When: Subnday, lecture 15:30 – 17:30, Tutorial 17:30 – 18:30
Where: Taub XX + zoom (*)
Credit: 3 academic points
Format: Frontal presentations + students’ presentations
Grade: Based on assignments and final project (see below)
Prerequisites: Deep learning course 236781 or a similar one, Computer Architecture course 236267 or a similar one
Where: Taub XX + zoom (*)
Credit: 3 academic points
Format: Frontal presentations + students’ presentations
Grade: Based on assignments and final project (see below)
Prerequisites: Deep learning course 236781 or a similar one, Computer Architecture course 236267 or a similar one
Attendance: Students are required to attend, however zoom attendance and recordings will be provided in special cases.
Syllabus:
This graduate-level course aims to allow students to experience the relationship between
deep-learning algorithms and the hardware on which they are deployed. The course
discusses theoretical, practical, and research aspects, and aims to encourage students to
pursue future research and projects in these directions.
deep-learning algorithms and the hardware on which they are deployed. The course
discusses theoretical, practical, and research aspects, and aims to encourage students to
pursue future research and projects in these directions.
Description:
The course will focus on the relationship between deep-learning algorithms and the hardware on
which they are deployed. We will study how different types of accelerators impact performance
efficiency and power. Students will gain hands-on experience with tools and algorithms on current
hardware solutions. Some classes may be given by industrial experts.
In the scope of this course, we will discuss various approaches to acceleration of deep-learning
algorithms and their implementations. Among other topics, we will discuss:
The course will focus on the relationship between deep-learning algorithms and the hardware on
which they are deployed. We will study how different types of accelerators impact performance
efficiency and power. Students will gain hands-on experience with tools and algorithms on current
hardware solutions. Some classes may be given by industrial experts.
In the scope of this course, we will discuss various approaches to acceleration of deep-learning
algorithms and their implementations. Among other topics, we will discuss:
- CPU-based accelerators
- GPU-based accelerators
- FPGA and ASIC accelerators
- Support for low power
- CUDA and PTX kernels
- Dataflow and memory management
- Binarization and quantization
- parallelization on multiple accelerators
- Support for LLMs
Course assignments and Grade distribution:
- Practical Homework assignments (40% total): submitting 2-3 code-based homework
assignments, in the scopes of CUDA kernels, quantization of DNNs and FPGA accelerator. - Final project (60% total): Each student will freely choose a topic for the final project based on
a paper relevant to the course, and correspondingly fulfill the following assignments under the
guidance of the course staff: - Presenting the original paper and the project proposal (15%) - During class
- Final Project presentation (15%) - At a date TBD, after the end of the semester
- Submission of project’s report and code implementation (30%) – Will be submitted
along with the final project presentations - Submission of project’s poster (optional)