An Energy-Efficient Heterogenous CPU-GPU Accelerator for Machine learning

2021 ELE Engineering Design Project (FM01)

---

Faculty Lab Coordinator

Farah Mohammadi

Topic Category

Consumer Products/Applications

Preamble

Machine learning (ML) algorithms are currently widely used for many artificial intelligence (AI) applications. Computer vision, speech recognition, and robotics are some examples of these AI-based applications. While ML techniques deliver state-of-the-art accuracy on many AI applications, it comes at the cost of high energy and computational complexity. Accordingly, designing energy-aware high-throughput platforms that enable efficient processing of ML algorithms without sacrificing application accuracy or increasing hardware overhead is critical to the wide deployment of ML techniques in smart-AI-based systems.

Objective

The goal of this work is architecting an energy-efficient heterogenous CPU-GPU platform with optimum number of CPUs and GPUs to improve the performance and energy consumption of ML algorithms. Popular AI projects in these days are looking for a low-power high-performance framework for running ML algorithms in these days. Heterogenous systems-on-chip are the efficient solution for improving the energy and performance of the ML algorithms due to the recent researches.

Partial Specifications

-An optimum CPU/GPU ratio model should be developed.
-The CPU/GPU ratio model should be verified for each application.
- ML algorithms characteristics (number of memory accesses and CPU/GPU usages) should be studied.
- Performance counters and lookup tables for each simulators (GPGPUSim and GEM5) should be studied and set.
- Gem5 should be installed.
-GPGPUSim should be installed.
- McPAT and HotSpot should be installed.

Suggested Approach

1- Developing an optimization model for obtaining the optimal ratio number of CPUs and GPUs.
2- Studying on different ML techniques and their characteristics such as memory accesses and CPU-GPU usages.
3- Simulating a Pure CPU framework on GEM5.
4- Simulating a Pure GPU framework on GPGPUSim.
5- Simulating a Heterogenous CPU-GPU platform for each ML technique.

Group Responsibilities

- Investigating available ML algorithms in existing Simulators such as GEM5 and GPGPUSim that Intel and AMD are working on them.
- Working on GEM5.
-Working on GPGPUSim.
- Integration of GEM5 and GPGPUSim.
-Deriving an optimum model of resources for ML algorithms.
-Prepare a technical report and present the results at the end of the program.

Student A Responsibilities

-Designing and verifying a Model for obtaining the optimal number of CPUs and optimal number of GPUs for each ML algorithm.

Student B Responsibilities

-Working with GEM5 Simulator for the CPUs.
- Working with McPAT as power and energy simulator.

Student C Responsibilities

-Working with GPGPUSim Simulator for the GPUs.
- Working with HotSpot as a thermal simulator.

Student D Responsibilities

-Working on integration of GEM5 and GPGPUSim for architecting the heterogenous platform.

Course Co-requisites

Digital Systems, Programming in C, Microprocessors

To ALL EDP Students

Due to COVID-19 pandemic, in the event University is not open for in-class/in-lab activities during the Winter term, your EDP topic specifications, requirements, implementations, and assessment methods will be adjusted by your FLCs at their discretion.