One of the key components to delivering the required performance for the Deep Learning (DL) applications are the hardware accelerators. VEDLIoT focuses on developing new dedicated hardware accelerators tailored explicitly towards specific applications requirements. However, the software ab-straction layers that have helped the independent development of both software and hardware in the past cannot be used any longer to achieve the best performance and efficiency for the most demanding workloads. The solution is to focus on hardware-software co-design. In VEDLIoT, four different types of DL accelerators are explored: (1) existing off-the- shelf; (2) statically configured; (3) dynamically reconfigurable; and (4) fully simultaneous co-design accelerators.