To evaluate the accuracies, the interpolations were made at each parameter point that we have. 
An RBF fitting was performed for every spectrum. A document is in preparation. Target value of the accuracy is <1%.
 
Model templates
All the AMBRE templates with 6000 K < Teff < 8000 K

• 2590 models of the plan-parallel model (p-model)
• 2125 models of the spherical model (s-model)

Computer
3.8GHz CPU, 62GB memory, CentOS7, 7 cores are used.
 
Processing time
~10 days for 2590 p-model spectra; ~5 days for 2125 s-model spectral. On average, 5 min for p-model, 3 min for s-model. Most of the interpolations takes only ~10 sec. Interpolations which use many neighbor spectra take much time. 

 
RMS of residuals
For each interpolation spectrum, a residual is calculated as a function of wavelength.
Residual (%) == (Interpolation - Reference) / Reference * 100, 
where Reference is an original AMBRE spectrum.
Overall, the accuracies is good enough. On average, they are 0.03 (/-0.03)% and 0.02 (/-0.04)% in p-mode and s-model, respectively. I found that accuracies become worse a bit at lower Teff and high metallicity in both p/s-models.
 
P-model






 
S-model






 

P and S-models
Both models reproduce with high accuracies. The results show that the accuracy of p-model is systematically higher than that of s-model. This is because the parameter grid of the p-model templates is dense like a regular grid while one of s-model is sparse. 

The results of the evaluations was provided and a code for this work was uploaded at github.