Preferred Networks Targets Faster AI Inference with MN-Core L2000

Unite.AI
Read full story· 4 min
Preferred Networks and SEMIFIVE are collaborating on the MN-Core L2000 AI inference chip series, aiming to double the memory bandwidth of the L1000 series to accelerate AI inference. The L2000 uses 3D-stacked DRAM and logic chips for higher bandwidth and power efficiency, with tape-out planned for early 2027 and manufacturing by Samsung Foundry.

What happened

Preferred Networks and SEMIFIVE announced a partnership on November 16, 2026 to develop the MN-Core L2000 Series, a new line of AI inference chips. The L2000 is designed to offer more than twice the memory bandwidth of the previous L1000 Series, using 3D-stacked DRAM and logic chips. Tape-out, which marks the completion of the chip design, is scheduled for the first quarter of 2027, with manufacturing to be handled by Samsung Foundry. The project is supported by Japan’s Acquisition, Technology & Logistics Agency and will be showcased at SC26 in Chicago in November 2026.

Key details

  • L2000 aims for over double the memory bandwidth of the L1000 Series.
  • Both L1000 and L2000 use vertically stacked DRAM and logic chips.
  • SEMIFIVE manages design implementation, supply chain, and advanced packaging.
  • Samsung Foundry will manufacture the logic chip and provide 3D-IC design infrastructure.
  • Development is funded under Japan’s Innovative Science and Technology Initiative for Security, running through fiscal 2028.

Why it matters

AI companies and developers who rely on fast inference for applications such as AI agents will have access to chips with higher memory bandwidth and improved power efficiency. The collaboration gives Preferred Networks a clear path from design to manufacturing, while SEMIFIVE expands its role in 3D-IC development. The new chip could reduce bottlenecks in AI inference workloads, but success depends on overcoming challenges related to power, heat, and manufacturing yield.


More in Chips & Compute