Businesses and devs main open opportunities for AI PCs

Omdia’s AI PCs: An Emerging Device Class finds that there are two basic technology options to deliver AI inference on PCs – one based on a traditional PC architecture enhanced with a high performance GPU, and one based on a smartphone-style system-on-chip incorporating dedicated AI acceleration. Some PC manufacturers have an early lead with the first option, having begun shipping powerful NVIDIA GPUs as early as 2019, while Apple essentially defined the second with its Apple Silicon MacBooks.

The PC concept was always intended to be personal, and Omdia researchers created a range of user personas to understand how different users’ needs for applications and performance requirements would differ. The “Corporate Warrior” persona, which describes business users who make heavy use of collaboration tools and travel frequently, is the biggest and the least well-served by the two dominant players. As such it is being heavily targeted by Intel’s Core Ultra CPUs as the company battles to break back into the AI space. However, Arm-based systems like Apple’s have an important edge.

“You can’t underestimate the importance of unified memory to AI on the PC,” said Alexander Harrowell, Omdia’s Principal Analyst for Advanced Computing. “The limiting factor for the kind of AI model you can even run is VRAM – the very popular 7B parameter class of AI models, such as Mistral-7B, need at least 8GB - and the limiting factor for performance is memory bandwidth. Arm’s investment in its unified memory architecture going back as far as the 1980s is a major advantage.”

The report is available now in Omdia’s Advanced Computing Intelligence Service and, as well as technology and user personas, covers the major silicon vendors and PC OEMs.