Hello! Welcome to the Spring 2023 edition of CMSC 818J at UMD CS!
Administrative Information
- Instructor: Bahar Asgari
- Email: bahar [AT] umd [DOT] edu
- TA: Ryan Synk
- Email: ryansynk [AT] umd [DOT] edu
- Lectures:
- Time: MW 2:00pm - 3:15pm
- Location: EGR 3102
- Office Hours:
- Time: MW 1:00pm - 1:45pm
- Location: IRB 5148
Description and topics
Domain-Specific Architectures (DSAs):
Today, most of the computations in the world, from large-scale data center applications to smaller programs on home appliances, are mainly running on general-purpose programmable CPUs, and more recently on GPUs, too. While programmability has made CPUs popular and easy to deploy in a wide range of application domains, it has made computation less power/cost-efficient and less scalable because even to perform the simplest logical operation (e.g., AND, OR) at hardware, several sophisticated steps are required to translate a software code to the hardware operation. For many years, while Moore’s law was still valid and the number of transistors on a microchip could double every two years, the costs associated with CPU-based computations had been acceptable. However, as Moore’s law has been dying (at least for certain technologies such as on-chip memory), to keep scaling performance efficiently, we must design domain-specific architectures (DSAs) that utilize a given hardware budget more efficiently by focusing on the implantation of only the operations that a particular application requires rather than supporting programmability. DSAs (a.k.a., hardware accelerators), have been proposed/designed for several applications with distinct characteristics at various scales including deep learning, computer vision, scientific computing, database systems, etc. Besides the growing research in academia, the industry has been heavily investing in developing DSAs. These days, any device you buy (e.g., cell phones or laptops) definitely includes at least one DSA. Just a couple of examples of many DSAs in today’s industry are Google’s TPU and Apple’s Neural Engine.
Course Objectives:
This course explores recent advances in both academia and industry in designing DSAs. This course overviews the characteristics of applications that enable them to greatly benefit from DSAs and the source of speedup in DSAs which mainly includes parallelism, memory access optimization, and data specialization.
Learning Outcomes:
As part of this course, students will get familiar with common techniques such as using systolic arrays and near-data processing in state-of-the-art DSAs to enable efficient computation. The course also summarizes the concept of hardware-software co-design. This course helps students understand key design-space trade-offs (e.g., balancing area-power-performance and/or specialization-generality) to design efficient DSAs and highlights important benchmarking metrics to evaluate the rapidly growing number of DSAs being developed in academia and industry.
Course Structure and Content:
This course includes lectures interspersed with heavy paper reading and discussions. The material for this course will be derived from recent papers from major computer architecture conferences (i.e., ISCA, MICRO, HPCA, ASPLOS) on DSAs. The semester-long assignments of this course will include paper reading/critique plus coding practices for DSA implementations or studies. The course also includes a research-oriented project that allows students to explore outstanding research topics in this field and propose their own solutions. To this end, each student first explores different application domains to choose one to work on. Then, they will be guided to take the following steps to proceed with their projects (each step has a partial credit). The selected final reports are encouraged to submit their work as a research paper and will also be supported to present their work at conferences.
- Write a proposal
- Explore the topics
- Browse the abstract of several papers and choose your favorite topic
- Select the state-of-the-art publications related to your favorite topic
- Choose metrics and compare prior work using those metrics
- Find out the unsolved challenges in prior work
- Develop a simple idea to solve the challenge
- Think about some simple solutions to solve the challenge
- Estimate your expected improvement compared to prior work
- Evaluate your work
- Define your infrastructure, simulator, programming language, etc.
- Choose your workloads and datasets
- Implement your work and run experiments
- Write your report
- Present your work to the class
Topics:
This course will cover topics including:
- Key concepts
- Performance measurement and modeling
- Perf tools, Roofline model, Little's law, ...
- Dataflow architectures and systolic arrays
- Data representation formats
- New technologies for DSAs
- Processing near/in/using memory, smart SSDs, CXL technology
- DSAs for:
- Deep learning
- Sparse problems
- Scientific computing
- Graph analytics
- Recommendation systems
- Data base systems
- Robotics
- NLP
- DNA Sequencing
Prerequisites
- You need to know some of the fundamental concepts of computer architecture. You may contact the instructor for more details.
