국내 엣지 AI 반도체가 성공하기 위해선 애플리케이션에 특화된 성능이 필요하다는 주장이 제기됐다. 더불어 SW 풀스택과 관련 생태계 지원 등에 대한 이야기가 나오며 성장 방향성에 대한 인사이트가 공유됐다.

▲2024 ICT Industry Outlook Conference
“Edge AI semiconductors, specialized rather than general-purpose, SW full-stack is competitive”
Opportunities for AI Semiconductor Fabless, Support Based on Peripheral Ecosystem Needed
It was argued that application-specific performance is necessary for domestic edge AI semiconductors to succeed. In addition, discussions were held on SW full stack and related ecosystem support, and insights on growth directions were shared.
At the 2024 ICT Industry Outlook Conference held on the 4th, the need for specialization of edge AI semiconductors and software optimization was raised as the direction of AI semiconductor development.
On this day, the AI semiconductor session was moderated by Cha Cheol-woong, director of the Korea Electronics Technology Institute (KETI), and presentations were given by Kim Byeong-su, director of the SoC Platform KETI Research Center, and Park Jun-young, CEO of UXFactory.
The two axes of artificial intelligence development can be broadly divided into △large-scale AI toward cloud/data centers and △edge computing/edge AI. Although the emergence of ChatGPT has brought attention to hyper-large-scale AI, edge AI is one of the fundamental elements of cutting-edge products applied to areas surrounding us, such as autonomous driving and mobility, CCTV and cameras, robotics, mobile, and unmanned vehicles.
In edge AI environments, efficiency is critical to utilizing limited resources, and targeting solutions specialized for specific applications is required.
■ Edge AI Semiconductor, “Not Universal” 
▲Kim Byeong-su, SoC Platform KETI Research Center Director
Center Director Kim Byeong-su explained, “At the edge, there are so many application areas, including aviation, mobile, factories, and ships, that the hardware (HW) and software (SW) specifications for each environment are different,” and “There is a problem that it is difficult to create a universal solution.”
In fact, Cambricon, which has been attracting attention as a representative NPU venture company in China, has been trying to create a general-purpose NPU structure. CEO Park Jun-young added, “Cambricon tried to create a general-purpose NPU like Intel’s CPU,” and “It tried to develop edge-use cloud semiconductors to achieve its goal of developing AI semiconductors that could be used anywhere, but due to the enormous development costs, it ended up with a net loss of 200 billion won last year.”
Park also cited Horizon Robotics, a Chinese NPU development company, as an example. He explained that the company was founded in 2015 by Yukai, a former Baidu employee, and started out developing general-purpose NPUs, but is now positioned as a company specializing in automotive semiconductor design.
Horizon Robotics is conducting joint development with SAIC Motor, supporting the development of autonomous driving SW full stack and sensors, not just supplying AI semiconductors, but also collaborating on development of solutions across the board.
■ Providing SW solutions is the key to competition 
▲Park Jun-young, CEO of UX Factory
CEO Park drew the line, saying, “AI semiconductors are not an industry that ends with just chip supply like the traditional semiconductor industry in the past,” and emphasized, “Once a chipset is released, it will inevitably require modularization and solution provision while supporting a deep learning framework.”
In fact, in the case of Nvidia, which reigns as the emperor of AI semiconductors with over 90% of the AI semiconductor market, there are considerable strengths in terms of superior HW specifications that come from fine processes and SoC design, but many developers say that it is the finishing touch in Nvidia's exclusive AI programming SW, 'CUDA'.
Park pointed out that, “Apart from the performance measured in terms of high TOPS within the chip, the key is whether the compiler (SW) is well designed to fully draw out the performance of the chip.”
Center Director Kim advised, “Companies developing domestic edge AI semiconductors need to develop SW well as well,” and “They need to focus on the advantages and characteristics that fit the target application and think about optimization such as latency, high performance, and low power.”
He also expressed his opinion that it is necessary to consider the following SW-side optimization methods: pruning, sparsity removal, quantization, preprocessing, workflow optimization, and parallelization, and that research should be conducted in this direction.
■ Edge AI Fabless Opportunity Field, Support for Peripheral Ecosystem Is Essential In server-oriented AI semiconductors, NVIDIA's H100 and A100 are ranked first as the best solution, and in edge-oriented AI semiconductors, NVIDIA's Jetson Nano platform is ranked first as the best solution, and opportunities are opening up for domestic AI semiconductor fabless companies due to market changes.
Especially in edge devices, the need for individualized AI semiconductors is increasing to respond to various application-specific characteristics that general-purpose solutions cannot individually address.
However, as the SoC paradigm becomes the mainstream, fabless startups will have to compete with global giants like Qualcomm.
CEO Park said, “In SoC, we need to use internal chip communication well and discuss how to install IP, but this is not easy for fabless startups.” He added, “In considering government policy, we need to approach it from an ecosystem perspective, and the goal is not simply to make the highest-performance AI semiconductor, but to create a system that can make AI semiconductors well.” He pointed out that “we must also work hard to support technology and the ecosystem.”