Intel's Arc Pro B70 Targets Local AIGC Production with MiniMax H3 Support

AI video tools are becoming a common part of creative work. An Adobe survey of 384 U.S. video creators found that 71% have used AI video generation or editing tools, and 41% use them weekly. But as usage increases, the cost of cloud-based services becomes more prominent. A video project often requires multiple rounds of generation, modification, and upscaling, making per-use fees and queue times a real financial concern.

A recent social media post circulated an itemized "July bill" from an AI comic studio, showing that out of about 280,000 yuan in revenue, computing costs accounted for nearly 200,000 yuan, leaving little profit after labor, copyright, and rent. Such examples highlight the economic squeeze that cloud-based AI can impose on high-frequency production.

Intel technical experts at the event argued that local AIGC is not merely about saving cloud fees. It also protects data privacy for unpublished material, enables creators to train custom LoRA models, and removes limits on trial and error. They also noted that local deployment does not mean abandoning the cloud; tasks can switch between both depending on needs.

The Arc Pro B70, with 32GB GDDR6 memory, 608 GB/s bandwidth, and 367 TOPS peak INT8 performance, is positioned to handle these local workloads. In an on-site test, running MiniMax's H3 optimized workflow on a B70 generated a 5-second 720p video in about 9 minutes, and upscaling it to 2K took about 7 minutes on the same card.

For batch production, Intel demonstrated a dual-B70 video agent setup where one card runs the local model and agent while the other handles generation. Digital Fold, a partner, presented an alternative: multiple cards processing different generation tasks in parallel. Both approaches point toward a small local production line.

Reducing the entry barrier is another focus. Under Windows, users can install drivers and deploy partner software; under Linux, pre-packaged Docker images simplify setup. Digital Fold has also integrated more than 20 ComfyUI nodes into three core nodes in its DFCine canvas, making complex workflows easier to manage. Still, as open-source models evolve, true usability requires the entire ecosystem—inference frameworks, workflows, and applications—to keep up with new model releases.

An Intel representative said the company is not chasing short-term opportunities, even as professional GPU supply is tight. Instead, Intel values continuous iteration of AI products and building a connected hardware-software-ecosystem, with the eventual goal of offering a "plug-in and use" solution for creators.