[Edaily Reporter Kim Hyun-ah ] #Nota Inc., a company specializing in AI model lightweighting and optimization, has validated a technology that boosts artificial intelligence (AI) response speeds by up to 52% by leveraging both AMD’s GPUs and NPUs. By distributing AI computations across the two types of semiconductors, the company reduced processing time and improved energy efficiency.

Nota Inc. announced on the 9th that it presented the results of its heterogeneous computing optimization technology based on the AMD Ryzen AI Max+ 395 at the “AMD Advancing AI World Tour Seoul 2026” held on the 4th.

According to the company’s internal tests, the total response completion time for the Qwen3-8B model was reduced by 28–52% compared to the existing AMD GPU-NPU hybrid execution environment. Energy consumption per token decreased by 31%, and generation time per token fell by 41–46%.

This technology involves the GPU and NPU sharing the processing of AI computations. The NPU handles “Prefill”—the process of analyzing the content of an input question—while the GPU handles “Decode,” which generates the actual response. The two computing devices are designed to share the same AI model weights, thereby reducing redundant memory usage.

Nota Inc. is also automating the optimization process by applying AI agents to its AI model optimization platform, “NetsPresso.” The AI agents analyze model and hardware characteristics to devise optimization strategies, then measure performance on actual devices to drive improvements. The company is expanding the scope of optimization to include not only the AI models themselves but also the software environments in which they run.

The company also released results from additional tests conducted on various language models. Tests on Llama, Qwen, Phi, and Mistral in the AMD Ryzen AI environment showed a 11–32% reduction in memory usage compared to the existing AMD hybrid execution environment. Compared to using only an NPU, response generation speeds increased by 2.5 to 3.8 times.

Optimizing Everything from AI Models to Runtime Environments… Boosting Semiconductor Utilization Efficiency

Launched in 2015 as a KAIST spin-off, Nota Inc. is a company that lightweight and optimizes AI models so they can run on devices with limited performance and memory, such as smartphones, CCTV cameras, and automobiles. It went public on the KOSDAQ in November 2025.

While AI semiconductor companies manufacture chips that handle computations, Nota Inc. develops software that enables AI to operate efficiently on those chips. Its core business is optimizing AI models for a variety of hardware, including NVIDIA GPUs, Arm-based processors, SamsungElectronics’ mobile semiconductors, and AMD platforms.

Its flagship product is the AI model optimization platform, Netzpresso. It supports model compression, quantization, computational optimization, hardware-specific conversion, and performance verification. This reduces the time AI developers spend manually modifying and testing models for each type of semiconductor.

The company also provides industry-specific on-device AI solutions powered by Netzpresso. Its flagship product, “Nota Inc. Vision Agent AI (NVA),” is an intelligent video surveillance solution that analyzes situations and behaviors captured on video. It goes beyond simply identifying people or objects to understand on-site situations and behavioral contexts, and can search video footage using natural language or generate analytical reports.

The Intelligent Transportation System (ITS) applies AI to existing CCTV cameras to analyze traffic volume and unexpected situations. The DMS Co.,Ltd. detects driver inattention, such as drowsiness or cell phone use. Face Recognition (FR) technology is used for access control and authentication. All of these solutions utilize an on-device approach, running AI directly on the device itself without connecting to a high-performance server.

At Nota Inc., CEO Chae Myeong-su oversees management and business operations, while founder and CTO Kim Tae-ho leads the development of core technologies. The two, who conducted research together at the KAIST Artificial Intelligence Research Institute, expanded the business following CEO Chae’s joining in 2018, after CTO Kim founded the company in 2015. CEO Chae is responsible for the transition to AI model lightweighting and global business expansion, while CTO Kim oversees the development of AI optimization technologies, including Netzpresso.

Nota Inc. has collaborated with domestic and international semiconductor companies such as SamsungElectronics, Arm, FuriosaAI, and Mobilint. Last July, it joined AMD’s Global Robotics Partner Network as one of the key early partners in the Asia-Pacific and Japan (APJ) region. In October, it showcased Netzpresso-based optimization technology on Arm’s AI developer platform.

Building on this AMD platform validation, Nota Inc. plans to expand its business in the AI PC and robotics sectors. Since these devices must run AI on limited memory and power, software that efficiently utilizes various computational resources—such as GPUs and NPUs—is crucial.

Kim Tae-ho, CTO of Nota Inc., stated, “In the era of agentic AI and physical AI, service performance and operational efficiency are determined not only by the performance of the model itself but also by how efficiently various computing resources—such as GPUs and NPUs—are utilized.” He added, “Based on our technology that optimizes everything from the model to the entire inference system, we will expand our collaboration with the global semiconductor ecosystem.”