The media noise in magnetic recording is strongly signal dependent, By modeling the medium as a collection of independent magnetic grains, the noise variance can be shown to vary as "1-m(2)", where m is the normalized magnetization, An approximate detection bound is developed for a magnetic recording channel dominated by such media noise. Four different coding techniques are compared: 16/17(0,k), 2/3(1,7), 5/6 MTR, and 8/10 trellis, It can also be deduced that, for constant areal density, low bit aspect ratios give better effective SNR Finally, it is shown that in the limit of very sparse grains, the channel capacity approaches 1 bit/grain.