AI Recommendation Systems: A Key Revenue Model for Big Tech Companies
KAIST Develops PIM-Based AI Accelerator Semiconductor System
PIM, a next-generation semiconductor that adds AI computing capabilities to memory.


A domestic research team has successfully developed an intelligent semiconductor system optimized for accelerating AI recommendation system learning algorithms based on Processing-In-Memory (PIM) technology.

A research team led by Professor Min-soo Yoo of the Department of Electrical and Electronic Engineering at the Korea Advanced Institute of Science and Technology (KAIST) announced on the 16th that it had developed a memory-centric AI accelerator semiconductor system based on PIM technology.

AI recommendation system algorithms are the AI technology that big tech companies like Google, Amazon, YouTube, and Facebook use to create content recommendations and personalized advertisements. As revenue from online advertising is a primary revenue model for Silicon Valley big tech companies like Google and Facebook, demand for advanced recommendation AI technology has been rapidly increasing in recent years.

According to data recently disclosed by Facebook, 70% of AI computations processed in Facebook data centers are used to process recommendation algorithms, and 50% of computing resources for AI algorithm training are used to train recommendation algorithms.

Accordingly, the research team succeeded in developing an intelligent semiconductor system based on PIM technology with AI computational functions added to memory semiconductors.

The von Neumann architecture, consisting of three elements—processor, memory, and program—is a standard computer architecture. Memory stores programs, and the processor retrieves and processes the stored programs from memory. However, because the processor is faster than the memory, delays are unavoidable.

PIM technology integrates CPU and RAM into a single semiconductor, effectively demonstrating a memory-centric design. Research team members explained that the developed system accelerates the learning process for AI recommendation algorithms by up to 21 times compared to existing AI acceleration systems using NVIDIA GPUs.

The research team stated that the results of this study are significant in that they suggest the possibility of commercialization and success of PIM technology designed around memory in the global AI semiconductor market, which is expected to see significant increase in demand and rapid growth in the future.

Meanwhile, the results of the study are scheduled to be presented at the International Symposium on High-Performance Computing Architectures (HPCA), hosted by IEEE, in February next year.