"Korea to Become an Exporter of Physical AI"... DeepX Unveils Roadmap for 'DX-M2' Featuring Samsung 2nm Technology

DeepX, which has declared its transformation from a fabless company to a physical AI enterprise, will begin mass production of its next-generation chip, the 'DX-M2', in 2027. This chip, the world's first to apply Samsung Electronics' 2-nanometer process, forms the basis of DeepX's strategy to expand generative AI beyond data center infrastructure and into real-world physical devices such as robots.

On the 14th, Kim Nokwon, CEO of DeepX, held a media briefing at the company's Pangyo headquarters and stated, "In the future, companies aiming to create artificial intelligence (AI) will turn to Nvidia, while those seeking to utilize AI will choose DeepX products." He also unveiled a three-stage full-stack strategy that connects chips, hardware platforms, and software ecosystems, as well as the development roadmap for the DX-M2.

Kim Nokwon, CEO of DeepX, held a media briefing at the Pangyo headquarters on the 14th. Photo by Jinyoung Kim

원본보기 아이콘

The DX-M2 is an 'edge AI chip' designed to perform AI computations directly on endpoint devices such as automobiles, robots, and home appliances, rather than relying on cloud servers. It is the first in the industry to apply Samsung Foundry's 2-nanometer (nm, one-billionth of a meter) process, with the goal of delivering ultra-low power consumption under 5W and achieving performance of up to 80 TOPS (80 trillion operations per second).

With the DX-M2, DeepX envisions powering generative AI models with hundreds of billions to trillions of parameters on battery-operated devices. This means bringing advanced capabilities-such as higher inference performance, complex command processing, and multimodal recognition-that were previously possible only in data centers, to ultra-low-power devices.

DeepX is recognized for meeting all three critical conditions that determine the success of physical AI chips: power efficiency, price competitiveness, and absolute heat generation. Its previous model, the DX-M1, achieves ultra-low power consumption averaging 2-3W, delivers 20 times greater power efficiency than Nvidia's on-device GPU 'Jetson Orin' at one-tenth the price, and features a die size approximately one-fourth that of competitors, allowing production of more than four times as many chips per wafer. The initial yield rate for Samsung's 5-nanometer process is 91%, significantly outperforming the industry average yield of 50-80%. Notably, in thermal tests where butter was placed on competing chips and melted, the DeepX chip maintained a temperature below body heat, making headlines worldwide.

DeepX DX-M1 Chip. Photo by Jinyoung Kim

원본보기 아이콘

Such competitiveness has led to a surge in interest from global clients. The number of mass production contracts increased from two in December last year to over 30 by March this year. In Korea, an AI computing solution co-developed with Hyundai Motor Group Robotics Lab has completed mass production validation and will be applied to delivery robot 'DAL-e' and mobility platform 'MobED', with mass production set to begin at the end of this year. In China, DeepX was selected as Baidu's AI chip supplier and recently secured an initial order for 40,000 units. Annual sales this year are estimated at 40 million dollars (about 5.93 billion won).

DeepX is also accelerating efforts to establish standards for software infrastructure. The company plans to complete its alternative solution 'DX-Newton' to Nvidia's 'Isaac', a software platform for physical AI systems, within this year. This will enable developers to maintain the familiar Isaac ROS (Robot Operating System)-based development flow, while switching only the AI inference acceleration segment to DeepX chips. Robot developers can utilize the familiar Isaac ROS environment during the initial development phase and transition to the DeepX platform for commercialization and mass production, thereby securing both power efficiency and cost competitiveness.

CEO Kim stated, "While Nvidia paved the way for AI with GPUs and CUDA, DeepX will pave the way for physical AI with DX-M1, DX-M2, and DX-Newton," adding, "Just as TSMC built today's Taiwan, DeepX will build Korea's physical AI semiconductor industry."

© The Asia Business Daily(www.asiae.co.kr). All rights reserved.