14:40 – 15:00
Memory Bandwidth is All You Need
(Language: Japanese, English interpretation provided)
As generative AI continues to evolve, the importance of inference computing is poised to grow steadily. In this session, the speaker Takahiro Ogura will introduce the 3D-stacked memory structure, a technological breakthrough for AI inference chips, and demonstrate its potential for social transformation through AI agents and edge AI. Ogura will also share his company Preferred Networks’s vision for democratization of AI and creation of new paradigms through their proprietary chip design, ecosystem development and industry collaborations.
Takahiro Ogura
Preferred Networks
Takahiro Ogura is a seasoned leader in high-performance computing with over 20 years of experience. He currently heads the AI Computing Division at Preferred Networks (PFN), overseeing the company’s critical computing infrastructure technologies including the MN-Core™ series of AI semiconductors and in-house supercomputers. Prior to joining PFN, he served as the after-sales service manager for domestic high-performance computing clients at Fujitsu, where he began his career as a systems engineer working on supercomputer systems projects. A notable career highlight was his instrumental role in the design and development of the Fugaku (Post-K) supercomputer at Japan’s national research and development agency RIKEN. Ogura received a Master’s degree from Tsukuba University’s Systems and Engineering program in 2005.
Preferred Networks
Company Profile
Preferred Networks (PFN) is a leading Japanese AI technology company known for its vertically integrated development of cutting-edge technologies—from custom AI chips, generative AI foundation models to real-world AI solutions and products. PFN plays a key role in advancing Japan’s AI sovereignty and has earned a strong domestic reputation. PFN’s AI solutions span across diverse industries including manufacturing, materials, energy, pharmaceuticals, healthcare, finance, retail, and entertainment. The company is currently developing a novel AI inference chip, leveraging its highly energy-efficient MN-Core™ logic and a new memory technology, with a release targeted for early 2027.
Company Products & Services
Preferred Networks (PFN) develops cutting-edge technologies across the AI value chain from semiconductors, supercomputers, generative AI foundation models to AI solutions and products:
MN-Core™ — a series of highly energy-efficient AI chips, which made a debut topping the Green500 list of the world’s most energy-efficient supercomputers three times in 2020/21. PFN is currently developing a new line of MN-Core chips focused on AI inference, set for release in early 2027.
Preferred Computing Platform™ (PFCP™︎) — a cloud-based service that provides users with access to the computing power of MN-Core 2. PFN has also announced a joint venture with Mitsubishi Corporation and IIJ aiming to launch an MN-Core™-based IaaS commercial service in 2026.
PLaMo™ — a series of Large Language Models developed fully from scratch by PFN, delivering top-tier performance in Japanese-language capabilities. These models power a wide range of AI applications in sectors such as finance, healthcare, and the public sector.
Matlantis™ — a cloud-based, high-speed universal atomistic simulator. Powered by a machine learning interatomic potential (MLIP), Matlantis significantly accelerates the simulation workflow in new materials discovery across industries such as semiconductor, automotive, and energy, by bypassing traditional heavy electronic state calculations.
