When possible, most DL frameworks take advantage of Nvidia's specialized CuDNN libraries, which often provide a 10x+ speedup (and are obviously not available via WebGL). So at least on the latest Nvidia cards, you will likely see a 10x slowdown, and probably even more.