Normally, one filters after normalization and before calculating differential expression to perform fewer tests. While this also speeds things up, the main benefit is due to not then having to correct for performing a bunch of tests that could never possibly have given significant results (this is why you need to filter carefully).

For WGCNA, I never tried filtering my data prior to use (granted, I have RNAseq rather than microarray data). Aside from faster computation (since you'll have smaller matrices), I'm not sure there's a big benefit to filtering with WGCNA.