We collect a global data set of PKIKP and PKiKP phases recorded by the Global Seismic Network and many regional seismic arrays to study seismic structure in the top of the Earth's inner core. The PKIKP and PKiKP observations show different characteristics between those sampling the eastern hemisphere (40¿E--180¿E) of the inner core and those sampling the western hemisphere (180¿W--40¿E). PKIKP phases (1) arrive about 0.4 s earlier than the theoretical arrivals based on Preliminary Reference Earth Model (PREM) for those sampling the eastern hemisphere of the inner core and about 0.3 s later for those sampling the western hemisphere (131¿--141¿); (2) bifurcate at smaller epicentral distances for those sampling the eastern hemisphere, compared to those sampling the western hemisphere; and (3) have smaller amplitudes for those sampling the eastern hemisphere. Waveform modeling of these observations suggests two different types of models for the two hemispheres of the top of the inner core, with a model in the eastern hemisphere having a P velocity increase of 0.765 km/s across the inner core boundary, a small radial velocity gradient of 0.000055 (km/s)/km, and an average Q value of 250, and a model in the western hemisphere with a P velocity increase of 0.633 km/s across the inner core boundary, a radial velocity gradient of 0.000533 (km/s)/km and an average Q value of 600. The hemispherical difference of seismic structures may be explained by different geometric inclusions of melt and/or different alignments of iron crystals with anisotropic properties in both velocity and attenuation. We speculate that this large-scale pattern of seismic heterogeneities may be caused by a large-scale heat flow anomaly at the bottom of the outer core and/or different vigorousness of convection in the top of the inner core between the two hemispheres.