@marc-alonso I think the drawback about their approach is about ReLU evaluation, since slot-wise multiplications are more difficult (check this recent thread)
Edit: ReLU was ok (is indeed computed after the CtoS step, on the re-encoded ciphertext), I think the way they consider convolutions as polynomial negacyclic convolutions is cool, it indeed avoids to run multiple multiplications and rotations. I guess the main issue is about packing LWE ciphertexts (which needs many rotation keys), but as soon as you have enough RAM, it is ok.