Performance Comparison of NVIDIA accelerators with SIMD, Associative, and Multi-core Processors for Air Traffic Management

Basic tasks for Air Traffic Management (ATM) will be implemented using NVIDIA's CUDA language on an NVIDIA device and compared to the performance of SIMD, associative, and multi-core processors doing the same tasks; CUDA is a parallel computing platform and application programming interface (API) model created by NVIDIA. To do this, we create a simulation of an airfield with constantly moving aircrafts. The basic tasks that will be used in the evaluation are: tracking and correlation, collision detection, and collision resolution. These are the most compute-intensive of the ATM tasks, so they will give us a good measure of the capabilities of the NVIDIA device. In our previous research, it was shown through a performance comparison between an associative SIMD processor and a multi-core processor that the associative SIMD processor can execute these tasks in linear time without missing a single deadline, whereas the multi-core processor required significantly more (exponential to be precise) time. Additionally, the multi-core processor regularly missed a large number of deadlines. Our goal in this paper is to determine whether we could get SIMD-like results using a CUDA implementation of the basic ATM tasks on NVIDIA accelerators. This will allow the novel use of increasingly popular NVIDIA accelerators for ATM tasks in place of a SIMD processor that was found to be a good fit for these tasks in the previous research. Our ATM implementation and evaluation on three different NVIDIA accelerators shows that all three NVIDIA accelerators can provide a SIMD-like implementation for ATM. Curve-fitting with MATLAB shows that the performance of NVIDIA accelerators increases only slightly faster than a linear graph.