To best utilize the compute cycles available on the devices of a heterogeneous platform, it is important to identify the tasks that are compute intensive and that can benefit from parallel execution. Consider an application that executes solely on a CPU, but there may be some tasks suitable to execute on a GPU. This can be determined using the Offload Modeling perspective of the
Intel Advisor estimates performance characterizations of the workload as it may execute on an accelerator. It consumes the information from profiling the workload and provides performance estimates, speedup, bottleneck characterization, and offload data transfer estimates and recommendations.
Typically, kernels with high compute, a large dataset, and limited memory transfers are best suited for offload to a device.