I need to develop an application that renders surface generated by CUDA. I looked at CUDA-Vulkan interoperation example:
It runs fine on a single NVIDIA GPU. Now I’d like to run it on multiple GPUs.
So my questions are:
- Is it possible at all? Because as I figured out both Vulkan and CUDA should be initialized using the same device. I know it’s possible to distribute kernel execution between multiple GPUs using CUDA streams. In this case each GPU stores some part of vertex buffer to generate data to. But is it possible to transfer those buffers to Vulkan without GPU-CPU-GPU round-trip as well as with single GPU?
- If p. 1 is possible, can I benefit utilizing NVLink interface?