The list is not exhaustive, this is just an invitation for further discussion between the executor and the instructor.
We intend to design and develop an accelerator that is power and energy efficient. These kind of accelerators could be ASIC or FPGA based. The interesting challenge is to figure out the trade-off between efficiency, power/energy, and reliability for specific applications. The project requires a knowledge about the machine learning concepts, digital VLSI, Processor design, Micro-controller, Architecture, hardware descriptor languages (Verilog and VHDL), simulation process, high-level synthesis, C/C++ and Python.
Three challenging problems: power, reliability, and complexity of the contemporary SoC design needs to be addressed in a drastically new way. The approach must be integral (or wholistic) where the information from front-end could be percolated down to the back-end design. The main idea is how smartly the front-end knowledge could be used in back-end to deal with the design parameters. The problem is challenging because of complex functionality. One of the areas that we intend to investigate is the machine learning approaches.
Our focus would be to evaluate the power efficiency of RISC-V processor. The project is intend to design application specific processors which are optimized for given parameters among power, energy, area, and reliability. To design new ISA (instruction set architecture) for such application specific processor is one of the agenda in this project.
This project involves writing a compiler which would parse the gate-level netlist and insert the scan architecture called Joint-scan which has been developed at our lab. One should have good knowledge about digital electronics, verilog, VHDL, C/C++/Python and compiler.
The open source idea is very powerful in the sense that it enables the design acceleration by many designers and it also give rise to developments of new tools and technique. So far we have in the VLSI CAD community most of the tools are proprietary, of course, there are few out there in the open domain, however these are not much active. We intend to develop, one by one with small efforts the open source CAD tools and make them available for use via cloud based environment.
The IoT devices, particularly the Edge devices, are being overloaded heavily with the data that are being sensed from the ambience. Sending those raw data over the network does not really make much sense as there is a limit in network bandwidth. Therefore, the wise is to send the essential data or information which are critical for the end application. The edge device now have to be equipped with some kind of processing capabilities and that demands for an energy efficiency. Energy efficiency is very much needed as most of the edge devices are powered by battery or by some mechanism of harvested energy. The project investigate the energy issues in processing elements primarily the micro-controller in different application domains. The project requires a skill set with respect to programming (in C/C++, python) and system design (verilog/vhdl and high level synthesis).
Dark silicon is real now. Many of the multi-core processor decides to operates at lower frequency even though technology permits to operate at higher GHz. The reason of course is the power dissipation. The project investigate the power problems with respect to different applications and look for a solution in the domain of domain specific architecture. Most of the work involves extensive use of architecture simulators. As part of this exercise we look at the applications in the domain of graph and machine learning.
If the machine learning has to reach the human accuracy at any point of time, the near data processing is going to play a very important role. As far as our understanding goes, the NDP technology is very close to the mechanism that the human brain process information (too big to talk about all these things). Our interest here is simply to understand the idea of NDP by solving the problem that is here. This project will be carried out using simulators and as needed the project will also make use of 3D memory.
More projects will be update here...