Let me clarify that here I am obviously talking about accuracy in terms of the discretization error of the numerical solution, not the model error. Just like a Monte Carlo simulation needs as many optimizations (and iterations) as one can afford in order to get an acceptably converged result, the same holds for finite difference/element methods (FDM/FEM) - just replace iterations with grid/mesh size.
Setting up the PDE-based TARF pricing engine properly is not trivial in its implemention and testing its convergencge properties with flat volatility should be the first step. But of course in most applications a volatility model is necessary. So in this post I will just take the engine I built for the previous post and instead of flat volatility use a (space and time-dependent) local volatility function a la Dupire. This is a pretty easy upgrade. After all the engine's building blocks are a bunch of individual 1-D BS PDE solvers and all we need to do is allow the volatility to be a function, nothing else has to change. The only question is how much would this affect the computational efficiency of the engine. As it turns out, not much at all, provided your local volatility surface is not discontinuous and too wild (and you cache and re-use some coefficients between the solvers).
Since I do not currently have access to market data, my brief tests here will be based on a sample market implied volatility (IV) surface sourced from  :
Local volatility surface construction
For the purpose of the present test I used 300 x 400 points in the strike and time dimension respectively (shown below). Most of the points in the strike dimension are placed around the spot. The above LV calibration procedure takes about 20 milliseconds. This surface is then interpolated (bi-)linearly by each solver to find the required local volatility function values on the pricing grid points. As for the quality of the calibration, using the same LV-enabled 1-D FD solvers making up the TARF pricing engine to price the market vanillas, produces an almost perfect fit in this case (IV RMSE about 0.001%).
I will follow this up by introducing some sort of stochastic volatility as well, still aiming to keep the valuation time in the milliseconds (but we'll see about that).
 Hakala, J. and U. Wystup, Foreign Exchange Risk, Risk Books, 2002