
Preferred Networks and SEMIFIVE target higher inference memory bandwidth with MN-Core L2000
The AMW Read
The planned successor advances a known chip program, but its bandwidth and manufacturing targets remain unproven.
Preferred Networks and SEMIFIVE target higher inference memory bandwidth with MN-Core L2000
Preferred Networks and SEMIFIVE are developing the MN-Core L2000, an AI inference chip designed to deliver more than twice the memory bandwidth of Preferred Networks' L1000 series. The L1000 is scheduled for release in 2027. The partners plan to complete the L2000 design and hand it to the manufacturer in the first quarter of 2027. Preferred Networks is responsible for the processor architecture and logic design; SEMIFIVE will handle design implementation and supply-chain coordination.
The project adds another specialized chip design to the AI infrastructure market, where inference performance can depend on how quickly a processor accesses memory. The L2000 design stacks DRAM vertically on MN-Core logic. Samsung Foundry will provide platform technology and 3D chip design infrastructure. The bandwidth target is a design goal, not a measured result, and the comparison is with the unreleased L1000 rather than a current competing chip.
For builders and investors, the next evidence to watch is whether the partners can complete the design and translate it into manufacturable hardware. Preferred Networks' CEO identified power delivery, heat dissipation, and manufacturing yield as challenges for stacked chips. The L2000 also has support from Japan's Acquisition, Technology & Logistics Agency under a security technology research program running from fiscal 2024 through 2028.
