The Generation Challenge Program global composite germplasm collection of 3372 wild and cultivated accessions includes 280 elite breeding lines and improved cultivars from public breeding programs, 250 germplasm lines of Chinese origin provided by CAAS, 64 wild accessions, and >3000 landrace accessions. These were selected from previously defined core collections, or for resistance/tolerance to various production constraints and/or variation in other agronomic and quality traits. A set of 48 sorghum simple sequence repeat (SSR) markers distributed across all ten linkage groups was chosen following preliminary analysis of 48 diverse genotypes with 104 available SSRs complemented by new markers from CIRAD and ICRISAT. Marker genotyping responsibilities were shared across collaborating labs (CAAS, CIRAD and ICRISAT). Diversity analysis was performed for 3367 accessions genotyped with 39 SSR markers, detecting 782 alleles (20.1 alleles/marker). Breeding lines and wild accessions clustered separately from landrace accessions, which exhibited structure explainable by geographic origin. Landrace population substructure was further characterized within racial groups (five basic races and ten hybrid races). Race bicolor showed little evidence of population structure, which is congruent with the hypothesis of it being the original domesticate. Race kafir (largely from Southern Africa) formed a distinct group and accessions of the durra, caudatum and guinea races each formed four distinct subgroups related to their geographic origins. The margeritiferum group within the guinea race formed its own cluster, suggesting an independent domestication event. Intermediate races followed similar patterns. Based on this diversity analysis, we have defined a reference subset for allele mining. (Texte intégral)