Digital In-Memory Computing: Design & Applications
DCIM EDA Tools
Digital in-memory computing technology can bring higher computational energy efficiency to AI chips. However, existing EDA tools do not support the automated design of in-memory computing modules. We has developed an automated platform tool, AutoDCIM, which is the world's first of its kind. It can generate corresponding in-memory computing modules for SoC design according to user requirements. It works with the CIM optimization module CIM-DSE. DSE conducts a design space exploration for CIM accelerators to find the optimal hardware design for given design objectives. CIM-DSE evaluates multiple hardware design parameters to achieve a design balance among performance, power consumption, and area.
Design Optimization
AutoDCIM takes user-defined design specifications as input and automatically generates a digital in-memory computing module. Its layout design is optimized and generated using a template-based layout generator and layout search algorithm.
Our center used the AutoDCIM tool to explore the design of digital in-memory computing modules with array sizes ranging from 4Kb to 256Kb in a 40nm process. The automatically generated digital in-memory computing modules achieve higher area and power efficiency compared to existing designs.
The digital in-memory computing unit optimized by AutoDCIM has been taped out and verified using 28nm process. The test results, reaching the world's advanced level, can provide computing units for different AI chips.
DCIM Chips
Our DCIM modules has successfully undergone manufacturing verification using 28nm technology. Our AC-DCIM module has demonstrated the best computing efficiency of up to 26 TOPS/W, which is 34% higher than TSMC's similar design (scaled to 22nm). This represents the most advanced performance and can be applied in high-performance edge computing chips.
DCIM Applications
The portable DNA analyzer we developed fully demonstrates the practical application potential of DCIM technology. This device enables faster and lower-cost genome analysis in remote or resource-limited environments, allowing advanced medical testing tools to break through geographical restrictions. By overcoming the performance bottlenecks of traditional platform, it provides smarter, more environmentally friendly, and more efficient solutions for various industries and communities, opening up brand-new possibilities for artificial intelligence-driven applications, healthcare, and other fields.