2018 CSL PhD Thesis Award Presentation: Mingu Kang

There is much interest in embedding data analytics and artificial intelligence capabilities into sensor-rich
platforms such as wearables, biomedical devices, autonomous vehicles, robots, and the Internet of Things (IoT) to
provide them with decision-making capabilities. Such platforms need to implement machine learning algorithms
under severe resource constraints on embedded battery- powered platforms.

However, the traditional digital von Neumann architecture suers from the “Memory Wall,” i.e., the interface separating memory and computation. The Wall dominates both the energy eciency and the throughput in digital architectures, especially in those that implement machine learning systems. Deep in-memory architectures (DIMA) overcome the Wall by having analog computations embedded in close proximity to a standard memory array (before sense amplification).

DIMA employs low-signal-to-noise-ratio (low-SNR) analog processing and delayed decisions to achieve both energy eciency and throughput gains without any loss in inference accuracy relative to a conventional digital architecture.The DIMA concept has been validated via three silicon IC prototypes in 65 nm CMOS, demonstrating up to 100X gains in the energy-delay product for decision-making systems.

Furthermore, a DIMA-based in-memory instruction set architecture with an LLVM-based compiler to provide an end-to-end application-mapping software framework has also been developed, demonstrating DIMA’s versatility. This talk will describe the DIMA concept, design principles, challenges, silicon-measured results, and future prospects.