Should we belive declared high XNOR speedups?

[MaratZakirov]

It is well known that both matrix multiplication and convolution operation (which are some kind of equivalent, convolution may be viewed as MM) are not straightforward. For example cuda MM uses Strassen algorithm conv2d(3x3) uses Winograd algorithm even bigger conv2d (9x9 for example) would use FFT. All methods of fast multiplication uses partial recurrent sums as essential component so XNOR operations could not be applied to them. Taking this into account it is obvious that XNOR will loose its advantage in speed staring form 6x6 in case of ideal implementation. What do think?