Chosun Ilbo reports AI infrastructure is shifting toward LPDDR memory and CPUs to improve cost efficiency.

The competition for cost-effectiveness in the artificial intelligence (AI) infrastructure market is intensifying. As demand for inference, the process of generating answers using already trained models, surges, reducing costs for equipment, power, and cooling has become critical. Accelerators for inference now incorporate low-power DRAM (LPDDR), the type used in smartphones, instead of high-bandwidth memory (HBM), which is expensive. Meanwhile, some AI models are being run on central processing units (CPUs) rather than graphics processing units (GPUs). Cost efficiency, not just peak performance, has emerged as a new benchmark for competition.

Intel, for instance, adopted LPDDR5X instead of HBM in its inference-focused "GPU D-Series," unveiled last month. While LPDDR has slower data transfer speeds, it offers larger capacity at a relatively lower cost. Anil Nanduri, Intel vice president, explained, "Companies want infrastructure that delivers the required performance at an appropriate cost," adding, "We focused on economically scaling high-capacity AI inference." Qualcomm also chose LPDDR for its inference accelerator "AI200," while U.S. startups Possiton and D-Matrix are developing similar technologies.

A shift toward using CPUs instead of GPUs is also underway. AMD has developed technology to handle certain AI tasks, such as document classification, using only its server CPU, "EPYC." By optimizing AI operations and memory usage for CPUs through software called "PACE," the company claims to have improved inference speeds by an average of 1.6 times compared with conventional methods. For businesses, this reduces the burden of purchasing expensive GPUs by leveraging existing CPU servers.

The memory industry anticipates diversifying AI demand. LPDDR, traditionally used in smartphones and laptops, is expanding into AI servers, creating a new growth market. Market research firm TrendForce forecasts that AI servers could surpass smartphones as the largest demand source for LPDDR between 2028 and 2030. The semiconductor industry expects the market to evolve toward dividing semiconductor and memory usage based on task characteristics, as GPUs and HBM remain advantageous for large-scale AI operations.