Innovation & Tech

Preferred Networks Doubles Down on Memory Bandwidth with MN-Core L2000

PFN and SEMIFIVE will build the MN-Core L2000 inference chip, targeting over twice the L1000's memory bandwidth, with Samsung Foundry manufacturing the logic die.

By Grace Kim

3 min read

Preferred Networks Targets Faster AI Inference with MN-Core L2000
Preferred Networks Targets Faster AI Inference with MN-Core L2000ElasticComputeFarm / pixabay

What's News

  • PFN and SEMIFIVE announced the MN-Core L2000 Series on November 16, 2026, targeting more than twice the memory bandwidth of the L1000 Series.
  • The L2000 tape-out is planned for Q1 2027, ahead of the L1000's scheduled 2027 release.
  • Samsung Foundry will manufacture the logic chip and supply 3D-IC design infrastructure.
  • Japan's Acquisition, Technology & Logistics Agency supports development via a Type S project from fiscal 2024 through fiscal 2028.
  • PFN says first-generation MN-Core supercomputers topped the Green500 energy-efficiency ranking three times.

Preferred Networks and SEMIFIVE announced on November 16, 2026, that they are collaborating on the MN-Core L2000 Series, AI inference chips designed to deliver more than twice the memory bandwidth of PFN's L1000 Series.

The L2000 succeeds the L1000, which is scheduled for release in 2027. Work on the newer series has already started, with tape-out — the completion of chip design and handoff of final data to the manufacturer — planned for the first quarter of the same year.

Why is memory bandwidth the bottleneck?

PFN's November 2024 L1000 announcement laid out the design logic. Training generative AI demands enormous computing resources, but inference asks something different: the processor must read hundreds of gigabytes of data every time it generates a unit of text, images or other output. The result is that limited memory bandwidth can throttle inference even though it needs fewer computing resources than training.

Both L-series designs stack DRAM vertically with logic chips built on MN-Core. PFN says this 3D structure achieves substantially higher memory bandwidth than GPUs and other AI processors, supporting faster inference and better power efficiency for applications including AI agents. The L1000 design also pairs densely mounted arithmetic units with distributed memory, both controlled by software to limit power consumption and heat.

The pedigree is real. PFN has been developing MN-Core since 2016, and the company says supercomputers using the first-generation architecture topped the Green500 energy-efficiency ranking three times. One caveat: that track record belongs to the earlier architecture. The L2000's bandwidth improvement remains a development target, not a shipped result.

The takeaway: the L2000 bets that stacking memory on logic beats raw compute for the inference workloads now dominating AI deployments.

Who does what in the three-way build?

The development splits three ways:

  • Preferred Networks owns the L2000's processor architecture and logical design.
  • SEMIFIVE, a turnkey provider for custom semiconductors, handles design implementation and supply-chain management through its 3D-IC design solution — spanning DRAM supply, volume manufacturing, advanced packaging and testing.
  • Samsung Foundry manufactures the logic chip on its platform technology and supplies 3D-IC design infrastructure, including optimized I/O interfaces that connect the stacked dies and design methods for optimizing the power structure.

SEMIFIVE CEO Brandon Cho called the project the company's "first full-scale 3D-IC initiative." The assignment puts SEMIFIVE squarely between PFN's processor design and the specialized memory, packaging and manufacturing operations required to implement it.

What makes 3D-stacked chips hard?

Plenty, according to PFN's own leadership. Co-founder and CEO Daisuke Okanohara identified power consumption, heat dissipation and manufacturing yield as the challenges that make 3D-stacked AI chips more complex to develop than conventional chips.

Those issues matter more here than usual. The entire performance approach depends on integrating logic and memory vertically rather than fabricating a standalone logic device, so a failure in any one of those three areas undercuts the architecture's core advantage.

In short: the harder the stacking, the harder the yield and heat problems — and the L2000 stacks aggressively.

Where does the money come from?

The L2000 development is supported by Japan's Acquisition, Technology & Logistics Agency under its Innovative Science and Technology Initiative for Security program. The support arrives through a Type S project running from fiscal 2024 through fiscal 2028.

The partnership also has a public date on the calendar. PFN and SEMIFIVE plan to jointly showcase the L2000 at PFN's booth during SC26 in Chicago, running November 17–19, 2026.

The deal gives PFN an implementation and manufacturing path for its bandwidth-focused architecture and widens SEMIFIVE's footprint in 3D-IC work. The execution test now is whether the companies can carry the design's bandwidth and efficiency targets into volume production while holding down the power, heat and yield problems Okanohara named.

Source: unite.ai

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,178 articles

Related articles

« Previous articleNext article »