PFN and SEMIFIVE Develop 3D-Stacked AI Inference Chip

PFN and SEMIFIVE Develop MN-Core L2000 AI Chip PFN and SEMIFIVE Develop MN-Core L2000 AI Chip

Japan’s Preferred Networks (PFN) is partnering with semiconductor design company SEMIFIVE to develop the MN-Core L2000 Series, a next-generation AI inference processor designed to address memory bandwidth constraints in generative AI systems. The chip is planned to deliver more than twice the memory bandwidth of PFN’s forthcoming MN-Core L1000 Series, with tape-out targeted for the first quarter of 2027. The project also brings Samsung Foundry into the manufacturing and 3D-IC design ecosystem, highlighting how chipmakers are increasingly exploring advanced packaging and memory integration to improve AI inference performance.

PFN targets the memory bottleneck in AI inference

The race to build faster AI processors is increasingly about moving data efficiently, not simply adding more computing power. Large language models and AI agents must repeatedly access model weights, intermediate data and growing key-value caches during inference. When memory cannot supply data quickly enough, expensive processing units can spend time waiting rather than computing.

Japan’s Preferred Networks (PFN) is betting on a different chip architecture to address that challenge. Its collaboration with South Korean semiconductor design company SEMIFIVE will produce the MN-Core L2000 Series, an inference-focused processor designed around vertically stacked logic and DRAM.

The L2000 is planned as the successor to PFN’s MN-Core L1000 Series, which is scheduled for release in 2027. PFN and SEMIFIVE are targeting more than twice the memory bandwidth of the L1000, with the L2000’s chip design scheduled to reach tape-out in the first quarter of 2027. Tape-out marks the handoff of completed chip design data to the manufacturing process; it does not mean the processor is ready for commercial deployment.

The announcement, made during the SC26 supercomputing conference in Chicago, brings together PFN’s processor architecture, SEMIFIVE’s custom-chip implementation and supply-chain capabilities, and Samsung Foundry’s manufacturing and 3D-IC design infrastructure.

A chip designed around data movement

The MN-Core L Series takes a different approach from conventional AI accelerators by placing memory bandwidth at the center of its design. PFN’s architecture vertically stacks logic chips based on its proprietary MN-Core design with DRAM, creating a three-dimensional integrated structure.

The idea is to shorten and expand the paths through which data moves between processing logic and memory. This can help reduce one of the constraints on generative AI inference, where repeatedly transferring large volumes of model data can limit throughput and increase latency.

The distinction matters because inference workloads behave differently from many traditional computing tasks. During model training, processors perform extensive calculations across large datasets. During inference, a model must continually retrieve parameters and manage intermediate information as it generates responses. Long-context prompts, multiple concurrent users and agentic workflows can intensify those memory demands.

A 2026 OECD assessment of computing infrastructure identifies memory bandwidth and latency as important constraints on AI performance, describing the widening gap between processor capabilities and memory data-transfer capacity as a memory-wall problem. The report also highlights power and heat management as related infrastructure limitations.

PFN’s 3D-stacked approach is intended to address several of these issues together. The company says the MN-Core L architecture can achieve substantially higher memory bandwidth than GPUs and other AI-processing chips, while improving power efficiency. Those are vendor claims, however; the L2000 announcement does not provide absolute bandwidth figures, power consumption measurements or independently verified performance comparisons.

Three companies, distinct roles

The collaboration divides responsibility across the chip development and manufacturing chain.

PFN is responsible for the processor architecture and logical design. The company has been developing its MN-Core technology since 2016 and has used earlier generations in its own computing services and products. PFN says supercomputers powered by the first-generation MN-Core architecture have topped the Green500 ranking of the world’s most energy-efficient supercomputers three times.

SEMIFIVE will handle design implementation and turnkey semiconductor development, including supply-chain coordination. Its role is particularly important because stacking logic and DRAM introduces design challenges that extend beyond conventional chip layout. The company will coordinate with partners across memory supply, manufacturing, advanced packaging and testing.

Samsung Foundry will manufacture the logic chip using its platform technology and provide design infrastructure for 3D integration. That includes support for die-to-die I/O connections and power-structure optimization.

