We've had a number of Atom C2000 machines die during the past 1-2 years. The most recent one ran continuously since 2015 until 2020/05. The randomly dying machines are SuperMicro 5018A-MLTN4 with A1SAM-2550F motherboard.
Via EEVBlog we now found this simple fix. And it works!
Amazingly SuperMicro or other vendors don't seem to advertize or talk about this simple fix very much...
Achieves around 4 MHz bandwidth, which should be OK for intensity monitoring etc. This will work only for quite small powers, since the maximum DC output swing is 5V, corresponding to a 5 uA photocurrent, or around 12 uW if we assume 0.4 A/W responsivity for the photodiode.
Again I needed to tweak the TIASim simulation to make the experimental data fit the model. I now suspect this is due to capacitive loading of the op-amp output. This should be possible to test by connecting the detector directly to the Spectrum Analyzer input, and then adding successively longer cables and seeing if the response changes.