That's primarily because Vulkan is a very low-level, but still portable, API. Different GPUs with different architectures provide memory heaps with different fu

That's primarily because Vulkan is a very low-level, but still portable, API. Different GPUs with different architectures provide memory heaps with different functional and performance characteristics (device local yes/no, host visible yes/no, host cached yes/no, host coherent yes/no), often even within the same device. For example, my Radeon RX 550 has access to 3 heaps: 1.75 GiB device local, 256 MiB device local + host visible + host coherent, 3 GiB host visible + host coherent + [optional] host cached. The last heap can be allocated either host cached or uncached; access from the GPU is presumably faster if it's treated as uncached. This gives the maximum amount of control over the performance/convenience tradeoff. If you want more convenience, then you can use a library or engine that handles the details for you. It's the same for CUDA: you get the convenience of malloc(), but you potentially lose a bit of performance.