GIGAZINE has tested the execution speeds of large language models (LLM) and video generation AI using the ASRock "Intel Arc Pro B70 Creator 32GB" graphics card equipped with the Intel Arc Pro B70.
HardwareASRockIntelQwen3.8-27B
Intel Arc Pro B70 Creator 32GB with 32GB VRAM can run local LLMs like Qwen3.8 27B at practical speeds
This article is a translation. Read the Japanese original
Despite its large 32GB VRAM capacity, this product is available on Amazon.co.jp for 298,054 yen (including tax) as of September 17, 2026. Compared to the NVIDIA GeForce RTX 5090, which can exceed 1 million yen due to price surges, this card is priced at less than one-third.
Regarding LLM performance, GIGAZINE used Unsloth Desktop to test "Qwen3.8 27B," achieving decoding speeds of 23.4tok/s for the 4bit quantized build and 15.9tok/s for the 8bit quantized build. The speed of the 8bit quantized build demonstrates practical performance for use cases that do not require real-time responsiveness, such as the continuous operation of AI agents.
In the image generation AI "Z-Image-Turbo," it operated at high speeds, taking 36.23 seconds for the first generation and approximately 5.3 seconds for subsequent images, with image quality reported to be equivalent to NVIDIA GPUs. On the other hand, in tests using the video generation AI "MiniMax H3," while the VRAM capacity was sufficient, the generation speed was reported to be slower than the NVIDIA GeForce RTX 5070 Ti. For content generation AI, since GPU computing performance is as important as VRAM capacity, NVIDIA GPUs are considered more suitable if video generation is the primary goal.