Last Friday was our last day with the trashcan Mac Pro, it still performs very well but, and was glorious when working with Intel PSXE but we moved to CUDA/SYCL/Linux while in a couple of years is very likely we could fully develop HPC compute applications from macOS, this is not the present state, to keep our edge from about a year ago we began to migrate our projects both to linux-based PSXE/codeXL and to CUDA (before we only developed accelerators for x86), our experience wasn't as smooth as it theoretically should have been (hpc vendors on easy multi platform development still overrates their arguments).
The part related to PSXE/codeXL native development wasn't difficult as log your IDE run local the same machine you compile/debug, for python/jupyter as long you can install docker it's somewhat easy, but then comes CUDA, it's tricky to setup it's development environment in Linux, it's time consuming, and the IDEs are also a mixed bag, theoretically you can code CUDA from any notepad app as with every c/c++ but debugging, code completion, insights etc are another question, the Best route (and maybe only practical) is to code/debug thru nsight (eclipse) ide, even you can do remote development as long you have eclipse at both ends, alternatives as clion or vs code, are not ready, remote development with PSXE/codeXL thru eclipse works as good as local in macOS, and python/tensorflow thru jupyter or vscode are really enjoying, but everything took a lot of time, but finally we have a productive workflow with mostly "remote" coding from the Mac, w/o macOS restrictions, it's not the same and the worst part is your target maybe won't run ever in a Mac, neither can integrate into a Mac app, but for those in HPC is nothing to care of, so until there comes a Mac pro with nVidia/Cuda SYCL TF2 support, i have nothing to do with a Mac pro an iMac even a Mac mini is up for us . We have a shared compute sever built around a 16core threadripper and a gp100 GPU, we study to try an Zen 3 with an Radeon instinct mi50 later once SYCL 2 on ROCm is good enough.