I am not sure what you mean by P50, the BOLD response is much slower than the responses measured by EEG, there won't be any meaningful BOLD response at 50ms. Also on a clinical 3T, a single EPI image will take at least 50ms to acquire. Whole brain coverage will take anywhere from a few hundred milliseconds to a few seconds, depending on the desired spatial resolution. So a typical BOLD fMRI experiment can't resolve signal variations on the time scale of 50ms.

It is usually not meaningfully directly compare signal intensity of images from different sessions and especially from different scanners. The actual signal intensity can be affected by many things, RF coil design, placement of the subject within the RF coil, receiver gain, scaling factors applied by the scanner, bias correction methods applied either by the scanner or in post processing, etc. For BOLD fMRI, the more meaningful thing is to look at the percent signal change of a stimulus correlated signal.

MRI scanners normally save images as DICOM files which will be with 12-bit (0-4095) or 16-bit (0-65,535). I am not sure about other vendors, but Siemens by default saves data as 12-bit, so after doing the image reconstruction, the data needs to be scaled to fit within this dynamic range. It is not straight forward, but you can force a Siemens scanner to save the data as 16-bit, so then a different scaling factor is used. It is also possible to force 16-bit as a default within the pulse sequence source code, on a Siemens scanner, which has been done in the Minnesota multi-band sequences that many people use.

Images with signal intensities around 20,000, were most likely saved by the scanner as 16-bit DICOM files, though it is possible that the data was scaled after it was converted to NIFTI. Images with signal intensities around 500, most likely were saved as 12-bit DICOM images, though the scaling factor applied by the scanner, would have been less than optimal.

Hope this helps.

I second Allen's response that raw MRI signals (including BOLD fMRI) are generally arbitrarily scaled and do not have a direct quantitative meaning. For this reason, percent signal change is more meaningful in BOLD studies (although it too is not a direct measure of a single physiological process).

One thing I would add is that the absolute value does start to become meaningful in the context of signal-to-noise ratio (SNR) or temporal SNR (tSNR). tSNR is one of the most important factors for predicting your ability to detect a significant signal change. If it is too low, noise in the images will dominate, and you will need to perform many more runs or subjects to achieve a desired statistical significance. For instance, if you are acquiring BOLD data on the same scanner with the exact same sequence, and with one set of parameters (e.g., voxel size, TR, etc.) you get intensities that are 100x greater than for another set of parameters, these two data sets will likely have very different SNRs (and probably tSNRs).

For more on these SNR relationships, you can check out papers from Christina Triantafyllou et al., e.g., Article Comparison of physiological noise at 1.5 T, 3 T and 7 T and ...