About TextMesh Pro's SDF generation presision Problem

Just as the above picture, we generate sdf atlas with params: point size 64x64,SDF32...and when compare the result(the right one) with the original font(the left one), it seems that all the corner part were smoothed which make it different form the original one.

Then we set bigger point szie until to 128 x 128, the result is almost the same with the original one.

So, if we want to get a accurate font from the original font design,we must set point size so large, and in that case the sdf altas will take a big part of the runtime memeory which is not practical in our game.

Finally, we want to know if the sdf generation algorithm can be improved according to above problem or imporving point size is only the solution to keep the original font shape information complete?