- For the projects some basic knowledge of C/C++ programming and Linux commands are required that you can learn along the way!
Communication and discussion forums
For communication and sharing resources we will be using:
- Elms-Canvas (to share lectures, homework assignments, the project, and recordings)
- Piazza (for discussion)
- Gradescope (for grading)
Material and textbooks
This course is mainly based on recent papers from major computer architecture conferences
Evaluation
- Midterm Exam 30%
- Assignments 10% + 2%
- Assignment 0: 2%
- Assignment 1: 5%
- Assignment 2: 5%
- Paper Reading Practice 10%
- Presentation and leading the discussion: 5%
- Critiques and participation in discussions: 5%
- Research Project 50%
- Literature Review 10%
- Proposal 5%
- Report 15%
- Presentation 15%
- Peer Review 5%
Tentative schedule
Week | Date | Lesson | Released | Due | Comments |
---|---|---|---|---|---|
1 | 01/25/2023 | Logistics + Introduction to DSAs | Assignments 0, 1 | ||
2 | 01/30/2023 | Dataflow I | Assignments 0 | Roofline model, pipelining, and systolic arrays | |
2 | 02/01/2023 | Dataflow II + Streaming I | Instructions for topic exploration (form) | Systolic Arrays for DNNS, Streaming, Little's law, and decompression. | |
3 | 02/06/2023 | Streaming II | The impact of data dependency challenges on streaming | ||
3 | 02/08/2023 | HW/SW Co-design | Assignment 1 | hw/sw co-design for efficient streaming, rethinking systolic arrays for solving PDEs | |
4 | 02/13/2023 | Near-Data Processing (NDP) | Assignment 2 | Presentation phase I paper selection (optional form) | We explore NDP solutions for recommendation systems |
4 | 02/15/2023 | Practice Questions for Midterm | Instruction for literature reivew and paper critique. Presentation phase I schedule and critique assignments. | Topic selection (form) | |
5 | 02/20/2023 | TBD | |||
5 | 02/22/2023 | Midterm | Date of midterm is subject to change | ||
6 | 02/27/2023 | DSAs for scientific computing | Presentation phase II schedule | ||
6 | 03/01/2023 | Analong-digital and sparse-dense accelerators | Assignment 2 (3/3) | ||
7 | 03/06/2023 | Sparse computation and compression I | Instructions for proposal | ||
7 | 03/08/2023 | Sparse computation and compression II | Literature review (3/10) | ||
8 | 03/13/2023 | Sparsity in neural networks plus systolic arrays I | Instructions for final report | ||
8 | 03/15/2023 | Sparsity in neural networks plus systolic arrays II | Proposal (3/17) | ||
9 | 03/20/2023 | No Class | Spring Break | ||
9 | 03/22/2023 | No Class | Spring Break | ||
10 | 03/27/2023 | Recommendation systems and near memory solutions I | |||
10 | 03/29/2023 | Recommendation systems and near memory solutions II | |||
11 | 04/03/2023 | Graph analytics | Template for final presentation and instructions for peer review | ||
11 | 04/05/2023 | Deep dive into sparsity I | Milestone I and experimental setup (4/7) | ||
12 | 04/10/2023 | Deep dive into sparsity II | |||
12 | 04/12/2023 | Neuroscience and genome processing | |||
13 | 04/17/2023 | Security and privacy | |||
13 | 04/19/2023 | Transformers | |||
14 | 04/24/2023 | Deep dive into HW/SW co-design and NDP | |||
14 | 04/26/2023 | TBD | Milestone II (4/28) | ||
15 | 05/01/2023 | Final Presentations (topic:TBD)) | |||
15 | 05/03/2023 | Final Presentations (topic:TBD) | |||
16 | 05/08/2023 | Final Presentations (topic:TBD) | |||
16 | 05/10/2023 | Final Presentations (topic:TBD) | Final Report |
Disabilities Support Accommodations
In case academic accommodations are needed, you must provide a letter of accommodation from the Office of Accessibility and Disability Services (ADS) within the first two weeks of the semester. For details, see the section titled “Accessibility” available at Course Related Policies.
