In decision trees more than half of the optimization time is spent doing sorting (sorting the set at each node to find the optimal splitting for the key at that node), in neural nets it's almost all matrix multiplies. That's where the speedup difference comes in the CPU v GPU comparison.

There are fast parallel sorting algorithms that should be able to take advantage of GPUs. Maybe they didn't implement them?

The LightGBM implementation on GPU is based on this paper: https://arxiv.org/abs/1706.08359 they use several smart techniques to make the computation faster. One is how they create histograms of features that are computed in parallel in the GPU

You need to sort it once for each dimension/predictor, after that it's log(n) to find the optimal split.

actually, the histogram-based tree learning algorithm doesn't need the sorting. The main cost of it is the memory access. The GPU version is slow (when #data is small) due to the additional memory copy cost between GPU and CPU. I think this issue can be fixed in the future.

In deep learning, the algorithm that is used, back propagation, is basically a chained matrix multiplication, which is good for the GPU structure. However, for boosted trees the optimization problem is a little bit different, there is a sorting and then a computation of a gradient and hessian. In this case, it is not as efficient for GPU than deep learning.