Artificial Intelligence (AI) is constantly evolving to make our lives better. NEC has set its sights on building a supercomputer for AI research that will be the largest ever built by a Japanese company and make it possible to reduce AI learning time to 1/4,600. But why do we need a large-scale AI research supercomputer now, and what can be done at higher speeds? We spoke with Akio Yamada, the Senior Vice President responsible for overall research and development at NEC, and Takatoshi Kitano, overseeing the development of this AI research supercomputer, to answer these and other questions.

The significance of NEC building its supercomputer: The responsibility and trust involved in being entrusted with data

──NEC is building an AI research supercomputer featuring the greatest computational capacity of any company in Japan, with full operation slated to begin in March 2023. What is the aim of this endeavor?

Senior Vice President Akio Yamada: The term digital twin is one heard frequently in recent years. It is an approach used to predict the future by modeling various phenomena that occur in the real world. We have recently seen a need to go further and propose optimal actions and solutions based on future predictions. For humans, it is the actual process of decision-making itself that we perform unconsciously, but underlying it is the knowledge we have learned for several decades since birth.

To utilize AI, much knowledge must be learned as humans do. The difficulty when trying to increase AI's popularity is not its utilization but rather its training process, which requires tremendous computational capacity. AI research supercomputers are ultra-fast computational systems that streamline the learning process, dramatically reducing the time spent on AI development.

Over the past several years, the utilization of AI has expanded to include areas such as the environment, urban infrastructure, and payment services. Even one AI project requires a long time for learning (i.e., massive computations) during the building process. Still, there are limits to approaches that rely solely on the brains of researchers to accelerate research and development when AI needs are expanding as rapidly as they are now. It can be said that the creation of this AI research supercomputer is necessary for the swift resolution of advanced social issues.

──NEC invested billions of yen to make a supercomputer of this scale a reality. It is incomparable to the typical amount invested by NEC for AI research. What is the significance of a company having its development environment?

We decided to use our own facilities rather than outside facilities to ensure the trust of our customers who would be entrusting their data to us.

Most data required for AI training is confidential corporate information and other data that must be handled with extreme care. By building its own AI research supercomputer, NEC can take data containing sensitive information in its robust security environment. It is this trust in the security of data, for which NEC itself will shoulder complete responsibility and carry out development, that will be crucial for the expansion of the co-creation of AI in the future.

──NEC's AI research system has now become one of the largest in the world.

NEC is researching AI at all its research laboratories in and outside Japan and employs hundreds of AI researchers. We also now rank eighth globally for paper acceptance at top-ranking conferences on machine learning for AI. It can also be said that this project came about because of NEC's desire to build an AI research supercomputer at a scale that can tolerate constant use by several hundred researchers to allow NEC's world-class researchers to take full advantage of their abilities. Creating the largest scale supercomputer of any domestic company was not our goal—it was merely the outcome.

Customers & Partners x Researchers x AI research supercomputer = Social Value Creation

──What social value can NEC provide by accelerating AI research and development?

Senior Vice President Akio Yamada: One example of a critical initiative is Green Transformation (GX) for achieving carbon neutrality.

Take taxi management, for instance. Efficient management necessitates AI to optimize vehicle dispatch planning based on past and present data such as traffic conditions, city congestion, and weather conditions. But optimization alone, such as showing where customers are located, does not constitute GX. Regarding electric vehicles (EVs), charging points and timing must also be considered, and suggestions must be based on efficiency and environmental impact on society.

Even concerning transportation alone, optimizing complex systems at a scale and speed beyond existing controls is crucial for city-wide traffic control, including trains, buses, and private vehicles. The realization of AI that satisfies these needs transcending human knowledge, will create a sustainable society. This AI research supercomputer will enable us to provide a system capable of accommodating a more comprehensive range of requests for AI construction than ever before.

──The key will be to co-create with our customers and partners.

With the acceleration made possible by our AI research supercomputer, trial-and-error by AI, which once took several months, can be completed within a single day. While thus far, we have only been able to accept a few consultations per year from customers and partners, we will now be able to respond to far more of these consultations. It will also enable us to rapidly increase the number of real-world examples of social implementation.

Combining an AI research supercomputer, NEC's outstanding AI researchers, and collaboration with our customers and partners will allow us to create social value. That is the actual value of NEC, and I am convinced that NEC will become a front-runner in AI.

Accelerating social value creation: Computational power will be a crucial source of competitiveness in the AI era

──Just how much will the performance of the AI research supercomputer improve?

Takatoshi Kitano (head of development): This AI research supercomputer has 928 high-end NVIDIA GPUs (image processing semiconductors). It means that what once took 4,600 hours for AI learning on a single high-end GPU can be done in just one hour with this AI research supercomputer.

AI development requires thousands of trial-and-error cycles as changes are made to algorithms, data, and more. And it is the shortening of these trial-and-error cycles that directly impacts the competitiveness of AI. In addition, learning on some of the latest large-scale AI supercomputers can take 355 years on a single GPU. With NEC's AI research supercomputer, large-scale AI models can also learn, and researchers can engage in cutting-edge AI research with no constraints. AI is at the core of the digital revolution, and computational power will be a crucial source of competitiveness in the AI era.

Possessing the greatest computational capacity of any company in Japan means that NEC can continue to create cutting-edge AI solutions. I believe this to be very significant for our customers, partners, and researchers.

──The main takeaway here seems to be that it's not just a competition to achieve superior performance.

Regarding AI research supercomputers, good specs don't necessarily equate to a good computer. It is only a good computer if you have sound problems and outstanding researchers use it. AI research supercomputers are only helpful when we have customers with advanced issues and talented researchers who need to be solved.

Innovation cannot be generated by NEC alone. We can co-create social value when we bring together customers and talented researchers in the best environment of any domestic company in Japan. By becoming an AI center of excellence, we hope to generate innovation and accelerate social value creation.

Recommended Articles

NEC Stories Top