Is Intel Arc Good for Local LLMs? B580, Pro B60 and B70

Overview

Intel Arc is good for local LLMs when you want the most VRAM per card and will accept a less polished software path: Intel’s spec pages list 12 GB on the Arc B580, 16 GB on the Arc A770, 24 GB on the Arc Pro B60 and 32 GB on the Arc Pro B65 and B70, but the software stack (SYCL, Vulkan, vLLM-based llm-scaler) is younger than CUDA and has open performance bugs. This page gives memory from Intel, the fit math from methodology, and the software status as reported in project docs and issues. Speeds are unknown here unless a project page publishes them with a named machine.

Is Intel Arc good for AI, and which card has how much memory?

Memory size is the first filter for LLMs, and Intel’s own pages (read 2026-10-13) give it:

Intel’s Battlemage ARK page agrees on the bandwidth and Xe-core columns for the B-series cards but leaves its memory-size column empty. Price: check current retailer price; none is quoted here. The Pro cards are workstation products, and the Pro B70 and B65 are covered in arc-pro-b70-llm and arc-pro-b70-vs-r9700.

Memory bandwidth matters because token generation reads the weights each step, but this page does not turn bandwidth into tokens per second. That figure is unknown here.

What LLMs fit on 12, 16, 24 and 32 GB of Arc?

Use the site formula: weights_GB = params × bits ÷ 8 ÷ 1e9, plus KV cache 2 × layers × kv_heads × head_dim × context × bytes, plus an assumed 1 to 2 GB overhead (llm-quantization-explained). Bits per weight are llama.cpp’s Q4_K_M 4.8944 from the quantize README, where that figure is measured on Llama-3.1-8B, so treat it as approximate for Qwen. Parameters and layers come from the Qwen2.5 cards and config.json (14B: 48 layers, 8 KV heads, hidden 5120, 40 heads, bfloat16; 32B: 64 layers, 8 KV heads), head dimension 5120 ÷ 40 = 128.

For example, 14.7e9 × 4.8944 ÷ 8 ÷ 1e9 = 8.9935 GB, and the 32B KV term is 2 × 64 × 8 × 128 × 4,096 × 2 = 1.0737 GB. At 32,768 context the 32B KV cache is 8.5899 GB, so weights 19.8835 + KV 8.5899 + 1 to 2 overhead = 29.4734 to 30.4734 GB, which is under 32 GB but with little margin. Intel’s pages write “GB” without saying decimal or GiB, so treat the margin as approximate. Bigger models are on vram-requirements and gpu-vram-list.

Why is Arc B580 slow on SYCL in llama.cpp?

Some users report that it is, and the cause is not settled. In llama.cpp issue #22413 (“brutally bad SYCL performance on Battlemage”, opened 2026-04-26, closed, labelled bug-unconfirmed) the reporter, using two Arc Pro B50 cards and a B70 on an older PCIe 3.0 Xeon platform, said SYCL generation was much slower than Vulkan on the same machine. A contributor reported much higher SYCL speeds on a B60 under Ubuntu 24.04. A contributor said PCIe and DDR speed mainly affect loading when the model fits in VRAM, recommended bare metal over VMs, said the SYCL optimizations are on by default, and later noted that Q3_K lacked SYCL optimization. These are user reports on mixed hardware, not a vendor benchmark, so no speeds are quoted here. The reporter named a B50 and B70, not a B580; the issue does not diagnose the B580 specifically.

The llama.cpp SYCL doc lists the Arc B580 and Arc A770 as verified devices and publishes one figure with a named card: in its 2025.2 news, llama-2-7b Q4_0 on Arc A770 went from 42 to 55 tokens/s. The rest of that machine is not named on that line, so it does not compare to other cards. The llama.cpp README lists SYCL as the “Intel GPU” backend and Vulkan as the “GPU” backend, so Vulkan is the other route to try.

Does Ollama support Intel Arc?

Ollama’s GPU doc says additional GPU support on Windows and Linux is provided via Vulkan, enabled by default when the backend is installed, and links Intel Linux driver instructions. It lists no Intel model table. The long-running request ollama #1590 is still open, and commenters there call Vulkan experimental, and one reports Vulkan on an Intel iGPU running about half as fast as CPU-only, so treat support as present but unevenly tested. Compare runtimes on llama-cpp-vs-ollama-vs-vllm.

