A global distribution of the thermospheric total mass density at 400 km altitude is derived from the high-accuracy accelerometer on board the CHAMP satellite with good temporal and spatial coverage. It shows two interesting features. One is the anomalous distribution at low latitudes. Instead of maximizing at the dayside equator, the thermospheric density shows maxima at about 20¿--25¿ geomagnetic latitude on both sides of the equator between 10 and 20 magnetic local time. This latitudinal distribution resembles fairly well the equatorial ionization anomaly, thus indicating strong magnetic control of the thermospheric mass density via ionosphere-thermosphere coupling. The thermospheric density shows a secondary maximum at the nightside equator shortly before midnight, in reminiscence of the well-known thermospheric midnight temperature maximum. Another feature to notice is that the thermospheric density is highly structured at high latitudes, with localized density enhancements possibly related to Joule/particle heating. This structure is, however, different from the cellular structure recognized by the National Center for Atmospheric Research thermosphere general circulation model at altitudes of 120--300 km. This indicates that the exact cellular structure may not necessarily extend to 400 km. A comparison between observations and the Mass Spectrometer Incoherent Scatter 1990 (MSIS90) model predictions shows that although the model describes the general structure of the observed density reasonably well, it misses the double peaks at low latitudes completely. This causes an underestimation of the total mass density by 15--20% in the crest region. At high latitudes an underestimation of 20--30% occurs in the midnight sector and the cusp region. Outside these localized areas the agreement between observations and predictions is quite good, with only ~5% difference on average under quiet geomagnetic conditions.