--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/RFFE-notes	Tue Sep 29 07:26:45 2020 +0000
@@ -0,0 +1,336 @@
+Every GSM phone handset or modem that is based on our beloved Calypso+Iota+Rita
+chipset also contains a functional section (can be a single component or a group
+of components) which we call the RFFE: radio frequency front end.  This RFFE is
+the functional block that provides frequency-filtered radio Rx paths from the
+antenna interface to Rita LNA inputs, as well radio Tx paths from the PA to the
+antenna.  The choice of particular RFFE implementation (specifically in the Rx
+direction) determines the device's GSM frequency band capabilities, i.e.,
+whether your GSM phone or modem is dual-band, triband or quadband.
+
+The least band-capable Calypso-based devices we are aware of are dual-band,
+i.e., we are not aware of anyone having ever made a Calypso-based GSM device
+that supports only one GSM frequency band.  The lowest-end configuration is
+single-region dual-band, supporting one low band and one high band, either EU
+bands or US bands.  The next step up is triband, supporting two EU bands and
+one US band or vice-versa, and the best RFFE designs are quadband, supporting
+all 4 possible GSM frequency bands.
+
+There are two particular RFFE designs that enjoy the privileged status of being
+most native to FreeCalypso: TI's original Leonardo+ quadband reference design
+and Openmoko's derivative design that is very polished, but is only triband.
+In absolute historical terms TI's Leonardo existed first and OM's version was a
+later derivative work, but for FreeCalypso the two got somewhat reversed with
+our peculiar timeline:
+
+* The Mother bought her first Openmoko Neo FreeRunner in 2011, and lots of other
+  people in the community have been playing with Openmoko devices - whereas TI's
+  Leonardo development board has always been legendary unobtainium.
+
+* Openmoko's PCB layout has been liberated (freely published) in 2015, whereas
+  TI's Leonardo PCB design remained as elusive as ever.
+
+* Between 2015 and 2017 we've successfully produced our FCDEV3B board based on
+  Openmoko's PCB layout, featuring OM's triband RFFE, and we gained a lot of
+  experience working with this RFFE as we recreated the previously lost theory
+  and tools for RF calibration.
+
+* In December of 2019 I was searching the world high and low for a real TI
+  Leonardo board so I could do some CMU200 experiments with its RFFE, but
+  instead I found the historical commercial Calypso modem module which we call
+  Tango, which is nothing less than a mass-produced version of the core of
+  Leonardo, the full quadband version.
+
+The discovery of Tango modules has finally brought TI's original quadband RFFE
+design back to its rightful place at the top of the quality hierarchy among
+Calypso GSM devices, and we are finally able to study it as much as we have
+studied OM's triband version in the preceding years.
+
+Leonardo and Openmoko RF sections
+=================================
+
+The RF sections of both Leonardo/Tango and Openmoko/FCDEV3B consist of TI's
+Rita RF transceiver, a standalone PA (RF3133 or RF3166) and the part we focus
+on as the RFFE.  The best way to illustrate the essential divisions and
+interconnections between components is with a block diagram:
+
+                                      +-------+
+                                      | RF PA |
+                                      |       |
+    +--------------- HB Tx path ------+-->>>--+------+
+    |      +-------- LB Tx path ------+-->>>--+---+  |
+    |      |                          |       |   |  |
++---+------+---+                      +-------+   |  |
+| HB Tx  LB Tx |                                  |  |
+|              |                                +-+--+--+
+| Rita         |                                | RFFE  |
+|              |                                |       |
+|       LB LNA-+----- LB Rx path ---------------+-     -+-- ANTENNA
+|      DCS LNA-+---- DCS Rx path ---------------+-      |
+|      PCS LNA-+---- PCS Rx path ---------------+-      |
+|              |                                |       |
++--------------+                                +-------+
+
+HB = high bands (DCS & PCS, 1800 & 1900 MHz)
+LB = low bands (GSM850 & EGSM, 850 & 900 MHz)
+
+The block marked as RFFE on the above block diagram is the one that differs
+between Leonardo/Tango and Openmoko/FCDEV3B, and it is this block that makes
+one quadband and the other only triband.  In the original Leonardo+ quadband
+reference design the RFFE block is a single component (Epcos M034F) that is
+called a Front End Module (FEM), whereas Openmoko replaced this integrated FEM
+with an Antenna Switch Module (ASM) plus 3 discrete SAW filters.  Both RFFE
+designs implement the following functions:
+
+* They switch the antenna interface between Rx and Tx - a GSM MS never needs to
+  receive and transmit at the same time, thus an antenna switch is used instead
+  of more expensive duplexer or diplexer arrangements.
