Intel Arc Pro B70 for Local LLMs: 32GB VRAM Explained

Overview

The Intel Arc Pro B70 for local LLMs is a 32 GB GDDR6 card on a 256-bit bus with 608 GB/s of memory bandwidth and a 230 W TBP, per Intel’s specification page. That 32 GB is enough for a 32B model at Q4_K_M with a long context, and not enough for a 70B model on one card. How fast it runs a model is unknown here, because no vendor page or project README read for this page publishes a tokens-per-second figure with a named machine. Fit arithmetic follows methodology; the wider Intel picture, including the B580, is on intel-arc-for-local-llm.

How much VRAM does the Arc Pro B70 have?

32 GB of GDDR6, a 256-bit memory interface and 608 GB/s of bandwidth. The same Intel page lists 32 Xe-cores, 256 XMX engines, PCI Express 5.0 x16 and a launch of Q1’26 with status Launched. The reference card is two slots, 10.5” x 3.9”, 1020 g, with one 8-pin power connector. Intel lists TBP as 230 W and a TBP-LP range of 160-290 (the page does not state the unit). Price is not on the spec page: check the current retailer price.

What is the difference between Arc Pro B60, B65 and B70?

Memory capacity and bandwidth are the numbers that matter for LLMs, and Intel’s pages give them for all three.

The B65 has the same 32 GB and 608 GB/s as the B70 but 20 Xe-cores instead of 32. For a memory-sized question, whether a model fits, they are equal. The B60 gives up 8 GB and bandwidth. Whether the extra Xe-cores change LLM speed is unknown here.

What fits in 32 GB?

Weights_GB = params x bits / 8 / 1e9, plus KV cache, plus overhead. KV bytes = 2 x layers x kv_heads x head_dim x context x bytes_per_value. Overhead of 1 to 2 GB is an assumption, not a measurement. Bits per weight are llama.cpp’s Q4_K_M 4.8944 and Q8_0 8.5008, as used on vram-requirements; the README gives them in its Llama-3.1-8B table, so applying them to a 32B model is an approximation. Source: llama.cpp quantize README.

Worked example with Qwen2.5-32B: 32.5B parameters, and from its config.json 64 layers, 8 KV heads, hidden size 5120 over 40 attention heads (head dimension 128) and bfloat16 (2 bytes).

Arithmetic: weights 32.5e9 x 4.8944 / 8 / 1e9 = 19.8835; KV per token 2 x 64 x 8 x 128 x 2 = 262,144 bytes, times 4,096 = 1.0737 GB and times 32,768 = 8.5899 GB. Per-model detail is on vram-for-32b-model and the 32 GB class on 32gb-vram-llm.

Can two Arc Pro B70 cards run a 70B model?

Memory-wise, possibly; whether it works well is unknown. The llama.cpp quantize README lists Llama 3.1 70B Q4_K_M at 43.1 GB, which does not fit one 32 GB card but is below the 64 GB two cards add up to (2 x 32). If layers split evenly, that is 43.1 / 2 = 21.55 GB of weights per card (an even split is an assumption). The KV cache for that model is unknown here because its config was not read, and speed and card-to-card scaling are unknown. The llm-scaler README lists tensor, pipeline and data parallel for vLLM; the vLLM Intel doc calls pipeline parallel a beta feature for online serving. Planning notes are on multi-gpu-llm, multi-gpu-pcie-and-psu and vram-for-70b-model.

Does the Arc Pro B70 work with Ollama, llama.cpp and vLLM?

What the primary sources say, and nothing more:

• Intel’s own stack. The intel/llm-scaler README describes a GenAI solution running on “Intel Arc Pro B60 and B70 GPUs”, with a vLLM-based text component. It publishes no benchmark figures and names no test machine.
• llama.cpp. The llama.cpp README backend table lists SYCL for “Intel GPU” and Vulkan for “GPU”. It does not name the Arc Pro B70.
• Ollama. The Ollama GPU doc says Vulkan is enabled by default when the backend is installed, can be turned off with OLLAMA_VULKAN=0, and links Linux Intel GPU instructions. Arc is not mentioned, so B70 support in Ollama is unverified here.
• vLLM. The vLLM Intel GPU doc lists “Intel Data Center GPU, Intel ARC GPU” and says it “initially supports basic model inference and serving”. The B70 is not named there.

No source read states that a given model runs at a given speed on a B70.

Is the Intel Arc Pro B70 good for AI?

