Benchmark Foundry Holds Key to Korea's AI Chip Ambitions

The government recently announced its "K-Nvidia" plan to foster the domestic artificial intelligence semiconductor industry. Recognizing that securing actual end-users is essential for domestic chips to gain market competitiveness, the plan includes incorporating domestically produced neural processing units (NPUs) into public procurement systems, supporting joint development and demonstration with demand-side companies, and prioritizing domestic NPUs in public sectors such as defense and security.

While this policy direction is sound, concerns remain that current measures may fall short of achieving meaningful adoption of domestic semiconductors in industry. A case in point is the National AI Computing Center project. Although domestic semiconductor adoption was initially pursued, negative feedback from cloud operators who would actually run the data centers ultimately led to a decision to procure only Nvidia graphics processing units (GPUs). This demonstrates that significant barriers to domestic chip adoption exist in the field despite government support.

Beyond direct demand creation, there are indirect ways to expand demand. The government could establish a trust foundation by certifying product performance and providing practical references. Similar to how the KS Mark once officially certified product performance through KS standards, a national verification system for AI semiconductor performance could be introduced.

Admittedly, AI semiconductor certification is far more complex than KS Mark standards. Semiconductor companies typically demonstrate their capabilities through benchmarks—standardized test programs that quantify chip computing performance. Geekbench scores used for smartphone chip evaluation are a prime example.

In the AI semiconductor field, MLPerf (Machine Learning Performance) serves as the leading benchmark. While domestic AI chip companies promote their MLPerf scores, these alone are insufficient for potential customers to make adoption decisions. Companies lack confidence in real-world performance within their own environments and long-term operational stability. To build customer trust, companies must go beyond simply presenting benchmark scores. Tests should be tailored to applications that potential customers actually use, with results verified in real operating environments. Data compiled this way functions as reliable references beyond mere benchmark scores. This represents a "benchmark foundry" role—providing customized reliability and performance assessments based on standard benchmarks, much as semiconductor foundries manufacture chips tailored to customer design requirements using process technology and basic designs.

An organization performing this "benchmark foundry" role could become the hub of a semiconductor ecosystem connecting end-users with chip manufacturers. If university students gain experience with domestic semiconductors through these benchmarks, they can drive broader adoption in industry after graduation. Government support for such a "benchmark foundry" could help domestic semiconductors earn initial market trust and establish a foundation for ecosystem development.