"2026 to Mark Korea's Rise as a Top 3 AI Powerhouse"... Government Launches NPU Performance Standardization Strategy (Comprehensive)

Expanding Measurement Scope to On-Device AI
and Establishing a New Certification System

Paik Kyunghoon, Deputy Prime Minister and Minister of Science and ICT, is delivering a congratulatory speech at the '2025 Artificial Intelligence Semiconductor Future Technology Conference' held at Lotte Hotel in Jung-gu, Seoul on the 10th. 2025.12.10 Photo by Kang Jinhyung

Paik Kyunghoon, Deputy Prime Minister and Minister of Science and ICT, stated that "2026 will be a turning point to determine whether Korea can emerge as one of the world’s top three artificial intelligence (AI) powerhouses." He announced plans to actively promote the widespread adoption of the K-Perf performance indicator as a core foundation for establishing a domestic AI semiconductor ecosystem.

At the '2025 Artificial Intelligence Semiconductor Future Technology Conference' held at Lotte Hotel in Jung-gu, Seoul on December 10, Paik said, "If this year laid the groundwork for Korea to become one of the top three AI powerhouses, then 2026 will be a decisive moment for us to rise as the AI hub of the Asia-Pacific region." He added, "We will elevate our proprietary foundation models to a world-class level and secure competitiveness in domestic neural processing units (NPUs), moving beyond reliance on graphics processing units (GPUs)."

To achieve this, the government will allocate approximately 10 trillion won of next year's 35.5 trillion won national research and development (R&D) budget to the AI sector. Paik emphasized, "We will actively support foundation models and services tailored to each field of expertise so that AI can be embraced not just as a simple tool but as a research partner." He added, "We will create an environment where companies and researchers can achieve results, enabling domestic AI semiconductors to write a 'second K-semiconductor miracle.'"

On this day, the K-Perf council, led by industry, was also officially launched. K-Perf is Korea's first integrated performance indicator designed to measure and compare the performance of domestic AI semiconductors (NPUs) based on actual service environments. This joint project between the government and industry aims to resolve the issue of varying benchmark standards among suppliers, which has made it difficult for client companies to objectively assess NPU performance. K-Perf provides Excel-based performance measurement tables and graphs that account for various variables such as precision and input/output sequence length. Its distinguishing feature is the segmentation of measurement models, conditions, and indicators to reflect the diverse service environments proposed by client companies.

Oh Yunjae, Semiconductor and Quantum Program Manager at the Institute for Information & Communications Technology Planning & Evaluation, explained the structure and necessity of K-Perf at the event. He said, "K-Perf is a starting point that standardizes the models and conditions actually used by client companies, enabling fair comparisons of chip throughput, token throughput, and latency."

Oh also introduced corporate application cases. FuriosaAI, in collaboration with LG AI Research, measured the performance of the 'Exaone 4.0' 32B inference model across various device combinations. Rebellions is conducting real-world performance evaluations by applying its framework in cloud environments. HyperExcel plans to provide K-Perf performance figures for all categories after launching its chip in the first half of next year.

Oh stated, "Going forward, we will expand the measurement scope to include on-device AI and establish certification and verification procedures within the first quarter of next year." He added, "We will open up the development environment to allow schools, research institutes, and developers to participate easily, thereby expanding the ecosystem."

The council will be operated as a public-private joint platform with participation from 12 client and supplier companies and 3 related organizations. Its goal is to create a performance data sharing system that reflects actual industrial operating environments and to establish standardized criteria so that domestic AI semiconductors can be widely adopted in both public and private services.

This content was produced with the assistance of AI translation services.