DeepX mass-produces edge AI semiconductors with high-efficiency neural network processing units (NPUs)
Participation in chip development by Apple and Cisco… Expectation of future widespread adoption of NPU-based chips
NPU hardware-software optimization… Maximize efficiency of specific AI algorithms
Maintains high performance while realizing low power consumption and low heat generation
Due to the remarkable development and evolution of artificial intelligence (AI), the popularization of AI is rapidly progressing. Individuals receive AI results processed from data center servers and use them for work. AI semiconductors have also begun to be applied to individual electronic devices such as PCs, mobile phones, closed-circuit (CC) TVs, and home appliances. As the need for not only people but also electronic devices to immediately process information on-site increases, the use of AI semiconductors is increasing.
DeepX is a leading startup in the design of AI semiconductors used in electronic devices. It is called an edge AI semiconductor. It comes from the meaning of processing data at the edge of the network, close to the data (field), rather than at the central data center on the network.
DeepX manufactures neural network processing units (NPUs). This is a method that allows specific AI algorithms to be processed with high efficiency by further subdividing the AI algorithm than the graphics processing unit (GPU) that NVIDIA focuses on. It was founded in 2018 by CEO Nokwon Kim (46), who designed chips for global companies such as Apple and Broadcom. CEO Kim, who met at his office in Pangyo New Town, Seongnam-si, Gyeonggi-do on the 18th, said, “Just as it is difficult for us to imagine life without Wi-Fi now, in the near future, humans will depend on electronic devices equipped with AI semiconductors.” “I started the business because I wanted to participate in an opportunity to change the lives of mankind,” he said.
● “If the existing GPU costs 100 won, ours costs 5 won.”
![I gave up 6.5 billion won and became independent from Apple… “Incorporating AI chips into all devices”[허진석의 톡톡 스타트업]](https://dimg.donga.com/wps/NEWS/IMAGE/2024/10/25/130298223.1.jpg "I gave up 6.5 billion won and became independent from Apple… “Incorporating AI chips into all devices”[허진석의 톡톡 스타트업]")
DeepX’s product line currently consists of three types (DX-M1, DX-V3, and DX-H1). According to DeepX, the DX-M1 chip is excellent for processing AI algorithms that recognize specific objects in images at low power and low cost. It also works on battery-powered devices. At the Semiconductor Expo held at COEX in Seoul until the 25th, DeepX did not melt butter as the temperature only rose to 35.5 degrees Celsius despite running the latest object recognition algorithm. On the other hand, chips from global AI semiconductor companies that are being compared rose to 60.7 degrees. CEO Kim said, “It can be applied to numerous types of devices that process video information, such as security CCTVs and process inspection cameras,” and added, “Compared to existing GPUs, the price and power consumption are only 1/20th.” did it This single chip can process video information from 16 CCTVs (channels) at a speed of 30 frames per second (FPS). It is produced at Samsung Electronics using a 5nm (nanometer, 1nm is 1 billionth of a meter) process, and DeepX announced that it is focusing on increasing the mass production yield.
DX-V3 is a chip that provides vision to autonomous driving and robots that require processing of camera and three-dimensional (3D) sensor signals. TSMC plans to produce it using the 12nm process. The DX-H1 chip is for AI servers and provides high performance with much less power and cost than existing general-purpose graphics processing units (GPGPUs). Samsung Electronics plans to produce it using the 5nm process.
DeepX’s design technology played a large role in achieving high performance, low power consumption, and low heat generation. DeepX possesses the original technology that produces the world’s highest effective AI computing performance ratio (FPS/TOPS) and overall performance ratio (FPS/W). Based on this technology, more than 300 patents have been applied for, and more than 70 patents have been registered. It is a technology that maximizes image processing efficiency when each has the same computational performance, and is a technology that can produce the highest computational performance at the same power.
CEO Kim said, “It contains technologies learned while designing chips at the world’s best companies,” and added, “We optimized the software-driven and hardware-driven parts so that specific AI algorithms can achieve the highest efficiency.”
● Turn Nobel Prize winner’s theory into a chip
CEO Kim, who received a bachelor’s and master’s degree in electronic engineering from Korea University, worked at a government-funded research institute in Korea before going to the United States to study. He said, “I heard that there are many best people in the electronics field in the United States, and I wanted to see for myself how talented they are.” While working on a doctoral degree in electrical engineering at the University of California, Los Angeles (UCLA), I went to the IBM Watson Lab as a visiting researcher in 2008 and had the opportunity to start a business. Watson Research Institute was in charge of a project to implement a paper introducing the concept of deep learning published in 2006 by Professor Geoffrey Hinton of the University of Toronto, Canada, into a chip. The term deep learning was not used at this time. CEO Kim devised a new neural network processing structural design and demonstrated performance that was 100 times faster than when using a central processing unit (CPU). When he returned to UCLA, he continued his research and published a related paper in the Journal of the World Institute of Electrical Engineers. CEO Kim said, “I first saw the possibility and value of predicting future situations by analyzing data with patterns in 2010, before the term deep learning even appeared.” Professor Hinton received this year’s Nobel Prize in Physics for his contribution to the development of AI.
CEO Kim designed semiconductors at Cisco Systems, a global communications equipment company, for three years from 2011, and worked as a senior researcher at Apple from 2014 to 2017, participating in the development of the iPhone chip processor.
CEO Kim said, “In order to leave Apple, I had to give up a stock bonus worth about $5 million, which could have been monetized over four years,” and “Nevertheless, I could see a future where NPU-based chips would be installed all over the world, and that future was promising.” “I thought that if I had the talent to advance my business, it was right to choose that talent, so I started a business,” he recalled.
● “We are in the process of building a distribution network for export”
CEO Kim’s heart still seemed to be racing when he pictured the future of edge AI semiconductors. He said, “This was stated in the data I saw when I received my doctorate in 2011 and worked at Cisco Systems. Twenty years ago, 5 billion devices around the world were connected to the Internet, and by 2020, there would be 70 billion devices connected to the Internet. Actually, that much is connected. If so many devices are equipped with AI semiconductors similar to human intelligence, what kind of world will that future be like? “A world of another dimension will open up,” he said, unable to hide his excitement.
He predicted that the need for edge AI semiconductors will grow, citing the three major trends in the AI industry: autonomy, unmanned technology, and personalization. A representative example of autonomy is self-driving cars, which require edge AI semiconductors to achieve high-performance, low-power performance without being affected by the communication environment. This means that personal information can be more safely protected by keeping information containing personal preferences and body information confined to the individual’s mobile phone.
Edge AI semiconductors will be used by various companies at home and abroad. CEO Kim said, “The response from the global market has been good, so I am currently busy with business trips to establish a global distribution network in each continent.” We are pioneering the market with a local subsidiary in Silicon Valley, USA. DeepX plans to provide innovative solutions in various global industries, including smart cities, surveillance systems, and smart factories.
CEO Kim said, “Our goal is to implement ‘intelligence that is cooler than humans.’ “It shows better performance with less energy than the human brain uses.”
