SK Telecom, NVIDIA & Rebellions Develop Integrated GPU and NPU System for Sovereign AI - The Fast Mode

The agreement was signed at NVIDIA AI Day Seoul, held at COEX in Seoul. The partnership will use NVIDIA's NVLink Fusion technology to connect Rebellions' next-generation NPU, the Rebel R200, to the NVIDIA platform.

The companies will also apply NVIDIA's MGX server rack reference architecture to develop the integrated system. NVLink Fusion extends NVIDIA's high-speed interconnect technology to support integration with AI semiconductors from other companies.

Disaggregated Inference Architecture

The focus of AI competition is increasingly shifting from model development to service deployment and operation. With the growing adoption of agentic AI and a surge in inference workloads, power efficiency and operating costs are becoming as critical as processing performance.

AI inference generally consists of two stages: prefill and decode. During prefill, the system processes and analyzes user input, a computationally intensive task that benefits from high-performance processors such as GPUs. During decode, the system generates responses sequentially, token by token, making fast memory access and data retrieval more important than raw computational power.

To address these different requirements, the three companies plan to implement a disaggregated inference architecture that assigns the two stages to different processors. NVIDIA GPUs will handle the prefill stage, processing and analyzing user queries, while Rebellions' NPUs will manage the decode stage, generating responses.

Performance and Cost Goals

By distributing inference workloads between GPUs and NPUs according to their respective strengths, the companies aim to improve processing performance and power efficiency while reducing the operational costs of AI infrastructure.