by

Japanese boffins want to give AI chips more memory muscle

Japanese chip designer Preferred Networks has teamed up with South Korea’s SEMIFIVE to develop an AI processor promising more than twice the memory bandwidth of its predecessor.

The MN-Core L2000 series is being designed specifically for AI inference, where moving data around can consume more time and energy than performing the calculations themselves.

The companies announced their partnership ahead of the SC26 supercomputing conference in Chicago, with the chip scheduled for tape-out during the first quarter of 2027. There is a slight catch. Its predecessor, the MN-Core L1000, is scheduled for release in 2027, so customers will have to wait before discovering whether either processor lives up to its billing.

According to Preferred Networks the answer to AI’s growing memory problems involves stacking DRAM directly above the processor logic. This three-dimensional arrangement shortens the distance data must travel between memory and processing circuitry, potentially increasing bandwidth while reducing energy consumption.

It is an increasingly attractive proposition as AI models grow larger and the industry finds that adding more processing power is of limited use when chips spend time waiting for data. The MN-Core L2000 builds on Preferred Networks’ proprietary accelerator architecture, which the Japanese outfit has been developing since 2016.

The company has some experience in this area. Supercomputers using its original MN-Core processors have topped the Green500 ranking of the world’s most energy-efficient supercomputers three times. The new L-series processors concentrate on inference, the process of running trained AI models, with particular attention to the memory requirements of generative AI and autonomous agents.

Preferred Networks claims its vertically stacked architecture can deliver substantially greater memory bandwidth than conventional AI processors, although independent performance figures for the L2000 remain some distance away. Samsung Foundry has joined the project and will manufacture the processor logic using its semiconductor technology.

The Korean chipmaker is providing design infrastructure for the three-dimensional arrangement, including interfaces connecting the stacked chips and techniques for managing power delivery. SEMIFIVE will handle the physical chip implementation and manufacturing arrangements, drawing on its experience with Samsung’s foundry processes. Preferred Networks will remain responsible for the processor architecture and logical design.

The arrangement gives Samsung another opportunity to establish itself in advanced semiconductor packaging, an increasingly important area of competition with TSMC. However, squeezing processor logic and DRAM into the same vertical structure creates some rather unpleasant engineering problems. Heat becomes particularly troublesome when multiple active semiconductor layers occupy a small space. Manufacturing defects can prove expensive when several chips must work together inside one package.

Preferred Networks co-founder and chief executive Daisuke Okanohara acknowledged the difficulties.

“Developing AI chips with a 3D-stacked structure presents more complex technical challenges than conventional chip development, particularly in areas such as power consumption, heat dissipation, and manufacturing yield,” Okanohara said.

These are problems semiconductor manufacturers have wrestled with for years, and announcing a partnership has yet to become an established way to solve them.

SEMIFIVE co-founder and chief executive Brandon Cho described the project as the company’s first full-scale 3D-IC initiative.

The technology promises improved data transfer efficiency, although converting an attractive architecture into commercially viable silicon will require considerable work. Japan’s Acquisition, Technology & Logistics Agency is supporting the development programme under a research initiative running from 2024 to 2028.

 

 

Latest articles

Share

Featured articles

Hot topics

No results found.

Latest reviews