사진 확대
“Artificial intelligence (AI) factories are the most complex systems on the planet right now. The key is how many tokens you pull out of a fixed power budget.”
Stephane Siru, vice president of Daso System’s high-tech industry, who recently visited Korea, met with Mail Business to diagnose the key tasks of building AI factories.
Daso System is an industrial software company that provides 3D design technology and virtual twins, and among them, the high-tech sector, which is headed by Vice President Siru, covers semiconductors, electronics, telecommunications, robots, data centers, and AI factories.
“Everything that semiconductors go into is high-tech,” he said. “The technology of the Daso system can be designed from shampoo bottles to satellites, and there is high-tech in between.” Daso System divides the industry into 12 major sectors to provide solutions tailored to each industry, and models and simulations are applied from macro units such as AI factories and smart cities to micro units at the material science level.
The most important area emerging in the high-tech sector is AI Factory. The AI Factory, which refers to an infrastructure specialized in AI learning and reasoning, is a concept that has taken a step forward from the existing data center and is a factory that produces ‘token’, the basic unit of AI.
Vice President Siru believes that a new supply chain is being formed in the high-tech industry. It is a structure in which a token is produced in an AI factory due to the input of semiconductors and power, and this token again drives physical AI systems such as robots and autonomous vehicles. “AI Factory is actually a factory that turns electricity into tokens,” he said.
The problem is complexity. Thousands to hundreds of thousands of graphics processing units (GPUs) are put into AI factories, and the demand for cooling, power, and networks that support them is also increasing. On the other hand, the power that each AI factory can secure is limited. “We need to increase the number of tokens that can be generated from a given power budget, that is, token efficiency,” Vice President Siru said. “We need to harmonize computing, network, power, and cooling to combine billions.”We can optimally design tens of billions of dollars worth of infrastructure,” he said.
The solution presented by the Daso system is model-based system engineering (MBSE). It is to apply the method of designing and verifying the entire system in a virtual space before actual production like an aircraft to the AI factory. Based on Daso Systems’ industrial virtual twin technology, Nvidia also uses Daso Systems’ MBSE to design gigawatt (GW) class AI factory references.
In February, the two companies unveiled the cooperation together with Nvidia CEO Jensen Huang and Daso System CEO Pascal Dalrose at Daso System’s annual event, “3D Experience World.” This year, Computex introduced a demo of integrated infrastructure systems with system engineering with Nvidia and power and cooling infrastructure company Vertiv.
Last month, Daso System also released ‘Virtual Companion’, an industrial AI agent that helps design. It is a solution that operates on the 3D Experience platform, the main solution of Daso System, and consists of three types of agents: “Aura,” which supports overall program management, “Leo” for engineering, and “Marie,” which is in charge of science. “Business, engineering, and science have separated the companion because they have different ways of reasoning,” Siru said.
Vice President Siru cited physical law-based design, securing security and sovereignty over intellectual property (IP) accumulated by companies over decades, and tracking possibilities throughout the IP cycle as conditions for industrial AI. For this, vertical integration from simulation to cloud infrastructure is necessary, but at the same time, openness such as interworking between MCPs and agents must also be taken.
Semiconductors are also a representative industry in the high-tech sector of Daso Systems. Although he did not disclose specific domestic customers, Vice President Siru said, “Daso System has developed a technology portfolio that can simulate many elements of semiconductor processes over the past 20 years,” adding, “Companies can use our solutions to model and simulate semiconductor fab processes.”
Meanwhile, Siru cited South Korea as one of the few bases where high-tech is implemented. Regarding Korea’s strengths, he said, “Korea has a deep-rooted engineering culture,” adding, “This engineering culture will be a great asset in the AI supply chain such as fab and AI factory, and Korea is a country ready for a new era.”
[Reporter Jeong Hojun]