Somewhat off topic from the article, but I'm curious about GPU vs CPU computing here:

So it looks like meshroom disables a feature (or features) when not having a Nvidia CUDA-enabled GPU, rather than e.g. still have the feature, but have it be slower.

Any idea, if it were computed (or emulated) on CPU, how much slower it would be? (like 10x slower, 100x slower, 1000x slower, even more?!)

And also, ignoring speed but instead looking at implementation effort, if it's feasible to emulate on CPU at all, or is GPU really so different for this that it's too much work to implement the same thing for a general purpose processor? At least matrix multiplication libraries (the main thing GPUs can do well afaik) exist for CPU too...