Overview / Usage

The project focuses on Heterogeneous Computing and its uses to speed up Deep Learning algorithms. To do so, we will use oneAPI to train a simple Neural Network algorithm and compare the training speeds of using oneAPI and CUDA. We will create a Neural Network code in CUDA, then translate it using the Compatibility Tool and run it using the oneAPI Toolkit compiler.

Methodology / Approach

Since OneAPI provides solutions to Heterogeneous Computing at a framework level (at the "CUDA" level), we will first attempt to create our own custom Neural Network using CUDA code. After creating and analyzing the creation of the code, we will use the Compatibility Tool to convert out CUDA code into DPC. After that, we will modify that translated code so that the compiler can be executed. After doing so, we will compare the performance of the CUDA code vs. the DPC oneAPI code.

Technologies Used

We will use an Intel Base Workstation with an i9-13900K Processor, the Intel UHD Graphics 770, and an Intel ARC A770 to test our DPC++ code. Moreover, we will be using an NVIDIA TITAN Xp and NVIDIA RTX 4090 to test out the CUDA code.

Repository

https://github.com/Goleys/oneAPI_NeuralNetwork_Research.git

Collaborators