Doctor of Philosophy (PhD)
The Division of Electrical and Computer Engineering
How to implement quality computing with the limited power budget is the key factor to move very large scale integration (VLSI) chip design forward. This work introduces various techniques of low power VLSI design used for state of art computing. From the viewpoint of power supply, conventional in-chip voltage regulators based on analog blocks bring the large overhead of both power and area to computational chips. Motivated by this, a digital based switchable pin method to dynamically regulate power at low circuit cost has been proposed to make computing to be executed with a stable voltage supply. For one of the widely used and time consuming arithmetic units, multiplier, its operation in logarithmic domain shows an advantageous performance compared to that in binary domain considering computation latency, power and area. However, the introduced conversion error reduces the reliability of the following computation (e.g. multiplication and division.). In this work, a fast calibration method suppressing the conversion error and its VLSI implementation are proposed. The proposed logarithmic converter can be supplied by dc power to achieve fast conversion and clocked power to reduce the power dissipated during conversion. Going out of traditional computation methods and widely used static logic, neuron-like cell is also studied in this work. Using multiple input floating gate (MIFG) metal-oxide semiconductor field-effect transistor (MOSFET) based logic, a 32-bit, 16-operation arithmetic logic unit (ALU) with zipped decoding and a feedback loop is designed. The proposed ALU can reduce the switching power and has a strong driven-in capability due to coupling capacitors compared to static logic based ALU. Besides, recent neural computations bring serious challenges to digital VLSI implementation due to overload matrix multiplications and non-linear functions. An analog VLSI design which is compatible to external digital environment is proposed for the network of long short-term memory (LSTM). The entire analog based network computes much faster and has higher energy efficiency than the digital one.
Zhao, Zhou, "Techniques of Energy-Efficient VLSI Chip Design for High-Performance Computing" (2018). LSU Doctoral Dissertations. 4702.