What are ipex-llm and llm-scaler?

ipex-llm was an Intel project for running LLMs on Intel GPUs. Its page says the repository was archived on 2026-01-28, is read-only, and that Intel will not provide or guarantee development or support. Its page also says the project has known security issues. Old guides that depend on it are therefore frozen.

llm-scaler is the current Intel project. Its page says it runs on Arc Pro B60 and B70, ships a vLLM-based component with INT4 and FP8 serving and tensor, pipeline and data parallelism, and an experimental ComfyUI and SGLang Diffusion component. It publishes no numeric benchmarks and names no test machine, so speed is unknown. It does not specify supported operating systems on the page read.

Is Arc A770 x2 better than one Radeon 7900 XTX?

Unknown. The Ask HN thread (2025-03-11) asked exactly that for QwQ 32B and received no tokens-per-second figures. Capacity is the part you can check: two A770 cards total 32 GB across two devices (2 × 16 GB, split across cards, see multi-gpu-llm), while AMD’s side is on amd-gpu-for-local-llm. A single 32 GB Pro B65 or B70 avoids splitting.

What about an Intel Arc eGPU?

Unknown. None of the Intel pages read here state external-enclosure or Thunderbolt support for LLM use.

Questions

Is Intel Arc good for AI?

Yes for capacity per card, with caveats on software. Intel lists 12 GB (B580), 16 GB (A770), 24 GB (Pro B60) and 32 GB (Pro B65, B70). llama.cpp supports Arc through SYCL and Vulkan, but users report SYCL performance problems on Battlemage, and speeds for these cards are unknown on this page.

Why do people keep investing in Intel for AI?

This page has no primary source for anyone’s motives, so it does not answer why. The factual draw visible in Intel’s own specs is more memory per card at the Pro tier: 32 GB on the B65 and B70 against 12 GB on the B580.

2x Arc A770 or 1x Radeon 7900 XTX for llama.cpp?

Unknown on speed; the Ask HN thread that asked this had no figures. Two A770 cards are 2 × 16 GB across two devices; compare the single-card alternative on amd-gpu-for-local-llm.

Why is Arc B580 slow on SYCL in llama.cpp?

Not established. llama.cpp issue #22413 reports slow SYCL on Battlemage cards (B50, B70) versus Vulkan, with a contributor pointing to missing Q3_K optimization, bare metal over VMs, and driver versions. The issue is closed and unconfirmed, and it does not test a B580.

Does Ollama support Intel Arc?

Through Vulkan, per Ollama’s GPU doc, which says Vulkan is enabled by default when the backend is installed. It lists no Intel cards, and issue #1590 is still open with mixed reports.

Is ipex-llm still maintained?

No. Its repository was archived by Intel on 2026-01-28 and is read-only. llm-scaler is the active Intel project for Arc Pro B60 and B70.

Sources read on 2026-10-13

• Arc B580 spec page: 12 GB GDDR6, 192-bit, 456 GB/s, 190 W, PCIe 4.0 x8, 20 Xe cores.
• Arc A770 16GB spec page: 16 GB, 256-bit, 560 GB/s, 225 W, 32 Xe cores.
• Arc Pro B60, Pro B65, Pro B70 spec pages: memory, bandwidth, power, PCIe, Xe cores.
• Intel ARK Battlemage page: bandwidth and Xe-core columns; memory-size column empty.
• intel/llm-scaler and intel/ipex-llm READMEs: supported cards, components, archive notice.
• llama.cpp SYCL.md, llama.cpp README, issue #22413: verified devices, backend table, reported SYCL slowdown.
• Ollama GPU doc and issue #1590: Vulkan support, open Intel request.
• Ask HN 43336272: the A770 versus 7900 XTX question; no figures.
• llama.cpp quantize README (bits per weight), Qwen2.5 14B and 32B config.json.

Related

• arc-pro-b70-llm and arc-pro-b70-vs-r9700 cover the 32 GB Pro card.
• budget-gpu-for-local-llm and best-gpu-for-local-llm place Arc among other options.
• amd-gpu-for-local-llm is the AMD side.
• llama-cpp-vs-ollama-vs-vllm compares runtimes.
• vram-requirements, gpu-vram-list, multi-gpu-llm, methodology.