Energy consumption of Graphic Processing Units with respect to automotive use-cases

With the introduction of API's like CUDA, Stream+ or OpenCL, modern Graphics Processing Units (GPU's) can be easily employed for general purpose computing. Plus, their comparatively low price per GFLOP makes them interesting candidates for coprocessors in future embedded Electronic Control Units (ECUs). Yet, as car manufacturers thrive to reduce the Thermal Design Power (TDP) of each and every ECU within the car, and GPUs are unfortunately commonly known for their high energy demands, it is important to further investigate the power consumption of GPUs for typical automotive applications. Thus, we measured the energy consumption of General Purpose GPU applications within the driver assistance and the multimedia domain, examined a prototypical implementation of a novel 3D navigation application and had a look at applications that make use of the GPU for H.264/AVC video decoding. These investigations show that the few existing prediction models for GPU power consumption are rather too simple for automotive purposes and have to be evolved in a future step.