SAFARI Project & Seminars Courses (Spring 2023)

The increasing difficulty of scaling the performance and efficiency of CPUs every year has created the need for turning computers into heterogeneous systems, i.e., systems composed of multiple types of processors that can suit better different types of workloads or parts of them. More than a decade ago, Graphics Processing Units (GPUs) became general-purpose parallel processors, in order to make their outstanding processing capabilities available to many workloads beyond graphics. GPUs have been a critical key to the recent rise of Machine Learning and Artificial Intelligence, which took unrealistic training times before the use of GPUs. Field-Programmable Gate Arrays (FPGAs) are another example computing device that can deliver impressive benefits in terms of performance and energy efficiency. More specific examples are (1) a plethora of specialized accelerators (e.g., Tensor Processing Units for neural networks), and (2) near-data processing architectures (i.e., placing compute capabilities near or inside memory/storage).

Despite the great advances in the adoption of heterogeneous systems in recent years, there are still many challenges to tackle, for example:

• Heterogeneous implementations (using GPUs, FPGAs, TPUs) of modern applications from important fields such as bioinformatics, machine learning, graph processing, medical imaging, personalized medicine, robotics, virtual reality, etc.
• Scheduling techniques for heterogeneous systems with different general-purpose processors and accelerators, e.g., kernel offloading, memory scheduling, etc.
• Workload characterization and programming tools that enable easier and more efficient use of heterogeneous systems.

If you are enthusiastic about working hands-on with different software, hardware, and architecture projects for heterogeneous systems, this is your P&S. You will have the opportunity to program heterogeneous systems with different types of devices (CPUs, GPUs, FPGAs, TPUs), propose algorithmic changes to important applications to better leverage the compute power of heterogeneous systems, understand different workloads and identify the most suitable device for their execution, design optimized scheduling techniques, etc. In general, the goal will be to reach the highest performance reported for a given important application.

Prerequisites of the course:

• Digital Design and Computer Architecture (or equivalent course).
• Familiarity with C/C++ programming and strong coding skills.
• Interest in future computer architectures and computing paradigms.
• Interest in discovering why things do or do not work and solving problems
• Interest in making systems efficient and usable

The course is conducted in English.

The course has two main parts:
1. Weekly lectures on GPU and heterogeneous programming.
2. Hands-on project: Each student develops his/her own project.

Course description page Moodle

Spring 2023 Lecture Playlist

Fall 2022 Lecture Playlist

Spring 2022 Lecture Playlist

Fall 2021 Lecture Playlist

• An introduction to SIMD processors and GPUs:
  • Video
• An introduction to GPUs and heterogeneous programming:
  • Video

• Example recent studies of FPGA and GPU implementation for bioinformatics:
  • GateKeeper: FPGA for bioinformatics (Bioinformatics 2017)
  • SneakySnake: Pre-alignment filter on FPGA and GPU (Bioinformatics 2020)
• An example recent study of a suite of heterogeneous benchmarks:
  • Chai: heterogeneous benchmarks (ISPASS 2017)

• An example recent study of a medical image application on GPU:
  • GPU for medical imaging (CMPB 2020)
• Example studies of programming tools and performance portability on heterogeneous systems:
  • Boyi: execution models for FPGAs (FPGA 2020)
  • Zorua: hardware support for GPU performance portability (MICRO 2016)
  • Locality descriptor: Cross-layer abstraction to express data locality on GPUs (ISCA 2018)
• Example studies of scheduling techniques for heterogeneous systems:
  • Thread scheduling (MICRO 2011)
  • DASH: memory scheduling (TACO 2016)

• HW0 handout