Hi all,
I’m very new to the GPGPU and Programming. I’m interested to know if Hypothesis testing(Statistical) like One sample Kolmogorov-Smirnov test (K–S test) and Levene’s test could be implemented in GPGPU using CUDA? If so what will be the limitations?
Sure. The main burden of the KS test is sorting, and GPUs can certainly do that.
Levene’s test is different and simpler, and its computation is effectively dominated by a large sum with some simple multiplies… something a GPU can do quickly and in parallel.
There wouldn’t be any practical limitations, though if you had more than 100M data points for the KS test you might have problems loading all the data onto the GPU for sorting.
The other considerations are in overhead… will the GPU give you a net speed win? For small number of points (or few parallel tests) the setup time would likely dominate computation, and it’d be better to do on the CPU.
Levene’s test itself may still be better on the CPU just because it’s so simple the data transfer may dominate total time. This wouldn’t apply if you’re doing other things with the data on the GPU or if the data itself were created by another GPU kernel.
Hi SPWorley, Its great to hear from you. The information that you have shared is of great help. Could share any articles or other materials that you have come across regarding this?