CIS 6010: Special Topics in Computer Architecture: GPGPU Architecture and Programming Fall 2023

Course Information

instructor: Joe Devietti
when: Monday/Wednesday 12-1:30pm
where: Towne 305
contact: email, canvas

office hours:

• by appointment

Course Description

Graphics Processing Units (GPUs) have become extremely popular and are used to accelerate an increasingly diverse set of non-graphics workloads. This seminar will examine modern GPU architectures, the programming models used to write general-purpose code for GPUs, and the complexities of programming such highly parallel architectures. There will be a special emphasis on concurrency correctness issues as they relate to GPUs, including GPU memory consistency models and GPU concurrency bugs. Graduate-level coursework in computer architecture (e.g., CIS 5710) will be very helpful.

Course Materials

No textbooks are required; links to all readings will be provided at this website.

Grading

• Project: 50%
• Participation: 30%
• Assignments: 20%

There will be no exams.

Submit homework via Canvas.

The class project can be done in groups of up to 3. The project is open-ended: it should be something related to GPUs but the specifics are up to you. Choosing a project that incorporates your interests (research or otherwise) is a great idea! Here are some project ideas:

• Rewrite your matrix multiply code from the homeworks to operate on 16-bit (“half”) floating point elements instead of 32-bit floats. Update the cuBLAS code correspondingly, and use Tensor Cores to accelerate your implementation.
• Build a series of scalable locking implementations in CUDA, from simple spin-locks to something like MCS locks. The lack of coherence on GPUs should add an interesting wrinkle. Useful resources are Michael Scott’s webpage and the SSync library from EPFL.
• Choose a GPU-related paper (e.g., one that we’ve read in class, though others are fine as well) that has source code available, and try to reproduce some of the results from it.
• Pick a non-trivial open Github issue for an application written in CUDA, and try to resolve it. As a starting point, here are some popular GH repositories with CUDA code.
• Port an application or algorithm of interest to you to CUDA, and benchmark its performance.

Course Schedule

This schedule is subject to change