The project also fits within Samsung’s broader SAFE ecosystem, which connects foundry customers with design-service, IP, electronic-design-automation and advanced-packaging partners. Samsung’s multi-die integration capabilities are intended to support complex 2.5D and 3D chip designs.

A different competitive path from GPU scaling

The L2000 enters a market dominated by general-purpose GPU platforms, particularly those from NVIDIA, while cloud providers and semiconductor companies are also developing custom accelerators for specific workloads. Google’s TPU and Amazon’s Trainium illustrate the broader move toward purpose-built AI silicon.

PFN’s strategy is narrower: optimize the processor and memory architecture specifically for inference rather than compete solely through peak compute figures. That could appeal to operators running high-volume generative AI services, enterprise models and agent workloads where memory movement, latency and power consumption affect operating economics.

But high memory bandwidth alone does not guarantee a faster or cheaper AI system. Performance also depends on compute throughput, memory capacity, software compilers, model compatibility, interconnects, thermal design and how efficiently workloads are scheduled. The L2000’s eventual value will depend on how those components work together.

The engineering risks are substantial. PFN’s CEO has identified power consumption, heat dissipation and manufacturing yield as particular challenges for 3D-stacked AI chips. Stacking memory and logic can improve data access, but it also makes thermal management, packaging and production more demanding.

Development milestone, not a finished product

The MN-Core L2000 is still under development. The immediate milestone is tape-out in early 2027, while the L1000 is itself scheduled for release in 2027. PFN has not disclosed the L2000’s final specifications, commercial availability, pricing or customer deployment plans.

The project is also receiving support from Japan’s Acquisition, Technology & Logistics Agency (ATLA) under its Innovative Science and Technology Initiative for Security. The funded research project focuses on accelerating deep learning and inference through 3D-integrated LSI technology and runs from fiscal 2024 through fiscal 2028. Preferred Networks, Inc.

For PFN, SEMIFIVE and Samsung, the collaboration is a test of whether a specialized 3D-stacked architecture can translate promising memory-bandwidth targets into manufacturable, energy-efficient AI silicon. If it succeeds, the L2000 could give AI infrastructure operators another option beyond conventional GPU-based systems. The decisive evidence, however, will come from production silicon, software support and real-world inference benchmarks.

Market Landscape

AI chip development is moving beyond the pursuit of raw compute performance. Memory bandwidth, capacity, latency, power consumption and advanced packaging are increasingly important factors in determining the economics of inference.

Technology approachMain focusMarket relevance
GPU acceleratorsBroad AI compute and mature software ecosystemsWidely deployed for training and inference
Custom AI ASICsWorkload-specific compute efficiencyCloud providers and specialist chip developers seek better performance per watt
High-bandwidth memoryFaster access to large volumes of dataImportant to modern AI accelerator architectures
3D-stacked logic and DRAMCloser integration of memory and processing logicPotential alternative architecture for bandwidth-sensitive workloads

PFN’s MN-Core L Series is positioned in the specialized inference-accelerator segment. Its emphasis on vertically stacked DRAM and logic differentiates it from conventional GPU-centered infrastructure, although the L2000’s commercial competitiveness cannot be assessed until detailed specifications and benchmarks become available.

The wider semiconductor ecosystem is also investing in advanced packaging. Samsung’s SAFE program provides design and integration support for multi-die and advanced 3D packaging, illustrating how foundries increasingly rely on partner ecosystems to develop complex AI silicon.

Top Insights

  • PFN and SEMIFIVE are developing the MN-Core L2000, targeting more than twice the memory bandwidth of the planned L1000 generation.
  • The architecture vertically stacks logic chips and DRAM, aiming to improve data movement and inference efficiency.
  • SEMIFIVE will manage design implementation and supply-chain coordination, while Samsung Foundry provides manufacturing and 3D-IC design infrastructure.
  • Tape-out is targeted for Q1 2027, but commercial availability, pricing and final performance specifications remain undisclosed.
  • The project reflects a broader shift toward specialized inference silicon as memory bandwidth and power become critical AI infrastructure constraints.

Power Tomorrow’s Intelligence — Build It with TechEdgeAI

Grow Your
Brand Visibility

Looking to publish a press release, guest article, interview or podcast? Connect with us.

GET FEATURED
Subscribe

Sign up today for exclusive insights and updates.

Newsletter Signup