1. ECE 8823: GPU Architectures
NVIDIA Keplar
ECE 8823: GPU Architectures
Sudhakar Yalamanchili
School of Electrical and Computer Engineering
Georgia Institute of Technology

2. Course Objectives Basic CUDA Programming
BSP Programming Model : Basic structures and idioms
Simple microarchitecture-targeted optimizations
Core concepts of general purpose graphics processing unit (GPGPU) architecture
Execution Model: How massively parallel program execution is orchestrated using the Single Instruction Multiple Thread (SIMT) model
Basic microarchitecture: Organization principles
Advanced microarchitecture concepts
Optimization of control flow, memory access behavior, and thread scheduling
Datapath organization, power/energy optimizations
Competing execution models: Systolic and Dataflow

3. What You Will Do Understand how GPU architectures work
Fundamental architectural concepts and challenges
Acquire knowledge to optimize systems
As a designer
As a user
As a researcher
Assignments (tentative)
CUDA Programming (2-3)
Microarchitecture (2-3)
Build an instruction set emulator for a GPU processor
Implement microarchitectural solutions for control and memory divergence
Project: may range from cycle-level microarchitecture evaluations to major application/algorithm development

4. Morgan Kaufmann Publishers
April 6, 2019
What You Will Learn
How the bulk synchronous parallel (BSP) model is represented in CUDA
Basic organizational principles of a CUDA program
To write some simple-to-moderate complexity CUDA programs
Will not cover advanced CUDA features
The Single Instruction Multiple Thread (SIMT) execution model
And how the hardware executes them
Determinants of SIMT program performance
And how it can be improved
Chapter 1 — Computer Abstractions and Technology

5. Course Information Web page:
Will be constantly updated, so check it regularly
Prerequisite: ECE 4100/6100 or CS 4290/6290
Course Materials:
D. Kirk and W. Hwu, “Programming Massively Parallel Processors: A Hands On Approach,” Morgan Kaufmann (pubs), Second Ed., 2012
Journal & Conference papers, patents, class notes
Online content for the text

6. Grading Policy Homework Assignments: 40% (5%, 10%, 10%, 15%) Exams
Individual work, no collaboration other than specified!
No late assignments will be accepted
You are expected to make a passing grade on the assignments to pass the course!
Exams
Midterm: 15% (Monday March 5th, 2017)
Assignments: 40%
Project: 30%
Final: 15%: Wednesday May 2nd, 11:30-2:20
Teaching Assistants: TBD

7. Academic Honesty
You are allowed and encouraged to discuss assignments with other students in the class
Any copying of code is unacceptable
Includes reading someone else’s code and then going off to write your own
Includes verbal specific directions on data structures and code
Giving/receiving help on an exam is unacceptable
Incidents of academic dishonesty will be referred to Dean of Students and subject to at least a letter grade penalty or an F in the course

8. Preliminary Course Layout*
CUDA Programming (CUDA Programming Guide)
Introduction to CUDA C (Chapter 3)
CUDA Data Parallel Execution Model (Chapter 4)
Memory Hierarchy (Chapter 5, Chapter 8.3)
Performance Issues (Chapter 6)
Basic GPU Microarchitecture (papers, patents)
Control Divergence (papers)
Memory Divergence (papers)
Scheduling (papers)
Power & Energy Optimization (papers)
Advanced Topics (Systolic, Dataflow, OpenCL, OpenACC)
3-4 Weeks
2-3 Weeks
2 Week
1 Week
2 Weeks
1-2 Weeks
1-2 Weeks
*Note that this is subject to change

9. Tentative Lecture and Assignment Schedule
Week 1
Jan 8
Overview & Introduction
Week 2
Jan 15
CUDA Execution Model
Assn 1
Week 3
Jan 22
CUDA Memory Model
Assn 1 Due/Assn 2
Week 4
Jan 29
CUDA Memory Optimizations
Assn 2a Due
Week 5
Feb 5
arch:kernel exec
Assn 2b due
Week 6
Feb 12
arch: SM microarchitecture and register file
Assn 3
Week 7
Feb 19
Control-Div
Week 8
Feb 26
Assn 3 due/Assn 4
Week 9
March 5
Memory Divergence
Midterm (March 5th)
Week 10
March 12
Scheduling
Project Writeups Due
Week 11
March 19
Spring break
Week 12
March 26
Assn 4 due/Feedback on Projects
Week 13
April 2
Power & Energy Optimizations
Week 14
April 9
Dataflow and Systolic Computation
Project status due
Week 15
April 16
Review OpenCL and OpenACC
Week 16
April 23
Special Topics
May 2nd
Project Report Due, Final Exam

10. Lets Get Started……….