Getting started

The first decision in this build is to select the engine that powers your AI: the GPU. Whether you’re giving new life to an used card or upgrading to something fresh, there are a few simple things to keep in mind:

We recommend AMD Radeon RX 6000 and 7000 series GPUs for their exceptional price-to-performance ratio and dependable open-source Linux drivers. An 8GB VRAM card is great for getting started, but 16GB or more lets you explore larger, more ambitious models.

You are of course encouraged to try other GPUs from the Raspberry Pi PCIe Database, or share your results for cards not listed, to help grow the community’s knowledge.

Finally, keep the GPU’s idle/maximum power and its physical size in mind, as these will influence your choice of PSU and case.

With your GPU selected, it’s time to house your build in a case that suits both style and substance. If you already have an ATX case, it can easily accommodate a Mini-ITX board. However, if you’re eyeing a new Mini-ITX case, keep these tips in mind:

• GPU fit: Check that your case can handle the size of your GPU, both in length and height. Many modern GPUs are over 25cm long and take up two or more slots, which may exceed the capacity of some Mini-ITX cases.

• PSU type: Determine whether the case requires an ATX or the more compact SFX power supply.

• Cooling options: Look for the number of case fans included or supported. The Sentinel Core can power up to two case fans, so you can keep your setup cool and efficient.

The last piece of the puzzle is your power supply unit (PSU), that suits your GPU and case:

• Case compatibility: Check whether your case requires an ATX or SFX PSU.

• Wattage: Make sure the PSU's 12V output comfortably exceeds your GPU’s maximum power draw. (The Pi’s 5V usage is negligible in comparison.)

• Connections: Confirm there are enough 6+2 PCIe power connectors for your GPU—some SFX models may fall short.

Other features you might want to consider:

• Detachable cables: A modular PSU allows for cleaner cable management.

• Efficiency ratings: 80 Plus or higher certification often means better power efficiency.

• Low idle power: An efficient PSU reduces idle power draw, saving energy when the system isn’t under load.

• Silent operation: Look for 0 RPM fan control to keep things whisper-quiet.

Once your Sentinel Core is up and running, you can customize and enhance it with a few optional extras to suit your needs:
Storage: Planning to store large media libraries or AI models? Consider adding a USB 3 hard drive.

Connectivity: Expand your setup’s capabilities with Raspberry Pi HATs or a ZigBee dongle for seamless smart home integration.

Silent fans: Keep things cool and quiet by upgrading your case fans to ultra-silent models for a more peaceful environment.

Start by going to settings -> Add-ons -> Add-on store.

From Official add-ons install Piper and Samba share. Optionally also install openWakeWord and Assist Microphone.

Sentinel Core additionally comes preconfigured withe the Home Assistant LLM add-on repository, from which we will install Ramalama and Ramalama Whisper for GPU accelerated Llama.cpp and Whisper.cpp. Additionally, install Sanctuary Updater to receive OS updates from our GPU-enabled fork of Home Assistant.

For most of these the default options are fine, but in particular you will want to set a Samba password and experiment with choosing different LLM models in the Ramalama configuration.

For all these add-ons, make sure that "Start on boot" and "Watchdog" are enabled.

There currently isn't an official Llama.cpp integration, so we will use Extended OpenAI Conversation. To install that, download the ZIP file from the above link and extract it, connect to the Samba share and open the "config" folder. Drag the custom_components folder from the ZIP to the Samba share config folder. Now restart Home Assistant in Settings -> System and select "Restart Home Assistant" or "Reboot system" from the power menu in the top right.

Now head over to Settings -> Devices & Services.

All the installed voice add-ons should be auto-detected by Wyoming. If not, manually add Wyoming and click on "add service", then enter "localhost" and the port from the settings of each voice add-on.

Now click "Add Integration" and search for Extended OpenAI Conversation and add it. For the Base URL enter "http://localhost:10202" where 10202 is the port configured in the Ramalama addon. For the API key enter anything you want, and then proceed.

You will probably want to tweak the prompt and make sure to enable "Control Home Assistant".

This is where it all comes together. In Settings -> Voice Assistants select the existing "Home Assistant" and configure the following items:

• Conversation Agent: the configuration added in Extended OpenAI Converation

• Speech-to-text: whisper.cpp

• Text-to-speech: piper

Then head over to the "Expose" app and select the entities you want you AI to be able to see and control.

That's it! You can now chat with you AI from the assist menu in the top right of the dashboard. For using your microphone it is best to use the mobile app.

From here you can explore different models, different voices, and pair microphones like the Home Assistant Voice, FutureProofHomes Satelite1, or Wyoming Satelite.