Overview / Usage

The purpose of this benchmark tool is to evaluate performance bounds of GPUs on mixed operational intensity kernels. Modern GPUs are able to hide memory latency by switching execution to threads able to perform compute operations. Using this tool one can assess the practical optimum balance in both types of operations for a GPU. CUDA, HIP, OpenCL and SYCL implementations have been developed.

Methodology / Approach

Each kernel is customized on a range of different operational intensity values. Four types of experiments are executed combined with global memory accesses:

1. Single precision Flops (multiply-additions)
2. Double precision Flops (multiply-additions)
3. Half precision Flops (multiply-additions)
4. Integer multiply-addition operations

Technologies Used

C/C, SYCL, OpenCL, CUDA, HIP, Intel DPC Compatibility Tool

Repository

https://github.com/ekondis/mixbench