Harald Welte wrote:

[...] the big difference is that the mobile has been calibrated to (if I remember) 2dB accuracy,

The GSM 05.05 spec indeed sets a tolerance of 2 dB under normal conditions or 2.5 dB under extreme conditions on the mobile station's maximum power output level for each band. These tolerances get looser for the lower power control level, up to 5 dB under normal conditions or 6 dB under extreme conditions for the lowest PCL.

However, as someone who is intimately familiar with the actual calibration procedures performed by MS manufacturers (just did it myself less than 48 hours ago for a new batch of boards), I can tell you how it is actually done. There is a table of target power levels in dBm, and the calibration station goes through the full set of allowed PCLs (5-19 for EGSM and GSM850, or 0-15 for DCS and PCS) and steers the Tx power output (as measured by the test instrument, usually R&S CMU200) for each PCL to the dBm number in the table.

But there is a hack in this setup which most MS manufacturers fail to disclose: with all GSM MS RF designs I have laid my hands on so far, including the one I've inherited from Openmoko (OK, I admit that all of my experience is with old stuff, not any of the recent stuff from MTK or Spreadtrum, but don't blame me, blame MTK and Spreadtrum for not publicly releasing their full chip docs and reference firmware source code), the MS hardware is oftentimes not actually capable of putting out the maximum power output prescribed by the spec (33 dBm for low bands or 30 dBm for high bands) while staying in the linear range of the PA (the range in which the relation between the APC control voltage and RF Vout is mostly linear), and thus the calibration procedure has to be fudged. In the case of Openmoko, they set the calibration targets for the highest power levels to 31.8 dBm and 28.8 dBm instead of 33/30 (for low and high bands, respectively), the next PCL down is set to 30.5/27.5 instead of 31/28, and then the proper official numbers from the spec further down. I have to do the same on my current GSM MS boards: when I tried putting out the full 33/30 dBm at the maximum PCL as the spec calls for, I can usually hit it if I crank the PA really high, but then it goes out of its linear range, and lacking better RF kung-fu, I have to assume that it's bad..

The maximum Tx power level fudging done by manufacturers in my or Openmoko's position is 1.2 dB, which is less than the spec tolerance of 2 dB, thus manufacturers like OM who did this fudging had no problems passing type approval tests, but I really wish I could find someone with more professional GSM MS design and manufacturing experience who could explain what actually happens, why we are not able hit the highest power levels while staying in the PA's linear range (the PA datasheet says it can put out 34.2 dBm in the low bands and 32.0 dBm in the high bands), and how the RF hw design might be improved so we can put out the highest power levels without fudging.

Aside from the just-described fudging, the actual calibration tolerance (the maximum error between what the calibration procedure aims for and what actually comes out as measured by the CMU200) is less than 0.7 dB - and those ~0.5-0.6 dB errors happen mostly toward the lower power levels where the spec tolerances are the loosest; the maximum power level is calibrated to the shifted target number to better than 0.1 dB.

M~