I've finally got AI working on my B70. I had searched a lot online and nothing was coming up due to software support not being available. Here is what I found was working, and what you'll need.


- The graphics card NEEDS to be its firmware updated. You can follow the guides here from Stupifier or by having it connected to a baremetal PC running windows first! The Intel driver installer contains a firmware update for the card, and these are the only delivery methods. Personally, I would not suggest doing this through a pass-through VM.
- Bios up-to-date, Resizable BAR (ReBAR) enabled, 'Above 4G' enabled.
- Running Unraid 7.3.1. or above.
- Have the plugin 'Intel GPU TOP' installed.

I have tried multiple AI docker containers, and found that only llama.cpp works best, others don't work, or have very very little functionality. Check previous findings here if you're interested. Here are my settings for this docker container.

Llama.cpp

Post Arguments:--models-preset /models/presets.ini --host 0.0.0.0 --port 8000 --models-max 1 -ngl 999 --device SYCL0 --flash-attn on

WebUI: (your desire port here)

Container Port: 8000

GPU device: /dev/dri

Container Device: /dev/dri

ZES_ENABLE_SYSMAN: 1

Container Variable: 1

GGML_SYCL_DEVICE: 0

Notes:
- The model location has to be inside of your selected models folder defined in "Model Storage Path". For the post argument, you only need to run the included script below, I use 'User Scripts' to run it as I add models, and that's it.
- My setup has an intel 14700K which has integrated iGPU, it is possible you may need to change the Post Argument last part to --device SYCL1 (or maybe remove that part), if you have more cards or a different setup.

Generation presets.sh

Edit this script to edit the path to your models, and to add your graphics card information (such as the VRAM, etc) so that it can accurately setup the settings for the models.

Con to this setup, when you have to change the model you have to do it in the post argument, it can't be done from the UI or from open-webUI.

My open-webUI connection looks like this:

Edited July 16Jul 16 by JustOverride
Instructions updated to add multi model sections!

Since Latest version of Unraid uses 6.x kernel and new GPUs like this show large improvements using a 7.x kernel......would it not be better to just use SR-IOV to pass the GPU into a VM and do your work there?

At least until Unraid kernel version catches up?

Unraid is limited to runinng LTS kernels becuase of ZFS, next LTS kernel shodul be out at then end of this year.

14 hours ago, Stupifier said:

Since Latest version of Unraid uses 6.x kernel and new GPUs like this show large improvements using a 7.x kernel......would it not be better to just use SR-IOV to pass the GPU into a VM and do your work there?

At least until Unraid kernel version catches up?

I thought of this but the best way to keep the most resources is to not have to split the card and not have to run a VM. This works pretty well and will hold me up until better support becomes available.

ONE NOTE: I forgot to add, this (and maybe other ARC cards as well) need to be connected to a baremetal PC running windows first! The Intel Drivers installer also contains a firmware update for the card, and this is the only delivery method. I would not suggest doing this through a pass-through VM.

13 hours ago, JustOverride said:

ONE NOTE: I forgot to add, this (and maybe other ARC cards as well) need to be connected to a baremetal PC running windows first! The Intel Drivers installer also contains a firmware update for the card, and this is the only delivery method. I would not suggest doing this through a pass-through VM.

OMG....That sucks! I don't have a baremetal PC lying around that can hold a card like that. Yikes......I technically could temporarily repurpose my unraid server to install Windows on an SSD, but that is a pain to do.

Hard to even find a changelog to see what is in any firmware update

Edited June 20Jun 20 by Stupifier

Here is the information I found, I've checked it for accuracy. Sources included. At the end of the day, you always want firmware up-to-date.


The bundled firmware update inside the Intel Arc Windows driver installer primarily delivers low-level Video BIOS (vBIOS) and Graphics Security Control (GSC) microcode updates directly to the hardware. Intel integrated this behavior to ensure cards function correctly without relying on separate flashing tools or Windows Update. [1, 2, 3, 4]

Because these updates are embedded directly within the standard driver packages, they do not usually have standalone release notes, but they provide the foundational mechanics for several critical GPU behaviors. [1, 2]

What is Included and Managed by the Firmware Update

• HDMI & DisplayPort Compatibility: Updates core display handshaking protocols to resolve "no signal" issues and blank screens when waking a computer from sleep or during a cold boot. [1]

• System Stability Fixes: Adjusts low-level power states to fix hard system hangs, sleep-to-wake crashes, and erratic power-cycling bugs inherent to the physical board. [1, 2]

• Fan Curves & Thermal Management: Recalibrates fan controller behavior on reference models to optimize thermal throttling limits and reduce fan-speed oscillation.

• Resizable BAR (ReBAR) Optimization: Tweaks how the GPU maps its VRAM memory allocation to the CPU, which is vital for the card's overall stability and frame-rate delivery.

• Security & Virtualization Adjustments: Modifies the integrated Graphics Security Control (GSC) processor. On enterprise variant cards (like the Arc Pro series), firmware changes are used to regulate and limit SR-IOV Virtual Functions. [1, 2, 3]

Important Installation Notes

• It Installs Automatically: The Intel Arc driver installer will flash the firmware automatically if it detects your card is running an older vBIOS version.

