Case Study: Naver reduces AI serving costs and maintains performance with PyTorch and Intel oneAPI-powered CPU optimization

Naver saves $340K annually and boosts model serving up to 7x with PyTorch

Naver, a major South Korean internet company, faced a challenge with its AI model servers. With demand growing and high-cost GPU resources limited, the company needed to transition its relatively lightweight AI workloads from expensive GPU servers to more cost-effective Intel CPU servers without sacrificing performance or service quality.

With support from PyTorch and by utilizing Intel® Extension for PyTorch, Naver implemented a solution involving optimal thread configuration and core pinning to overcome CPU bottlenecks. The company also used knowledge distillation to lightweight its models. This allowed Naver to successfully transition three key models to CPU servers, achieving performance improvements of up to 33x and annual cost savings of approximately $340,000 by freeing up 30 GPUs.

Naver