Hi everyone,

I picked up an Intel Arc B70 for Ollama and others.

Since I haven’t found anything about this in the forum yet and it wasn’t exactly plug-and-play, here’s how I got it working. I’m no expert and got a lot of help from gemin; it took me about two days to get it up and running.

1. Check if your motherboard even supports the Intel Arc B70—for me, it only worked after a BIOS update.

2. Configure the BIOS:

- Re-Size BAR Support: Enabled

- Above 4G Decoding: Enabled

- CMS: Disabled

- Boot Mode: UEFI

- Secure Boot: Disabled

I’m not 100% sure if the last three settings have to be this way, but I have them set this way in any case

3. Unraid: I’m using Unraid 7.3.0-Beta.2 because the current version apparently doesn’t yet include the new Linux kernel that supports the card

4. The Intel-GPU-TOP plugin does detect the card, but not yet.

5. Ollama Docker: The official ones in the Unraid App Store didn’t work for me. Maybe I did something wrong, but I couldn’t get them to run—I always got error 2 in Ollama. I’m currently using uberchuckie/ollama-intel-gpu from Docker Hub.

Repository: uberchuckie/ollama-intel-gpu

Settings to add:

Type: Variable Name: ZES_ENABLE_SYSMAN Key: ZES_ENABLE_SYSMAN Value: 1

Type: Variable Name: SYSCL_CACHE_PERSISTENT Key: SYSCL_CACHE_PERSISTENT Value: 1

Type: Port Name: Ollama Port Container Port:11434 Host Port:11434 Connection Type: TCP

Type: Device Name: Intel GPU Value: /dev/dri

Type: Path Name: Ollama Models Container Path:/root/.ollama Host Path:/mnt/user/appdata/ollama Access Mode: Read/Write

I don't know if you really need all these variables, but it works and I'm just happy right now :D

More info:

- Gemini said you could also get the B70 to run with Vulkan; in his estimation, the card would then run at 30–50% performance. I decided to go with IPEX-LLM. So if anyone's having trouble, it seems there are several ways to do it.

- Also, it’s apparently recommended to enable ASPM in the BIOS to save power, but I’m not sure if it’s on or off for me right now.

- By the way, the card is listed as renderD128 when I run ls /dev/dri.

I hope this helps someone, and that an update comes soon for the Intel GPU TOP plugin, as well as drivers or an official plug-and-play Docker container for apps :)

And please don't forget—I'm no expert either. I spent ages trying to get this to work, but this is how it's currently running for me. Maybe I have something in there that's too much or too little—feel free to leave a comment about it.

Greetings from here to there

Translated with DeepL.com (free version)

Edited April 19Apr 19 by xel87
Screenshot

Followup questions for you:

Have you done any benchmarking with any models on that container?

Or What models have you been using and how do they perform?
Run into any issues since setting it up?