• Longer Install Times: When a firmware bundle is being applied, the installer will take significantly longer to complete. Your screen may flicker or remain black longer than a normal driver update.

• Do Not Interrupt: It is critical that you do not power down, restart, or unplug your PC while this installer is running to avoid corrupting the GPU's onboard flash memory. [1]

Maybe this can eventually be used to update firmware on Intel Arc GPU from Linux

GitHub

GitHub - fwupd/fwupd: A system daemon to allow session so...

A system daemon to allow session software to update firmware - fwupd/fwupd

Hi guys. I have just purchased but not yet received an intel arc pro b50. I intent on using this in my unraid server to use the sr iov feature to split the card between running windows vm's and docker. The system has a supermicro h11ssl V2 motherboard with epic 7k62 48 core CPU and 128gb ram etc.

After doing more research ( perhaps too late!) I find my motherboard does not support resizable bar, at least not without hacking the bios which I don't want to do. It does support the above 4gb decode and sr iov. I also read that with some operating systems resizable bar is not important, as the os can deal with that if you add a command to the startup sequence. Resizable bar is an issue more for window machines? And as I am not passing the entire card to any one Windows VM.

So my query is will I need resizable bar or not with the latest unraid? Do I need to a command to the go file or similar?

Thanks in advance, des.

Using Unraid, I successfully updated the Firmware on my ASRock Intel Arc Pro B70 GPU!

@JustOverride Both of these guides outline the procedure to update firmware.
https://forum.level1techs.com/t/remember-to-update-your-intel-arc-firmware-on-linux/208736
https://gist.github.com/mploschiavo/9968c883c4a872a74e0f38edd7cda2ef

You just need to make sure you pay close attention to the Device IDs. For the ASRock Intel Arc Pro B70 (creator edition):

Before update
```
❯ sudo igsc fw version

Device: FW Version: BMG__31.1058

❯ sudo igsc list-devices

Device [1] '/dev/mei0': 8086:e223 1849:6025 0000:10:00.00
```

After update
```
❯ sudo igsc fw version

Device: FW Version: BMG__31.1062

❯ sudo igsc list-devices

Device [1] '/dev/mei0': 8086:e223 1849:6025 0000:10:00.00
```

So my device IDs are: e223 and 6025. But on the standard Intel version of Intel Arc Pro B70, the device IDs are e223 and 1707.

Important thing here is that Intel firmware is divided into four independent firmware bin files (fw-code, fw-data, oprom-code, oprom-data). People say each can be flashed independent of the other. The latest firmware update files did not contain a compatible oprom-data bin file for me. All I could find is the 1707 variant, not the 6025 variant I need for the ASRock version I own.

So I ultimately only flashed fw-code and fw-data. Rebooted my unraid server. And it worked!

On 6/26/2026 at 9:05 PM, Dieseldes said:

Hi guys. I have just purchased but not yet received an intel arc pro b50. I intent on using this in my unraid server to use the sr iov feature to split the card between running windows vm's and docker. The system has a supermicro h11ssl V2 motherboard with epic 7k62 48 core CPU and 128gb ram etc.

After doing more research ( perhaps too late!) I find my motherboard does not support resizable bar, at least not without hacking the bios which I don't want to do. It does support the above 4gb decode and sr iov. I also read that with some operating systems resizable bar is not important, as the os can deal with that if you add a command to the startup sequence. Resizable bar is an issue more for window machines? And as I am not passing the entire card to any one Windows VM.

So my query is will I need resizable bar or not with the latest unraid? Do I need to a command to the go file or similar?

Thanks in advance, des.

Just responding to my own queries. I sent the GPU back without testing as it seems resizable bar is a must. Thanks

Instructions updated to add multi model sections to the WebUI!
Simply update the post argument to the new one added in the main post, run (or add to User Scripts) the attached above script to generate your 'presets.ini' file, and then that file will automatically list and pass all of your models to WebUI. This will enable you to have switchable models on WebUI without having to manually reload the llama.cpp container.

Update for folks. There is a custom VLLM container regularly being updated and It is custom-tailored for Intel Arc Pro B70 GPU!
It is blowing any custom llama.cpp container out of the water.

I would highly recommend folks try this over llama.cpp.

https://gist.github.com/urakozz/810b7c642b689f252665f9d7278b763f

📊 **Intel Arc Pro B70 (Battlemage Xe2) LLM Benchmarks**

🔹 vLLM (GPTQ + MTP Speculative)
• Model: `palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4`
• Prompt Processing (PP4096): 10,676 t/s
• Time to First Token: 454 ms
• Token Generation (TG256): 96.51 t/s *(Peak: 112.67)*
• Method: MTP Speculative (`num_tokens: 4`)

🔹 llama.cpp (SYCL Baseline)
• Model: `Qwen 3.6 35B MoE (GGUF)`
• Prompt Processing (PP4096): ~1,500 – 2,500 t/s
• Time to First Token: ~1,800 – 2,500 ms
• Token Generation (TG256): ~60 – 70 t/s
• Method: Standard Sequential

Edited July 25Jul 25 by Stupifier