It is good for capacity: 32 GB of VRAM is the headline. The AMD rival is compared on arc-pro-b70-vs-r9700 and radeon-ai-pro-r9700-llm. It is unproven for speed on this site, and the software route (llm-scaler, SYCL, Vulkan) is narrower than the CUDA path. Pick it for fit per card, then check current retailer price and the backend you will use before buying. All cards are listed on gpu-vram-list.

Questions

Is the Intel Arc Pro B70 good for AI?

It has 32 GB of GDDR6 and 608 GB/s of bandwidth per Intel’s specification page, which is enough to hold a 32B model at Q4_K_M with 32,768 context (29.4734 to 30.4734 GB with assumed overhead). Speed is unknown because no source read publishes a tokens-per-second figure for it, and the software path is narrower than CUDA.

Does the Arc Pro B70 work with Ollama, llama.cpp and vLLM?

Intel’s llm-scaler README names the B70 for its vLLM-based stack. The llama.cpp README lists SYCL for Intel GPUs and a generic Vulkan backend, and the Ollama GPU doc describes Vulkan and Intel GPU instructions without naming Arc. vLLM’s install doc lists Intel Arc GPUs and says it initially supports basic inference and serving. None of these pages names the B70 except llm-scaler.

How much VRAM does the Arc Pro B70 have?

32 GB of GDDR6 on a 256-bit interface with 608 GB/s of bandwidth, per Intel’s specification page. The Arc Pro B65 has the same 32 GB.

What is the difference between Arc Pro B60, B65 and B70?

The B60 has 24 GB, a 192-bit bus, 456 GB/s and PCIe 5.0 x8. The B65 and B70 both have 32 GB, 256-bit and 608 GB/s, but the B65 has 20 Xe-cores and 200 W TBP against the B70’s 32 Xe-cores and 230 W. Speed difference between them is unknown.

Can two Arc Pro B70 cards run a 70B model?

Their combined 64 GB is above the 43.1 GB the llama.cpp README lists for Llama 3.1 70B Q4_K_M weights, so it fits on paper before KV cache and overhead. Whether it runs, and how fast, is unknown here.

Is there a tokens-per-second figure for the B70?

No. The llm-scaler README publishes no benchmark numbers and names no test machine, so speed is written as unknown.

Sources read on 2026-10-13

• Intel Arc Pro B70 specifications: 32 GB GDDR6, 256-bit, 608 GB/s, 230 W TBP, TBP-LP 160-290 (unit not stated), 32 Xe-cores, 256 XMX engines, PCIe 5.0 x16, Q1’26, reference card size, weight and connector.
• Intel Arc Pro B65 specifications: 32 GB GDDR6, 256-bit, 608 GB/s, 200 W, 20 Xe-cores, PCIe 5.0 x16, Q1’26.
• Intel Arc Pro B60 specifications: 24 GB GDDR6, 192-bit, 456 GB/s, 200 W, 20 Xe-cores, PCIe 5.0 x8, Q2’25.
• intel/llm-scaler README: B60 and B70 named, vLLM component, parallelism features, no benchmark numbers.
• llama.cpp README: SYCL “Intel GPU” and Vulkan “GPU” backend rows.
• Ollama GPU doc: Vulkan default, OLLAMA_VULKAN=0, Intel GPU instructions link, Arc not mentioned.
• vLLM Intel GPU install doc: Intel Data Center GPU and Intel ARC GPU, basic inference and serving, tensor parallel, pipeline parallel beta for online serving.
• llama.cpp quantize README: Llama 3.1 70B Q4_K_M 43.1 GB; bits per weight 4.8944 and 8.5008 (bits per weight come from its Llama-3.1-8B table; re-verified 2026-10-13).
• Qwen2.5-32B model card and config.json: 32.5B parameters, 64 layers, 8 KV heads, 40 attention heads, hidden size 5120, bfloat16.

Related

• intel-arc-for-local-llm covers the wider Intel story and the B580.
• arc-pro-b70-vs-r9700 and radeon-ai-pro-r9700-llm cover the AMD rival.
• 32gb-vram-llm and vram-for-32b-model cover what fits in 32 GB.
• multi-gpu-llm, multi-gpu-pcie-and-psu and vram-for-70b-model cover two cards and 70B.
• gpu-vram-list, llama-cpp-vs-ollama-vs-vllm and methodology cover lists, runtimes and the citation rule.