Solving partial differential equations (PDEs) with physics-informed neural networks (PINNs) has been popular since 2017. A search on Google Scholar revealed a rapid growth from 60 results in 2017 to 3,340 results in 2021. Such a blooming raised our interest in PINNs' feasibility in practical computational fluid dynamics (CFD) problems. In this work, we investigated data-free PINNs' convergence behaviors, scalability, cost-performance ratios, and the cost of non-computing tasks. We also investigated several training strategies and PINNs' capability to solve flow problems with instability. We confirmed that PINNs exhibit a qualitative converging behavior with respect to model complexity and identified that adding hidden layers is more efficient than adding neurons to increase accuracy. However, PINNs lack a quantitative metric to measure the model complexity, which makes cost-performance estimation difficult for engineering purposes. We also found that the accuracy-loss relationship becomes more nondeterministic when the model complexity increases, which may also hurt its engineering use. On the other hand, our results show that PINNs have good weak scaling in general, which helps scale per-batch training points when using lower-end GPUs. Nevertheless, our benchmarks also indicate that PINNs are generally insensitive to the number of per-batch training points. Increasing per-batch training points hurts the time-to-solution for no apparent benefit, while PINNs' time-to-solutions were already found to be several orders slower than traditional CFD code. Moreover, our results show that PINNs could not predict vortex shedding in a cylinder flow at Re=200. The unsteady data-free PINN solver converged to a steady-state solution. Another data-driven PINN solver also transitioned to the steady-state solution beyond the range of the given observed data. The Koopman analysis shows that the PINN is dispersive and dissipative. These results cast doubt on both data-driven and data-free PINNs' capability to solve flow with instability. We also compared the differences and similarities between data-free PINNs and conventional numerical methods. We found PINNs can be deemed as a special case of least-square finite element methods but with a more complicated nonlinear approximate solution. PINNs were relatively new, and many numerical properties were still unknown. Our work identified several challenges that need to be tackled for PINNs to be feasible in practical engineer...