I found this document finally. it says thatpaper:
"Efficient GPU-Based Texture Interpolation using Uniform B-Splines"
http://www.mate.tue.nl/mate/pdfs/10318.pdf
Which states:
It should be noted that there are some precision issues associated with the hard-wired linear texture interpolation. When, e.g.,an eight-bit texture is filtered, most people would expect that first the neighboring texture knots are queried, casted to floating point, and then weighted and added. This is, however, not the case; the texture knots are first weighted and added, and then casted to floating point, which limits the precision to the least significant bit of the texture data format [Ruijters et al. 08], as is illustrated in Figure 4. As a consequence, higher accuracies can only be obtained by using larger texture words, and thus at the cost of texture memory consumption.
A further precision issue of the linear texture interpolation is caused by the fact that the accuracy of the texture coordinates is limited to a fixed-point format with eight bits of fractional value [NVIDIA 08]. This means that there are only 254 discrete coordinate positions between two texture knots, as shown in the zoomed graph in Figure 4, which is especially of interest when the knots are far apart...
The two references it uses 8 bit fixed point for interpolation between pixels, even for floating point. That was 2008 so not sure if things have changed, but at least here is some solid documentation.has are:
[NVIDIA 08] NVIDIA Corporation. “Linear Filetering.” In NVIDIA CUDA Compute Unified Unified Device Architecture: Programming Guide, Appendix D.2, 2008.
http://www.uni-graz.at/~liebma/CUDA/NVIDIA_CUDA_Programming_Guide_2.0.pdf
and:
[Ruijters et al. 08] Daniel Ruijters, Bart M. ter Haar Romeny, and Paul Suetens. “Accuracy of GPU-based B-Spline Evaluation.” In Computer Graphics and Imaging (CGIM), pp. 117–122. Calgary, AB, Canada: ACTA Press, 2008. http://www.dannyruijters.nl/docs/GPU_AccuracyBSpline.pdf