The current HPC landscape is changing. More and more HPC systems are being accelerated by GPUs that are optimized for flops per watt and throughput. Flops and throughput are important for most scientific applications, therefore, GPU support in scientific applications is a major advantage.
The development of applications for GPUs is not standardized as it is for multi-core CPUs. Each vendor has its own SDK with a different programming model for running algorithms on its own GPUs. This makes it difficult to port applications to GPUs from other vendors. Learning a vendor-specific GPU language also involves a lot of work. Alpaka 3 is a modern C++ 20 solution that allows you to write your GPU-accelerated code once and then run it on GPUs from different vendors. Alpaka 3 was developed with zero-overhead abstraction to achieve near-native performance. The new C++ 20 standard allows us to provide a user-friendly API that helps developers write their applications. Alpaka 3 also provides utilities to avoid common GPU programming mistakes and simplify writing fast GPU code. It also supports the ability to offload code to CPUs, making it perfectly portable.
The presentation provides an introduction to Alpaka 3. No prior knowledge of GPUs or GPU development is required.