Life with Arc Pro B70 + B60, part 1

The build#

I wanted to build my own RAG system, where AI draws on my documents when answering questions. I also found the safety restrictions of public models such as Fable and Astra limiting. That made me want to run models locally and control their settings myself.

I chose two Intel Arc Pro cards: a B70 with 32 GB and a B60 with 24 GB of memory. Together they have 56 GB of VRAM, the memory on the graphics cards that holds the models.

The rest of the build includes a Ryzen 9 9900X, 64 GB of RAM and a 1 TB SSD. The complete computer cost CZK 100,331 including VAT in May 2026. It runs Windows 11 Pro.

How fast it runs#

I use llama.cpp to run the models. I tried Vulkan and SYCL, two ways to run calculations on the graphics cards. Vulkan was faster in most tests. GPT OSS 20B showed the biggest gap: on the B70, it generated 132 tokens per second with Vulkan and 56 with SYCL. A token is a short piece of text.

These were the answer generation speeds with Vulkan:

The large 70B model did not fit on one card. Its compressed Q4_K_M version took 39.6 GiB and worked across both. SYCL managed up to 11.6 tokens per second, but I kept Vulkan for everyday use. It processed prompts faster and left more memory free.

What slowed me down#

Windows reported incorrect details about the cards’ connection and memory. I had to check the actual state with other tools and measurements.

Vulkan changed the device order between runs, so one test accidentally used the integrated graphics. Selecting the card by name solved that problem. Older PowerShell also sent Czech characters with the wrong encoding, which I had to fix in the requests.

After the first day#

Chat, coding help and a large model across both cards all work. There were no driver resets during the day of testing. Most of the setup work went into checking that the software used the right card.

Vulkan has worked best for my needs so far. These results apply to this machine and these software versions.

Measured on 28 August 2026 with Windows 11 26200.9168, llama.cpp b10666 and driver 32.0.101.8805. The benchmark used 512 input tokens and 128 output tokens, repeated five times. I also tested actual conversations.

Detailed results#

B70 alone: Vulkan and SYCL#

All numbers are in tokens per second. Reading means processing the prompt, and writing means generating the answer. Higher numbers mean faster performance.

The 70B model across both cards#

DeepSeek R1 Distill Llama 70B in Q4_K_M format. Layer mode assigns whole model layers to each card. Tensor mode also splits calculations within layers. The 4:3 and 5:3 ratios set how work is divided between the B70 and B60.

I chose Vulkan layer for everyday use. SYCL tensor generated answers faster, but with a context of 16,000 tokens it left only about 0.5 GiB of free memory on the B60.

The complete build#

Prices are from the May 2026 invoices, in CZK and rounded to whole crowns.

Operating system: Windows 11 Pro 25H2. Its price is not included in the hardware total.