A neural network application may be provided in step 601. For example, the architecture and parameters of the neural network may be provided. The parameters may be weights, biases, and types of input/output data of the neural network. The code of the neural network application may be received in step 602. The code may, for example, be in a programming language such as Tensorflow, Keras, pyToch, etc. The received code may be adapted in step 603 in order to be executed by the hardware accelerator. For example, the code may be lowered to an intermediate representation (IR) level that can be executed by the hardware accelerator. Replication units may be created in step 604 from the first computation unit according to different precisions. The parameters of the neural network may be forwarded in step 605 to the replication units so that processing of the application may start at the first computation unit and the replication units. It may be determined in step 606 if the controller has found the replication unit that can deliver the result of the operation. If it is determined in step 606 that the controller has not found the replication unit that can deliver the result of the operation, the result of the first computation unit may be used in step 608. If it is determined in step 606 that the controller has found a replication unit that can deliver the result of the operation the result of the found replication unit may be used in step 607.