There is no reason that SDDs need to be round. In fact, many aren't. Custom SDDs could be designed to enhance throughput at reduced resolution with better matched shape.
You are right, I could not find an example previous, Today I came across this so I am going to leave the link below:
https://www.hitachi-hightech.com/hhs-us/product_detail/?pn=ana-vortex-me3The detector itself and form factor is irrelevant, but existence of such shape of SDD is promising for our cause.
We would need solid state detector with 4 cm length, where width should be 2-3 mm max:
1) That is more less the 2R of cylinder of active volume of G(F)PC, and so the solid state detector should contain not larger width to not introduce the worsening of spectral resolution achieved by diffraction.
2) while it would be possible to add the aperture in front of solid state detector, the blanked portion of detector would still require cooling - narrow solid state detector would require less cooling.
With PIN detector You can forget about sub 1µs dead times, it requires few to tens of µs shaping time to get something sensible. SDD is the only detection technology currently allowing to gets anything sensible with sub 1µs shaping and going > 2Mcps in input count rate.
Interestingly AMPTEK recently added to its portfolio thicker SDD (1mm) which has extended efficiency past 10keV (That would be really good for LiF range), with no impact for lower energies.
But I really wonder if G(F)PC could also not go easily to >2Mcps with better charge sensitive preamplifier and faster Shaping amplifier.
What about debris from moving parts (from this rotating screw moving the assemblage of turret) - the construction and maintenance of G(F)PC is much more simple and possible to DIY (basically changing counter windows), where fixing or cleaning of solid state detector (and It fails unexpectedly, i.e. from 8 SDD EDS know to me 5 had to be sent back to manufacturer for fixing in 4 years from installment, with downtime of more than 2 months) is requiring sending it back to manufacturer.
Downside of GFPC is the gas. But how much bottles are bothersome for machine anyway being fixed at room and not being portable? We are changing bottle of P10 gas which supplies gas to 5 spectrometers of SXFiveFE about every 12 months. It is important to check for gas leaks at junctions, as that can significantly slow down gas bottle getting empty. Is it really so bothersome to change gas every 12 months? It can be bothersome if bottles are in different room with worse air conditioning, and bottles being attached to wrong (thermal unstable) wall, the thermal seasonal temperature variations will show up in PHA. However if bottles are at the same room (as probe), situated by inner wall (thermally stable) and room has good AC, then the observed PHA shifts due to seasonal temperature, pressure and humidity will be visible at some small degree only on low pressure GFPC.
PHA seasonal shifts (like atmospheric front pass over geographic area/campus where probe resides at summer) could be overcome with automatic bias gain readjustment in software. It could be enough to add temperature humidity and pressure sensors to initiate such readjustment in case of condition changed. But wait a moment - from my experience SDD are also affected by atmospheric front passing. I saw many times that EDS needed re-calibration after that, so moving to solid state is not going to fix that problem, but getting precise dehumidifier for room would be more beneficial.
The only real benefit moving away from G(F)PC I see is for FEG machines, where silent little Argon leak from counter can silently contaminate the simple diode ion pumps and then unexpectedly do boom! - Your pricey FEG tip just got ruined. But actually even there after experiencing such events once, now I am confidentially sure to be able to detect such event by monitoring vacuum of both ion pumps, where secondary pump will hint of that very early enough to be able identify and fix the issue.
Finally, do we really need sub-µs shaping time (dead time)? SDD EDS have that... but there live time is estimated, not directly measured compared to G(F)PC with fixed time signal blanked/fixed dead time. The first is worse estimated with high impulse count and with increasing dead time (pulse density), where GFPC counting with fixed blanking dead time has no paralysable behaviour even with very high count rates (I know that somewhere here in forum was written contrary, but that is due to wrong calculation of dead time at differential settings and unaccounted pile-up, which BTW is going to be fixed in next Cameca Peaksight 5.6). Also SDD EDS dead time estimation is really bad (imprecise in few orders compared to current GFPC counting implementation) for very short counting times. I think it is much more important to have no paralysable behaviour AND very precise dead time/ live time measurement (not mere estimation, with blocking behavior what SDD counting electronics have) – even if we get sub 1Mcps but having very precise live time estimation that would allow to estimate precisely input count rate even above 2 Mcps. I think for any quantitative work that is much more important. Telling that, I think it would be possible to improve the current WDS throughput and linearity and PHA stability significantly with not touching detector design at all and improving/replacing only analog amplification and counting electronics.
P.S. I nearly finished gathering factual pieces, and soon am going to open new thread in this forum showing currently used counting technologies and solutions by Bruker and Cameca WDS (measured stuff), AMPTEK, CREMAT published and available hardware. To go with improvements, we should first fully understand what are exact shortcomings of currently used technology (but also what are advantages), and what consequences would be by technological shift. (it not always brings in progress, but can bring unintended regress which in some usage cases can be crucial). The shortcomings in software/firmware of current technology is often overlooked, the lack of possibility to discuss the hardware (no schematics, or schematics provided to end-user with strict NDA clause) freely is one of main obstacles. I.e. there is this new thread about dead times of WDS. I could not convince anyone there why dead time is calculated wrong at widely used diff mode, and I am still not sure probeman had understood me correctly in numerous exchange of replays in PM. Contrary to that, I could easily convince there is error in diff mode dead time calculations the Cameca Software engineers by pinpointing/describing to exact place in schematics with a single e-mail, and that is going to be fixed in next Peaksight release (5.6). What bothers me a lot is that many one here are easy about of moving to much more complicated technology, before knowing where exactly current technology fails to deliver and if it is actually easy fixable.