Artificial intelligence (AI) semiconductor corporations Rebellions will supply server-type Neural Processing Unit (NPU) devices to the closed-circuit (CC)TV control centers of South Gyeongsang Province and Ulsan Metropolitan City.
Rebellions said on the 3rd that it won the server infrastructure projects for the South Gyeongsang Provincial Government and Ulsan City in the "domestic NPU-based public AI CCTV transition" initiative led by the Ministry of Science and ICT and the National Information Society Agency (NIA). Starting this month, it will deliver the high-performance server-type NPU "ATOM-Max."
The AI CCTV control system detects objects such as people or vehicles first at on-site devices like cameras, then reanalyzes the footage at the control center server. The server determines situations such as fires or crimes and generates confidence scores for the analysis and Korean-language explanations.
The ATOM-Max server runs large vision-language models (VLM) of 30B parameters or more at the control center. A VLM is an AI model that analyzes on-screen situations by processing images and text together.
Rebellions said, "Because large VLMs require substantial memory and compute performance, there are limits to conducting precise, simultaneous analysis of multiple video feeds using only small edge semiconductors," adding, "ATOM-Max can process multiple video feeds in parallel when unexpected situations occur at the same time and can operate on closed networks isolated from the external internet."
It also supports "vLLM," software for large language model (LLM) inference. This allows public institutions to adopt Rebellions NPUs without making major changes to their existing AI development environments. It also provides functions to replace or deploy new AI models.
Rebellions recently ran SK Telecom's large-scale AI model "adot X K1 (A.X K1)" on a single unit of its AI server "Rebel Server." The model uses a Mixture of Experts (MoE) architecture with more than 500 billion parameters, and Rebellions said it handled multiple users' requests simultaneously by distributing computation across multiple NPUs.
Park Sung-hyun, CEO of Rebellions, said, "Following the proof-of-concept for a domestic large-scale AI model, we have now applied a server-type NPU to public safety sites," adding, "We will ensure that domestic AI Semiconductor chips are used not just for demonstrations but as infrastructure for public safety."