Crowd　simulation　has　been　very　significant　and　widely　applied　in　many　fields　of　virtual　reality.　How　to　greatly　increase　the　efficiency　and　realism　becomes　the　focus　of　the　research　about　crowd　simulation.　In　this　paper,　a　real-time　crowd　simulation　system　is　designed　on　parallel　architectures.　The　system　is　applied　to　simulating　the　evacuation　of　massive　crowd　in　an　Olympic　venue　and　then　experiments　are　done　to　compare　and　analyze　their　different　simulation　efficiencies　on　different　architectures.　The　best　simulation　performance　comes　from　the　hybrid　architecture　in　combination　of　CPU　and　GPU　with　parallel　computing　on　multi-core　CPUs　and　the　it　can　reach　52　fps　when　the　number　of　crowd　is　10000.　GPU　architecture　can　be　used　in　the　crowd　simulation　system　with　few　data　transmission　times　between　CPU　and　GPU,　branch　statements　and　texture　look-ups　better.　©　2013　Binary　Information　Press.