Hot news

Implementation of an In-Memory-Compute Circuit in 16 nm FinFET Technology [mt]

Ulm University

Tutor: Johannes Stark - Stichworte: Artificial Neural Networks, Inference Accelerators, CMOS Circuit Design

The trend towards Artificial Intelligence (AI) has gathered significant impact and demands new hardware solutions. The distinct program structure of neural networks, requires highly parallelized and data-intensive operations to be executed which is facilitated by dedicated inference processors.
A promising approach for building such processors is mixed-signal hardware. Such circuits conduct the most commonly used multiply-accumulate (MAC) operation using analog quantities (charge, voltage, …). In-memory-compute (IMC) systems are an example for such circuits. They elegantly merge the needs of the inference hardware for memory and a smart data-flow with analog compute circuits.
The goal of this thesis is to reimplement an existing In-Memory-Compute Design (planar, 28nm) within a 16nm FinFET technology. Therefore, the reusability of existing design resources has to be evaluated and proper adaptions has to be made regarding FinFET peculiarities. Furthermore, a design layout should be developed including the assessment of layout-related parasitic influences. The architecture to work on comprises standard SRAM enhanced by Switched-Capacitor circuitry (various CDACs) and a SAR ADC.

What we expect:

  • Good understanding of Switched-Capacitor circuits and passive CDAC architectures (e.g. Integrated Analog Circuits, Integrated Interface Circuits)
  • Ideally some knowledge of Cadence Virtuoso (e.g. Project Analog CMOS Circuit Design)
  • Organized and well documented research and dedication to successful work

What we provide:

  • Experience in the field of digital, analog and mixed-signal design using industry-standard tools and leading-edge simulation concepts
  • All tools provided and made easily available, also for remote-work
  • An insight into state-of-the-art design techniques for ASICs

 

Implementation of an In-Memory-Compute Circuit