As　a　deep　residual　network　model,　Resnet50　has　significant　practical　significance　in　image　classification,　target　recognition,　and　image　semantic　recognition.　In　this　paper,　Nvidia　RTX　4090　GPU　is　used　to　conduct　detailed　performance　testing　and　bottleneck　analysis　for　Resnet50　training　and　inference,　including　specific　calculation　delay　and　data　processing　delay　under　different　batch　sizes.　In　order　to　verify　the　overall　acceleration　effect　of　Resnet50,　we　use　two　optimization　methods　on　the　basis　of　GPU　computing　acceleration:　the　first　is　to　use　mixed　precision　to　improve　GPU　training　and　inference　efficiency,　and　the　second　is　to　use　DALI　to　optimize　data　preprocessing　and　reduce　data　loading　delay.　The　experimental　results　show　that　when　the　batch　size　is　256,　the　mixed　precision　is　improved　by　about　90%　compared　with　FP32,　but　the　overall　performance　improvement　is　not　obvious.　When　using　mixed　precision　and　DALI　for　GPU　computing　and　data　loading　optimization　at　the　same　time,　it　can　bring　1.4　and　2.5　times　improvement　in　the　overall　performance　of　training　and　inference.　The　experimental　results　show　that　only　using　the　mixed　precision　can　not　improve　the　overall　computing　efficiency　of　the　system,　and　the　data　loading　time　cost　frequently　limits　the　end-to-end　performance.　Therefore,　only　by　optimizing　GPU　computation　and　data　loading　delay　at　the　same　time　can　end　users　get　a　significant　speed　increase.　This　paper　is　of　great　significance　to　evaluate　and　improve　the　computational　acceleration　performance　of　GPU-based　deep　neural　networks.　©　2024　SPIE.