GPU Architectures, Spring 2016 (CS 780 / CS 680)

General Information

• Time and Location: Tuesday/Thursday 3:30PM-4:50PM, Washington Hall 310

• Instructor: Adwait Jog (Personal Website)

• Office hours: Tues/Thu 10:30AM-noon or by appointment, McGl 111

• Email: adwait@cs.wm.edu
  

• Deadlines: Jan 29 (add/drop deadline) and Mar 18 (withdraw deadline)

Prerequisites

Students are expected to have a good understanding of the basic computer organization and design. Please talk to the instructor if you do not satisfy this requirement.

Course Description

This course provides an in-depth understanding of the architectural and micro-architectural details of a general purpose graphics processing unit (GPU). It also provides an in-depth coverage of important research issues associated with GPU computing by covering a range of latest papers that have appeared in leading international journals and conferences. Students are expected to read a variety of papers, critique them, and present them in the front of the class. In addition, students are expected to complete a semester-long research project and simulation-based assignments.

Grade Distribution

• (I) Paper critiques: 10%

• (II) In-Class Presentations: 20%

• (III) Simulation-based Assignments: 30%

• (IV) Semester-long Research Project: 40%

Submission Format

• In-Class Presentations will be uploaded in a shared Box folder visible to all students.

• Critique and Report Submissions via Box. A separate Box folder will be shared with the instructor and the student.

• Final grade submission via Banner.

(I) Paper Critiques

All students are required to submit a detailed critique for each paper we discuss in the class. However, the student who presents the paper in-class is not allowed to submit the critique for that particular paper. Deadline for critique submission is one week from when the paper is discussed completely in the class. Please submit any FIVE critiques in the PDF format.

Submission Format: Each critique should not exceed one-page and must consists of four sections: 1) paper summary (2-3 lines), 2) strengths (2-3 lines), 3) weaknesses (2-3 lines), and 4) detailed comments (rest of the page). More details are already discussed in class and associated slides are submitted to the box folder (shared with students).

No collaboration is allowed on critiques.

(II) In-Class Presentations

Each student will present a maximum of two papers throughout the semester. If you plan to audit the course, you are required to present at least two papers. When you present a paper, be prepared to answer a variety of questions asked by the instructor or other fellow students. The goal is to make class lively. A list of papers will be provided to students. They can choose from that list or come up with their own suggestions. Suggestions would need approval from the instructor.

In-Class Paper Presentations: I expect students to first present necessary background (~15 minutes) and then paper details (~25 minutes). The remaining time will be for discussion driven by the fellow students and the instructor.

(III) Simulation-based Assignments

Assignments will be given based on the GPGPU-Sim simulator. GPGPU-Sim Installation guide (for bg machines at CS, WM) is here. No collaboration is allowed on assignments.

(IV) Semester-Long Research Project

Students are expected to perform a semester-long research project. All projects need to be approved by the instructor. Please contact the instructor early to brainstorm potential project ideas.

Project Timeline

• Phase 1 – Project Determination: (Deadline: Feb 6); Please send an email to the instructor by the deadline containing: 1) Project Name (think of this as your paper/report title), 2) Problem Statement, 3) Expected Infrastructure Platform Required, 4) Possible Outcomes and Deliverables.
  

• Phase 2 – Determination of (more) concrete project goals and Infrastructure Setup (Deadline: Feb 25 ); Please meet the instructor during office hours to discuss the status of your project.

• Phase 3 – Motivation Results (Deadline: April 5); Please meet the instructor during office hours to discuss the status of your project.

• Phase 4 – Final Project Report (Deadline: May 10); Please email your final project report in PDF format. Please use the standard LaTeX or Word ACM templates. The PDF should have these sections: 1) Problem Statement, 2) Introduction, 3) Background and Related Work, 4) Motivation Results, 5) Implementation Details, 6) Infrastructure Details, 7) Final Results, and 8) Conclusions.

Areas for In-class Presentations and Projects

• GPU Core Design

• GPU Warp Scheduler Design

• GPU Thread block Scheduler Design

• GPU Cache Design

• GPU Memory System Design (incl. emerging memory technology design)

• GPU Processing-in-Memory

• GPU Interconnect Design

• GPU Reliability

• Multi-GPU systems

• Multi-context GPU Architecture

• Mobile GPU Architecture

• Approximate GPU Architecture

• Heterogeneous GPU platforms that incorporate GPUs

Useful Simulators and Tools

• GPGPU-Sim, A GPU Simulator (models NVIDIA-style GPUs). Also, look at GPU-Wattch, A GPU Power Model

• MacSim, A trace-driven CPU/GPU simulator

• SASSI, A Flexible GPGPU Instrumentation Tool

• MAFIA, A Multiple Application/Context Framework for GPU architectures

• Multi2sim, A CPU/GPU Simulator (models AMD-style GPUs)

• CPU-GPU Simulator, A Trace/Execution Driven CPU-GPU Heterogeneous Architecture Simulator

• GemDroid, A Mobile Architecture Simulator (GEM5 + Android Emulator)

Semester Schedule

Readings

• P1: Architectural Support for Address Translation on GPUs, ASPLOS 2014

• P2: Improving GPU Performance via Large Warps and Two-Level Warp Scheduling, MICRO 2011

• P3: Cache-Conscious Wavefront Scheduling, MICRO 2012

• P4: WarpPool: sharing requests with inter-warp coalescing for throughput processors, MICRO 2015

• P5: Dynamic Warp Formation and Scheduling for Efficient GPU Control Flow, MICRO 2007

• P6: GPU Register File Virtualization, MICRO 2015

• P7: Throughput-Effective On-Chip Networks for Manycore Accelerators, MICRO 2010

• P8: Cache Coherence for GPU Architectures, HPCA 2013

• P9: Equalizer: Dynamic Tuning of GPU Resources for Efficient Execution, MICRO 2014

• P10: A Case for Core-Assisted Bottleneck Acceleration in GPUs: Enabling Flexible Data Compression with Assist Warps, ISCA 2015

• P11: Neural acceleration for GPU throughput processors, MICRO 2015

• P12: A Complete Key Recovery Timing Attack on a GPU, HPCA 2016

Academic Accommodations

It is the policy of The College of William and Mary to accommodate students with disabilities and qualifying diagnosed conditions in accordance with federal and state laws. Any student who feels s/he may need an accommodation based on the impact of a learning, psychiatric, physical, or chronic health diagnosis should contact Student Accessibility Services staff at 757-221-2509 or at sas@wm.edu to determine if accommodations are warranted and to obtain an official letter of accommodation. For more information, please click here.

Honor Code

Students are required to follow the Honor System of the College of William and Mary.