Mask Policy
Please check the latest campus policies regarding COVID protocols: https://umd.edu/4Maryland. There has been a recent update in the mask mandate: “Effective Monday, August 29, 2022, wearing a mask will not be required while indoors in most situations, including classrooms. However, wearing a KN95 mask is recommended while indoors for added protection.”
Academic Integrity
Academic dishonesty includes not only cheating, fabrication, and plagiarism, but also includes helping other students commit acts of academic dishonesty by allowing them to obtain copies of your work. In short, all submitted work must be your own. Cases of academic dishonesty will be pursued to the fullest extent possible as stipulated by the Office of Student Conduct.
It is very important for you to be aware of the consequences of cheating, fabrication, facilitation, and plagiarism. If the student is found to be responsible for academic dishonesty, the typical sanction results in a special grade “XF”, indicating that the course failed due to academic dishonesty. If you have any doubt as to whether an act of yours might constitute academic dishonesty, please contact your TA or the course coordinator.
The CS Department takes academic integrity seriously. Information on how the CS Department views and handle academic integrity matters can be found at Academic Integrity.
A few examples of academic integrity violations
- Hardcoding of results in a project assignment. Hardcoding refers to attempting to make a program appear as if it works correctly (e.g., printing expected results for a test).
- Using any work available on the internet/web or any other source. For example, using work from StackOverflow, Sourceforge, or GitHub.
- Hiring any online service to complete an assignment for you.
- Discussing project implementations (everything beyond clarifications) with your classmates.
- Sharing your work with any students.
- Providing ideas/suggestions on how to solve/implement a programming assignment.
- Looking at another student’s work.
- Using online forums to ask for help regarding our assignments.
Additional information can be found in the sections titled “Academic Integrity” and “Code of Student Conduct” available at Course Related Policies.
For more information on the Code of Academic Integrity or the Office of Student Conduct, visit https://studentconduct.umd.edu.
Excused Absences
If you need to be excused for an absence from a single lecture due to a medical reason, you shall make a reasonable attempt to inform the instructor of your illness prior to the class. Upon returning to the class, you will present a self-signed note attesting to the date of your illness. Each note must contain an acknowledgment by the student that the information provided is true and correct. Providing false information to University officials is prohibited under Part 9(i) of the Code of Student Conduct (V-1.00(B) the University of Maryland Code of Student Conduct) and may result in disciplinary action.
Missing an exam for reasons such as illness, religious observance, participation in required university activities, or family or personal emergency (such as a serious automobile accident or close relative’s funeral) will be excused so long as the absence is requested in writing in advance and the student includes documentation that shows the absence qualifies as excused.
A self-signed note is not sufficient for exams because they are Major Scheduled Grading Events. In the case of medical absence, you must furnish documentation from the health care professional who treated you. The documentation must clearly include verification of (1) treatment dates and (2) the time period for which the student is unable to meet academic responsibilities. In addition, it must contain the name and phone number of the medical service provider to be used if verification is needed. No diagnostic information will ever be requested. Note that simply being seen by a health care professional does not constitute an excused absence; it must be clear that you were unable to perform your academic duties.
For additional details, see the section titled “Attendance and Missed Assignments” available at Course Related Policies.
Copyright
All course materials are copyright UMCP, Department of Computer Science © 2022. All rights reserved. Students are permitted to use course materials for their own personal use only. Course materials may not be distributed publicly or provided to others (except other students in the course), in any way or format.
Although every effort has been made to be complete and accurate, unforeseen circumstances arising during the semester could require the adjustment of any material given here. Consequently, given due notice to students, the instructor reserves the right to change any information on this syllabus or in other course materials.
Concerns
If you have any class concerns, please feel free to contact the instructor. If an issue arises with the instructor, report it using the form available at https://www.cs.umd.edu/classconcern.