Why is larger GPU memory required for GridInterpolationKernel in 3-dimensional space?

[hanyang-hu]

I'm using DKL combined with KISS-GP and monitoring GPU usage during my experiments. With the same dataset, setting the feature dimension to 2 results in consistently low GPU memory usage. However, when I increase the feature dimension to 3, I observe a significant increase in memory consumption. Is this behavior expected, or might there be an underlying issue? Any insights would be appreciated.

Just to complement: for the 2D case, my grid size is 100 (so supposedly 10000 inducing points); for the 3D case, my grid size is 16 (so supposedly 4096 inducing points).