk/tree/v5.2.26
- Capabilities: npu, ethos_u, inference_acceleration
<!-- /nsx:facts -->

## Add it to an app

```bash
nsx module add nsx-npu
```

## Capabilities

- npu
- ethos_u
- inference_acceleration

## Use cases

- Ethos-U85 NPU inference offload
- Vela-compiled TFLite model execution
- NPU power and performance mode control

## Requires

- [`nsx-core`](https://ambiqai.github.io/neuralspotx/modules/nsx-core/)
- [`nsx-soc-hal`](https://ambiqai.github.io/neuralspotx/modules/nsx-soc-hal/)
- [`nsx-ethos-u-driver`](https://ambiqai.github.io/neuralspotx/modules/nsx-ethos-u-driver/)

## Compatibility

:::note[How compatibility is declared]
The boards, SoCs and toolchains listed here are the ones the module manifest declares.
That is the module author's statement of intent, not evidence from hardware. Confirm on your
own target before you depend on it.
:::

| Kind | Declared |
| --- | --- |
| Boards | atomiq110_fpga_turbo |
| SoCs | atomiq110 |
| Toolchains | arm-none-eabi-gcc, armclang, atfe |

## Agent keywords

`npu`, `ethos`, `ethos-u`, `ethos-u85`, `vela`, `accelerator`

## Source

| Field | Value |
| --- | --- |
| Type | Backend specific |
| Version | 5.2.26 |
| Project | nsx-ambiq-sdk |
| Revision | `v5.2.26` |
| Manifest | `modules/nsx-npu/nsx-module.yaml` |
| Repository | [nsx-ambiq-sdk](https://github.com/AmbiqAI/nsx-ambiq-sdk/tree/v5.2.26) |

Back to the [module catalog](https://ambiqai.github.io/neuralspotx/modules/catalog/).