Preferred Networks Bets on Stacked Memory for Faster AI Inference
PFN and SEMIFIVE announced the MN-Core L2000 on November 16, 2026, targeting over twice the L1000's memory bandwidth for AI inference, with Samsung Foundry building the logic chip.

Why it matters
- PFN and SEMIFIVE announced the MN-Core L2000 collaboration on November 16, 2026, targeting more than twice the memory bandwidth of the L1000 Series.
- The L1000 ships in 2027; the L2000's tape-out is planned for Q1 2027.
- Samsung Foundry will manufacture the logic chip and provide 3D-IC design infrastructure.
- Development is backed by Japan's ATLA under a Type S project running fiscal 2024 through fiscal 2028.
- SEMIFIVE CEO Brandon Cho called it the company's 'first full-scale 3D-IC initiative.'
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 a chip that has not even shipped yet: the L1000 is scheduled for release in 2027, and work on the newer series is already underway, with tape-out planned for the first quarter of the same year.
Tape-out, for readers who don't live in semiconductor land, is the moment a chip design is finished and the final design data goes to the manufacturer. PFN is planning the L2000's tape-out before its predecessor reaches the market. That schedule tells you where the company thinks the inference market is heading, and how fast it believes it has to move.
Why does memory bandwidth matter so much for inference?
The reasoning comes from PFN's own November 2024 announcement of the L1000. Training generative AI demands substantial computing resources, but inference works differently: the processor has to read hundreds of gigabytes of data whenever it generates a unit of text, images, or other output. Limited memory bandwidth can therefore become a bottleneck even though inference needs fewer computing resources than training.
That's the mechanism behind the whole product line. If the chip can't feed data to its compute units fast enough, those units sit idle, and every generated token costs more time and power than it should.
Both L-series designs attack this by stacking DRAM vertically with logic chips based 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 likewise combines densely mounted arithmetic units with distributed memory, both controlled by software to limit power consumption and heat.
The takeaway: PFN isn't competing on raw compute. It's competing on how quickly a chip can move data, because that's what inference actually spends its time doing.
How credible is the engineering pedigree?
PFN has been developing MN-Core since 2016, and says supercomputers using the first-generation architecture topped the Green500 energy-efficiency ranking three times. That's a real track record in a specific discipline — squeezing more computation out of each watt.
One caveat belongs here, though. That Green500 history applies to the earlier architecture. The L2000's bandwidth improvement remains a development target, not a demonstrated result. Past efficiency awards don't guarantee that a new 3D-stacked design hits its goals.
Who does what in the partnership?
Three companies split the work, and the division of labor is worth understanding because it shapes who carries the execution risk.
- Preferred Networks is responsible for 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. Its coordination spans DRAM supply, volume manufacturing, advanced packaging, and testing.
- Samsung Foundry manufactures the logic chip on its platform technology. Samsung Foundry is also providing 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 described the project as the company's "first full-scale 3D-IC initiative." The assignment places SEMIFIVE between PFN's processor design and the specialized memory, packaging, and manufacturing operations needed to implement it — a middleman position with a lot of coordination riding on it.
In plain terms: PFN draws the chip, Samsung builds the logic, and SEMIFIVE makes the whole stack fit together and actually get manufactured.
What makes 3D-stacked chips hard?
PFN 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. These aren't footnotes. They're the central engineering risk of the project.
The reason is structural. The performance approach depends on integrating logic and memory vertically rather than fabricating a standalone logic device. Stacking dies means heat has fewer places to go, power delivery becomes harder to engineer, and a defect in one layer can sink a whole assembled package. Yield problems in 3D-IC production are more expensive than in conventional chips precisely because so much finished work gets thrown away when one layer fails.
That's why Samsung's contribution isn't just manufacturing capacity. Its 3D-IC design infrastructure — the optimized I/O interfaces between stacked dies and the power-structure design methods — exists to manage exactly the problems Okanohara named.
Is there government money behind this?
Yes, and it signals how Japan views AI silicon as a strategic capability. The L2000 development is supported by Japan's Acquisition, Technology & Logistics Agency under its Innovative Science and Technology Initiative for Security program. The support comes through a Type S project running from fiscal 2024 through fiscal 2028.
A five-year defense-linked funding commitment tells you the L2000 isn't framed purely as a commercial product inside Japan. It sits in a national effort to build domestic AI compute capability, at a moment when most of the world's inference hardware comes from a small number of foreign suppliers.
What happens next?
Two dates matter. The L1000 ships in 2027, and the L2000's tape-out is planned for the first quarter of that same year. Before any of that, PFN and SEMIFIVE plan to jointly showcase the L2000 at PFN's booth during SC26 in Chicago, from November 17 to 19, 2026.
For PFN, the partnership supplies what a fabless AI startup typically lacks: an implementation and manufacturing path for its bandwidth-focused architecture. For SEMIFIVE, it broadens the company's 3D-IC work into a full-scale, high-visibility project. The central execution test is whether the companies can carry the design's bandwidth and efficiency goals into production while controlling the power, heat, and yield challenges Okanohara identified. If they can, PFN will have a credible answer to the inference bottleneck that its November 2024 analysis identified; if they can't, the L2000's "more than twice the bandwidth" claim stays a target on a roadmap.
Source: unite.ai
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
749 articles
Related articles
- AI Compute in 2027: Serving Economics Will Reshape Data Centers
- Memory Customers Can Get Only 65% of What They Need, VanEck Manager Says
- Inside Samsung's Massive Pyeongtaek Chip Complex, Where AI Demand Fuels a Building Boom
- Huawei tells AI chip customers to buy their own memory on the black market
- Deepseek open-sources DeepInfer, the engine behind its cheapest model