CUDA

HPC-HPDA on GPU with CUDA

Description: This course aims to introduce high performance algorithmics and programming on GPU, with experiments on Machine Learning algorithms run on GPU servers.

Content: GPU architecture Algorithmic principles of fine grained GPU parallelism (SIMD and SIMT models) CUDA programming Usage of CUBLAS library Optimization of GPU and CPU-GPU CUDA codes Design and experiment of K-means algorithms on GPU

Prerequisites: 1st year common course: “Systèmes d’Information et Programmation” (1CC1000) 1st year common course: “Algorithmique & Complexité” (1CC2000) C++ Advanced programming course (3MD1020) of SDI mention at Metz Automatic Learning course (3MD1040) of SDI mention at Metz

Learning outcomes: At the end of this course, students will be able: Learning Outcome AA1: to analyse the adequacy of a mathematical solution with an implementation and execution on GPU, Learning Outcome AA2: to design a GPU algorithm, or to adapt an algorithm to increase its efficiency on GPU, Learning Outcome AA3: to design hybrid algorithms for CPU-GPU systems, overlapping data transfers and computations, Learning Outcome AA4: to implement algorithms and to debug codes on GPU, Learning Outcome AA5: to analyse and to summarize GPU software.

Means: Development and execution platform: GPU servers of the Data Center for Education of CentraleSupélec Metz Campus. NVIDIA CUDA development environment.

Evaluation methods: Evaluation of Lab results about parts 2 and 3, and final and individual exam Reports of the Lab about parts 2 and 3 (the content and the number of pages of the reports are constrained, in order to force the students to an effort of synthesis and clarity) In the event of unjustified absence from a practical work, the mark of 0 will be applied, in the event of justified absence the average mark of other labs will be applied. The remedial exam will be a 1 hour oral exam, which will constitute 100% of the remedial mark.

Evaluated skills:

  • Be operational, responsible, and innovative in the digital world
  • Know how to convince

Course supervisor: Stéphane Vialle

Geode ID: 3MD4030