Japan's Preferred Networks and SEMIFIVE to develop next-generation AI inference chip with double the memory bandwidth

Japanese AI company Preferred Networks and SEMIFIVE said on Monday they will jointly develop the MN-Core L2000 Series, a next-generation AI inference chip targeting more than twice the memory bandwidth of the previous L1000 Series, with tape-out planned for the first quarter of 2027.
Japanese artificial intelligence company Preferred Networks (PFN) and semiconductor design firm SEMIFIVE said on Monday (November 16) they will jointly develop the MN-Core L2000 Series, a next-generation AI inference chip targeting more than twice the memory bandwidth of the previous L1000 Series, with tape-out — the completion of the chip design and handoff of final design data to the manufacturer — planned for the first quarter of 2027.
How the work is divided
Under the companies' press release, PFN is responsible for the processor architecture and logical design; SEMIFIVE handles design implementation and supply-chain management through its 3D-IC design solution — chief executive Brandon Cho called it the company's “first full-scale 3D-IC initiative” — and the logic chip will be manufactured using Samsung Foundry's platform technology. The L2000 Series stacks DRAM vertically on top of the logic chip, the same 3D-stacking approach as the L1000 Series.
Aimed at generative AI inference
The MN-Core L Series is PFN's family of chips built specifically for generative AI inference, with memory bandwidth — a key determinant of generative AI inference performance — as the design priority. The L1000 Series is scheduled for release in 2027, and the L2000 is its successor. PFN has been developing the MN-Core architecture since 2016, and supercomputers using the first-generation architecture topped the Green500 energy-efficiency ranking three times. The project is supported by Japan's Acquisition, Technology & Logistics Agency under its Innovative Science and Technology Initiative for Security, as a Type S project running from fiscal 2024 through fiscal 2028.
On show at the supercomputing conference this week
PFN and SEMIFIVE will jointly showcase the L2000 Series at PFN's booth at SC26, the supercomputing conference in Chicago, from November 17 to 19. Competition in the inference-chip market is heating up: according to supply-chain analyst Ming-Chi Kuo, Google's TPU has won large orders from OpenAI and Anthropic, which plan to buy about 12 million and 14 million units respectively across 2028 and 2029. Whether PFN can win a share on the strength of its bandwidth advantage will depend on the L2000 meeting its first-quarter 2027 tape-out target.
N環球 
💬 Comments
No comments yet. Be the first!