How does oneAPI/SYCL compare to CUDA? We certainly need an alternative to OpenCL, but every day, I can't help but notice the widening gulf between CUDA and any other GPGPU API out there.