Arc Pro B70 and openclaw

[MortchTheAIGuy]

Difficulty Using Dual Intel Arc Pro B70 GPUs with OpenClaw and Local AI Models

I recently purchased two Intel Arc Pro B70 GPUs with the intention of using them for local AI inference and agent workloads. Both cards are installed in a Windows 11 system with an Intel Core i5-14600K, and I also use WSL2 with Ubuntu 24.04.

My primary goal was to use the B70 GPUs as the local inference hardware for an OpenClaw agent. Unfortunately, I have been unable to find a stable and straightforward software configuration that allows OpenClaw to reliably use the cards.

The main difficulty appears to be the software stack rather than the hardware itself. OpenClaw generally expects to connect to an inference backend such as Ollama or another OpenAI-compatible local server. NVIDIA hardware has a well-established path through CUDA, but the equivalent workflow for the Arc Pro B70 is much less clear.

I have encountered several practical issues:

• It is unclear which Windows or WSL inference backend currently provides reliable support for the Arc Pro B70.

• Many local-AI installation guides and OpenClaw examples assume an NVIDIA GPU and CUDA.

• Ollama does not appear to offer the same simple or mature Intel GPU workflow that it provides for NVIDIA hardware.

• Intel-specific options involving oneAPI, Level Zero, OpenVINO, IPEX-LLM, llama.cpp, or SYCL are difficult to evaluate because support varies between projects, operating systems, GPU generations, and model formats.

• Even when a backend can potentially access the GPU, it may not expose the OpenAI-compatible API or model-management features that OpenClaw expects.

• Using two B70 cards adds another layer of uncertainty. It is not clear which available inference engines can reliably divide a model between both GPUs or otherwise make effective use of both cards.

• Documentation specifically addressing the Arc Pro B70, Windows 11, WSL2, multi-GPU inference, and agent frameworks is currently limited.

At this point, I am considering moving the main OpenClaw inference workload back to an older Ubuntu system containing an NVIDIA RTX 3090 and Ollama. That combination is considerably easier to configure and is already functioning as a local inference endpoint. I would then try to use the two B70 cards for separate worker tasks rather than as the primary OpenClaw “brain.”

I purchased the B70 cards specifically because their memory capacity and professional AI positioning seemed well suited to local models. I would like to continue using them, but I need a clearer supported path.

I would appreciate guidance from Intel or other community members on the following:

1. What is the recommended inference stack for the Arc Pro B70 on Windows 11 or WSL2?

2. Is there a supported OpenAI-compatible inference server that works reliably with the B70?

3. Can two B70 cards currently be used together for a single model, either through tensor parallelism or model layer splitting?

4. Is native Windows or native Linux currently the better-supported environment?

5. Are there any official installation guides, validated model lists, containers, or reference configurations for local LLM inference on the B70?

6. Does Intel have a recommended way to connect an agent framework such as OpenClaw to an Arc Pro inference endpoint?

I am comfortable working with Windows, Linux, WSL, Docker, command-line tools, and local model servers. I am not necessarily looking for a one-click solution, but I would like to know which software path Intel considers stable and supported for these cards.

Any guidance, working configuration examples, or clarification regarding the current state of B70 multi-GPU LLM support would be greatly appreciated.

[ChristianA_Intel]

Hello MortchTheAIGuy,

I will further investigate this matter to ensure that I can provide you relevant information. Once I have an update, I will make sure to post the response here on this thread as soon as it becomes available.

In the meantime, if you have any additional questions, clarifications, or new observations that you would like to share, please feel free to let us know.

Thank you for understanding.

Best regards,

Christian Louierico A.

Intel Customer Support Technician

[ChristianA_Intel]

Hello MortchTheAIGuy,

Thank you for your patience.

I have reviewed the concerns you shared and noticed that there are multiple concerns involved. To ensure we can properly address each concern, could you please let me know which specific concern you would like to focus on in this thread? For any additional concerns, please create a new thread so we can address them separately.

Thank you for your understanding, and I look forward to your response.

Best regards,

Christian Louierico A.

Intel Customer Support Technician

[MortchTheAIGuy]

I guess the main issue for this thread can be how to get the B70s working with an agent harness. It would be nice to have the Agent's brain running on the B70.

[ChristianA_Intel]

Hello MortchTheAIGuy,

Thank you for sharing the requested information.

I will continue investigating this matter and will get back to you as soon as I have an update or if any additional details are needed.

Best regards,

Christian Louierico A.

Intel Customer Support Technician

[Vaultia]

Since you narrowed it down to "getting the agent's brain running on the B70," I can share a config that's been stable for me — I run 2x Arc Pro B70 (same as you, 64 GB total) as a production local inference endpoint for agent workloads, so this is from daily use rather than theory.

The short answer: llama.cpp built for SYCL, running llama-server, on native Linux. That combination gives you a stable OpenAI-compatible endpoint your agent harness can point at directly. Here's the detail.

Stack

• Native Ubuntu (not WSL2 — see below), oneAPI 2025.3, Level-Zero driver 20.2.0.
• llama.cpp built with -DGGML_SYCL=ON -DCMAKE_CXX_FLAGS_RELEASE="-O3 -DNDEBUG".
• Launch llama-server — it exposes an OpenAI-compatible API at /v1, so your agent (OpenClaw or anything expecting an OpenAI endpoint) just points at http://<host>:<port>/v1. No CUDA-specific glue, no Ollama Intel-workflow gaps.

Native Linux vs WSL2
I'd move off WSL2 for the inference host. The SYCL/Level-Zero path is much more predictable on native Linux; WSL2's translation layer caused more trouble than it was worth for multi-GPU. If the B70 box can dual-boot or run Linux natively, the stability jump is significant. (Your RTX 3090 + Ollama fallback is fine as a stopgap, but you won't need it once this is set up.)

Using both B70s
Two working patterns, depending on the model:

• One model per card (--main-gpu 0 / --main-gpu 1), two independent llama-server instances. Best for dense models — avoids inter-card overhead.
• Single model split across both cards with --split-mode layer (never row — segfaults on SYCL). For a large MoE this is nearly free: Qwen3-Coder-Next 80B-A3B (Q4_K_M, ~45 GB) runs ~71 t/s split across both B70s here, because only ~3B params are active per token. For dense models the split costs throughput, so pinning per card is usually better.

So yes — two B70s work together for one model via layer split in llama.cpp today. Tensor-parallel via vLLM XPU also gets the GPUs to talk (host-shared-memory CCL config), but model support is still shaky in my testing, so I keep llama.cpp for anything real.

A speed bonus for the agent brain
Gemma 4 ships MTP speculative-decoding drafters. With the matched QAT pair (QAT target + its QAT MTP drafter), I get 2.2x on the 31B and 2.95x on the 12B, lossless, on a single B70 — e.g. Gemma 4 12B goes from ~32 to ~94 t/s, and the 31B from ~23.5 to ~51-61 t/s. For an agent that makes lots of calls, that roughly doubles what one card can serve. Command shape:
llama-server -m <target-qat>.gguf --model-draft <mtp>.gguf --spec-type draft-mtp --spec-draft-n-max 4 -ngl 999 -fa on
(Gotcha: target and drafter must be the matched QAT pair, or acceptance collapses and you get no gain.)

I put together a full reproducible writeup (exact versions, build, model downloads, benchmark commands) while setting this up — happy to share if it'd help. And glad to help troubleshoot specific errors if you post them; the B70s are genuinely capable once the stack is right.

— Vaultia (sovereign local AI infra, Québec)