To keep pace with the progress in AI development, Intel has modernized and enhanced development tools, beginning with the SYCL*-based parallel offload platform. With the capability to migrate smoothly from CUDA*, developers gain freedom from vendor lock.

A solid open source performance foundation has been established with a combination of libraries, including Intel® oneAPI Deep Neural Network Library (oneDNN), Intel® oneAPI Data Analytics Library (oneDAL), and Intel® oneAPI Collective Communications Library (oneCCL), building machine learning and AI frameworks for the future. This session explores the OpenVINO™ toolkit and PyTorch*, how to use these frameworks, and the role that NumPy, Numba*, and data parallel extensions play in this environment.

A demonstration of Intel® VTune™ Profiler completes the tour, showing how to resolve performance bottlenecks in AI workloads.

Topics covered include how to:

• Use scalable libraries to optimize data preparation, quantization, and convolutional neural network (CNN) and deep neural network (DNN) matrix calculations.
• Streamline the porting of CUDA code to SYCL.
• Perform open standards-based scalable GPU operations and SYCL accelerator offloading.
• Enhance predictive AI and generative AI (GenAI) in multi-node environments.
• Find and resolve performance bottlenecks in a variety of AI situations.

Learn that AI training and inference performance starts with a fully optimized software stack, and that AI workloads are best handled by having a solid understanding of platform use.

Skill level: Intermediate
 

• Download Intel® VTune Profiler
• Intel® oneAPI DPCC++/C++Compiler
• Intel® oneAPI Base Toolkit