SKT to Verify GPU-NPU Combined Inference Infrastructure

According to AI TIMES, SK Telecom announced on November 9 that it has begun joint development and verification of a Korean-style sovereign AI inference infrastructure that combines GPU and domestically produced NPU technology together with Nvidia and Rebellions. The key is to test in actual data centers whether a split inference architecture, in which Nvidia GPUs handle prefill and Rebellions NPUs handle decode, can improve performance and power efficiency. SK Telecom, drawing on its experience operating AI data centers, will examine the stability and applicability of the integrated system, and it is also pursuing ways to connect it to proprietary AI models and public AI services. The collaboration is also an effort to expand the use of domestic AI semiconductors by linking them with the global AI computing ecosystem. The article noted that, more than the technical connection itself, the variable that will shape future competitiveness is whether the workload can be divided smoothly in real service environments and lead to improved operational efficiency. It also highlighted the possibility that this could become a test case for creating a Korean-style AI infrastructure that can be used in both the public and private sectors.