The present work reports progress in the development of our heterogeneous simulation technology for supercomputer simulations of turbulent flows based on high-accuracy schemes for unstructured meshes. The core multilevel parallel algorithm is implemented by means of MPI, OpenMP and OpenCL frameworks, allowing for co-execution on CPUs and GPUs of various vendors on hybrid cluster systems. The significant speedup we are used to on GPUs makes computing on CPUs rather dull and frustrating. Therefore, we are trying to port all new key features to GPUs as soon as possible. However, it turns out to be not so easy to obtain an efficient enough implementation that does not degrade the high performance of the core algorithm. In this paper, several new features for scale-resolving simulation of turbulent flows are considered. These include a synthetic turbulence generator for hybrid RANS-LES approaches, a novel robust modification of a hybrid RANS-LES approach. Optimization techniques and implementation details are presented with a comparative performance evaluation on a base of demonstration simulations. The performance and memory consumption issues are considered. Some of the solutions include using mixed single and double precision floating point formats, and certain management of GPU memory buffers. To be more specific, for instance, mathematical functions in double precision are extremely resource-intensive on the GPU, but there are a lot of them in turbulence modeling functions, which ruins performance. Here we show how this problem can be mitigated.