Editing by Kate Yuan

A Chinese team from Fudan University in Shanghai proposed COMB-MCM, a Computing-In-Memory (CIM) AI chip at ISSCC 2022, a global solid-state circuits forum known as the "Chip Olympics", according to Fudan University’s Academy for Engineering & Technology website.

This research paper was written by Doctorate student Zhu Haozhe and two other graduate students. They intend to avoid the calculation errors of analog computing circuits and uses multi-chiplet integration to achieve the scalability of computing power after tape-out while keeping the performance and energy efficiency of the "non-von-Neumann" architecture of CIM.

The Fudan University information shows that this chip is manufactured in 65nm and 28nm processes respectively. The feasibility of the solution is verified by the MCM system of 2.5D IC packaging under the 65nm process, and better performance is achieved under the 28nm process.

The Fudan team said the CIM architecture can eliminate unnecessary data movement by integrating the data memory and the computing units, and breaking the von Neumann bottleneck caused by separated CPUs and memory.

Fudan University has mainly carried out technological exploration and innovations. Architecturally, an SRAM-based computing-on-memory-Boundary (COMB) architecture is proposed to further reduce the data movement and system power consumption.

At the system level, it builds a Multi-Chiplet-Module (MCM) to support AI tasks of different complexity. By configuring the number of chiplets, the computing and memory requirements can be increased proportionally.

This technology exhibits ultra-high energy efficiency in AI applications that require intensive memory access and is considered a crucial technology for next-generation AI chips.

Starting in 1954, ISSCC (International Solid-State Circuits Conference) is the foremost global forum for the presentation of advances in solid-state circuits and systems-on-a-chip, according to its website. The Conference offers a unique opportunity for engineers working at the cutting edge of IC design and application to maintain technical currency and to network with leading experts.