News

Preferred Networks plans vertically stacked DRAM for its next inference chips

SEMIFIVE will implement the MN-Core L2000 design and Samsung Foundry will make the logic die. Tape-out is targeted for early 2027.
Macro photograph of a packaged processor.

Illustrative image: Akshat Sharma on Unsplash

Preferred Networks and SEMIFIVE are developing the MN-Core L2000 inference processor with DRAM stacked above the logic, targeting more than twice the memory bandwidth of the preceding L1000 series. Tape-out is planned for the first quarter of 2027, adding a second design to a programme whose L1000 processor is itself scheduled for release next year.

Preferred Networks is designing the processor architecture and logic, while SEMIFIVE handles implementation and coordinates the route through manufacturing, DRAM supply, packaging and test. Samsung Foundry will manufacture the logic and provide the associated 3D design infrastructure, bringing the memory connections and power delivery into the physical implementation of the stack.

The arrangement places DRAM directly over the processing logic, shortening the electrical paths between the two. Preferred Networks illustrated this approach in an L series presentation earlier this year and linked it to the low power consumption of the underlying processor. Bringing memory and logic together offers scope for many closely spaced connections, but also puts the memory above a source of heat that must be managed through the assembled package.

Heat dissipation, power consumption and manufacturing yield are among the challenges Preferred Networks identifies for the project. These extend across the stack: the logic, memory and their connections have to work together after assembly, with the package carrying away heat while supplying power to both types of die. SEMIFIVE’s remit covers those implementation and production stages as well as the logic design.

In inference, the intended bandwidth increase would accelerate movement of model data and working state between memory and processing units. The resulting application performance will also depend on the model’s memory access patterns, working set and arithmetic requirements, so the companies’ target is expressed as a bandwidth improvement over L1000 rather than a measured increase in token generation.

The partners are presenting L2000 at SC26 in Chicago, ahead of its planned tape-out early next year. They have yet to set a delivery date or disclose the absolute bandwidth and memory capacity of the design.

Sources: Preferred Networks and SEMIFIVE release; Preferred Networks’ L series architecture presentation, page 49.

Continue reading