CSL 2018 PhD Thesis Award Presentation: Mingu Kang

Selected annually, the CSL PhD Thesis Award is a prestigous honor bestowed to one graduate annually. This year's winner, Mingu Kang, will return to campus to accept the award from the committe and his adviser Naresh Shanbhag. Kang will also give a presentation of his award winning research, a summary of which can be found below.

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 suffers from the “Memory Wall,” i.e., the interface separating memory and computation. The Wall dominates both the energy efficiency 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 efficiency 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.