Abstract: Intersection computation is the most expensive part of the ray-tracing while BVH is the main acceleration structure.To improve the construction speed of BVH,a new BVH construction method based on dynamic task scheduling and threads optimization of warp is presented and it is optimized according to some main GPU architectures.The algorithm performed multiple rounds of bottom-up traversal over the BVH based on the SAH value.At the beginning of each cycle,we assigned the task according to the global variables that recorded the progress of the task if the current thread was idle,otherwise we continued to traverse the BVH.We reconstructed the treelet that used the node as treelet root when the node met the conditions during the traversal.At this stage,we leveraged 32 threads in the same warp to reconstruct the treelet collaboratively and we could adjust the number of parallel tasks according to the number of treelet leaves.According to the experiments on some classical scenes,our solution can be faster while the number of triangles is greater than one hundred thousand and can be about 47%faster when the number of triangles is above one million in the optimization stage of AtrBVH,thus increasing the overall speed by 25%,in the case of BVH construction quality being the same.