Intel Arc for local AI: can the B580 and A770 run LLMs?

Intel Arc cards can run local LLMs, and on paper they're good value. The A770 16 GB and the B580 12 GB have more memory or bandwidth than NVIDIA's entry-level cards, like the RTX 4060 with 8 GB. The trade-off is software: fewer tools support Arc, setup takes more effort, and real speeds often trail what the hardware should manage.

What the hardware offers

Those estimates come from our calculator, which assumes the software uses the card's memory bandwidth well. On Arc that's less certain than on NVIDIA, so treat them as a ceiling.

The A770 16 GB is the interesting one: 16 GB fits gpt-oss 20B, which our calculator puts at about 98 tok/s.

How to run models on Arc

There are a few routes, and which is best changes as drivers and tools improve:

• Vulkan. llama.cpp's Vulkan backend works on Arc, and LM Studio offers a Vulkan runtime. This is the simplest way to start, on Windows or Linux.
• SYCL. llama.cpp also has a SYCL backend, built on Intel's oneAPI toolkit. It can be faster than Vulkan but takes more setup.
• Intel's own builds. Intel has published optimized builds and containers of popular tools for its GPUs. Check Intel's current documentation for what's maintained.

Ollama's support for Intel GPUs has varied between versions. Check the current Ollama documentation for Arc support before you rely on it. If it isn't there, LM Studio with Vulkan is the easy fallback. See Ollama vs llama.cpp vs LM Studio.

The catches

• Software support is the narrowest of the three GPU brands. Many AI tools assume NVIDIA's CUDA. Some have Intel support, some don't.
• Speeds vary by backend and driver. The same card can be much faster on one setup than another. Update drivers before judging it.
• Resizable BAR matters. Arc cards perform poorly without it. Make sure it's enabled in your motherboard's settings. Older PCs, including many used workstations, may not support it, which makes Arc a poor fit for an old Xeon box.
• The 8 GB A770. Check listings carefully: the 16 GB version is the one you want.
• Multi-GPU with Arc is less well trodden than with NVIDIA.

Should you buy one for AI?

• Yes, if you want a new card with warranty, on a budget, and you're comfortable trying a couple of backends. The B580 with 12 GB runs the same models as an RTX 3060 12 GB, with more bandwidth on paper.
• Maybe, if you want 16 GB cheaply: compare the A770 16 GB against a Tesla P100, which has much faster memory but needs a shroud, and against 16 GB cards from NVIDIA and AMD.
• No, if you want everything to just work, or plan to use tools beyond chat. The RTX 3060 12 GB is the safe starter, and a used RTX 3090 is the safe step up.

Already own an Arc card? Start with Qwen3 8B or Qwen3 14B in LM Studio, then send us your numbers. Real results help others decide.