“The AI chip market is gradually moving from the server to the edge. I think the demand at the edge will continue to grow to further enhance the real-time operation of AI applications.”

The environment surrounding domestic fabless companies is not easy. Even in the barren land of system semiconductors, domestic fabless companies are sparing no effort in developing cutting-edge products that respond to market trends.

Kim Hyun-do, head of NeoWine’s new business development division, said NeoWine is looking for new food in the edge AI market. After expanding its scope from existing security semiconductors to edge AI chip semiconductors and AI software solutions, it is introducing an end-to-end service that integrates hardware and software to the market.

On May 11, our magazine met with Hyundo Kim, Head of New Business Development at NeoWine, at the 2023 International Artificial Intelligence Exhibition and heard about NeoWine's edge AI solutions.

■ Are you developing AI solutions as well as semiconductors?

In addition to hardware (HW) chip design, we also develop software (SW). With support from the National IT Industry Promotion Agency, AI SW development also began in 2021.

■ Your original specialty is security and encryption chips. I am curious about what led you to enter the AI field.

It is true that Neowine's main product is security semiconductors and that it has a long history in the related field. Usually, when developing a product, not only HW but also SW is developed. At this time, as many AI algorithms are applied to the SW side, we have developed AI solutions and launched services.

For example, if you develop an around-view parking system for a vehicle and export it overseas without a security chip, you may be exposed to the risk of hacking. NeoWine has a solution that responds to related risks by applying a security chip next to the MCU mounted on the system hardware. The security chip acts like a lock and blocks incorrect access.

From these anti-copying and encryption security chips, it has expanded to IoT security and data encryption, and is currently developing into AI semiconductors. It has completed two years of semiconductor IP development through government-supported R&D, and is currently in the chip manufacturing stage.

■ Please explain the AI solutions currently being showcased at this AI exhibition.

We are developing and introducing to the market an image search and interpretation solution using artificial intelligence as an illegal copy interpretation system. It was developed as a solution applicable to the Customs Service, etc. to interpret suspicious items and determine whether design patents are infringed.

We are also developing and commercializing AI models for flame and smoke detection applied in AI CCTV video security systems. A system where AI models are applied can be built to analyze and process collected images by installing an AI system on an edge device camera or applying an AI system on an edge server between the edge and server sides.

We are currently working on developing chips that can be installed in CCTV cameras, and we are also focusing on developing AI models that can be installed on top of the chips to commercialize and advance them.

■ What areas of technology and market do your current products focus on to differentiate themselves?

Domestic semiconductor manufacturers are mainly jumping into the server market. NeoWine is focusing on the edge and introducing products.

Looking at video cameras, we are seeing demands for miniaturized chip sizes and low-power specifications. Accordingly, we are focusing on product development by focusing on the IoT and edge device markets by commercializing edge AI chips that perform real-time calculations and processing.

Likewise, compared to ST and TI, which manufacture edge chips, the company sees its strengths in that it can provide technical support to customers with small quantities.

Currently, we have developed AI semiconductor design IP for real-time high-speed processing of edge AI models and compiler SW required for optimizing AI models installed on AI chips. In addition to simply supplying products, we can also provide IP and SW customization services tailored to the customer environment, and we also provide technology consulting services.

■ Please explain the features of the developed AI chip product.

The NPU IP consists of an NPU core consisting of 1,024 MAC array processors and multiple operators. It supports numerous operations such as operators Conv3D, Conv1x1, Depthwise, Convolution, and Matmul, and its basic performance is at the 4TOPS level.

The compiler SW is structured as ONNX and can accommodate AI models developed in various frameworks such as TensorFlow and PyTorch, and has the following features: AI model quantization, lightweight through operator fusion, and a compiler for high speed.

■ What do you think is the most difficult point in commercializing a product in the domestic fabless industry?

I think cost is the biggest problem. Since a lot of development resources such as manpower and resources are invested in product development, the development cost issue is the most sensitive. Large companies have a wide range of business sectors and a diversified sales structure, so they can generate stable profits and secure related funds, but this is often not the case for small and medium-sized companies.

The three elements of time, cost, and manpower are in a trade-off relationship, and if R&D costs are not involved, it will take a considerable amount of time and difficulty to develop a high-quality product. I expect that the situation will improve if domestic semiconductor and AI talent training businesses are activated and manpower support is provided in such areas in the future.