+
+* They accept RF Tx signal from one of two separate inputs (one for LB, one for
+  HB), and they provide the necessary low pass filters to suppress unwanted
+  harmonics present in the PA output.
+
+* Rx signals are fed to the RF transceiver on one of 3 separate Rx signal paths
+  as appropriate for the band in which reception takes place (LB, DCS or PCS),
+  and for each supported band the incoming signal is passed through an
+  appropriate SAW - a very narrow frequency-specific bandpass filter.
+
+RF Tx paths
+===========
+
+There are only two RF Tx paths: one for GSM850 and EGSM (850 & 900 MHz) bands
+combined, and the other for DCS and PCS (1800 & 1900 MHz) bands combined.  Thus
+the RF Tx section of a Calypso+Iota+Rita GSM phone or modem is exactly the same
+between dual-band, triband and quadband designs - only Rx paths differ.
+
+Both Leonardo/Tango and Openmoko/FCDEV3B use PAs from the same family by RFMD,
+just different evolutionary versions: Leonardo and Tango use the slightly older
+RF3133 PA whereas OM and FCDEV3B use the slightly newer RF3166.  The two PAs
+have exactly the same controls, and they are wired in exactly the same way
+between Leonardo and Openmoko.
+
+Besides PA controls, there are two very important Tx path control signals that
+are driven by Calypso and go to the FEM or ASM: RFFE LB & HB Tx switch controls.
+Both the M034F FEM used in Leonardo and Openmoko's triband ASM have two
+baseband-controlled internal switches selecting between Rx and Tx.  When these
+switches are inactive (no voltage applied), both LB and HB RF paths are set up
+for Rx; one of these Tx switches enables LB Tx path and the other enables HB Tx
+path.  There is also a third switch present both in M034F and in OM's ASM, but
+that one is for different Rx bands as explained later in this article.
+
+RF Rx paths
+===========
+
+TI's Rita RF transceiver supports all 4 GSM frequency bands internally for both
+Rx and Tx, but in the Rx direction it has only 3 LNA inputs.  GSM850 and EGSM
+share the same LNA input, whereas each of DCS and PCS gets its own.  There is
+also a requirement that whenever reception takes place in a given band, the Rx
+signal has to pass through a SAW filter specific to that band: thus dual-band
+designs have two SAW filters in them somewhere, triband designs have 3 and
+quadband designs have 4 of them.
+
+Triband design with Rita and other similar transceivers like Silabs Aero I/I+
+is straightforward: the antenna switch module (ASM) puts out 3 Rx paths
+corresponding to LB, DCS or PCS, each of the 3 is passed through its own SAW
+filter, and then they go to the 3 LNA inputs.  SAW filters in the DCS and PCS
+Rx paths are unambiguous, whereas the third SAW filter in the LB Rx path can be
+populated in one of two ways: populating an EGSM downlink band filter produces
+a tri900 GSM device, populating a GSM850 DL band filter produces a tri850 GSM
+device.  Openmoko's triband RFFE which we've reproduced on our FCDEV3B is built
+in exactly this manner.
+
+Quadband design with the same Rita transceiver is more complicated: there need
+to be a total of 4 SAW filters, but they need to go to just 3 Rita LNA inputs.
+One can make a discrete design: use an antenna switch module that puts out 4
+switched Rx paths, pass them through 4 discrete SAW filters, and then somehow
+combine the outputs of GSM850 and EGSM filters onto the single LB LNA input.
+Huawei's MG01GSMT design appears to do something similar (except that it
+supports a non-standard band in the place of GSM850), but I don't really
+understand how this magic works in terms of PCB layout.
+
+The M034F integrated FEM that was chosen by TI Back In The Day is much neater:
+it is a quadband FEM, containing 4 SAW filters inside, but it was specifically
+designed to interface with transceivers like TI Rita or Silabs Aero that have 3
+Rx input paths - thus the magic that combines the output of GSM850 and EGSM SAW
+filters is contained inside this FEM component.
+
+We do have an M034F.pdf datasheet for this magic component (came along with
+Calypso and Leonardo docs), and the block diagram on page 6 shows the magic
+quite clearly: there is a baseband-controlled switch selecting between EGSM Rx
+and GSM850 Rx (in addition to the two usual Tx switches), this switch directs
+the low band Rx path toward one of two different SAW filters, and the outputs
+of those two filters are then joined.  The high band Rx path always goes to both
+DCS and PCS band SAW filters, and each of those high band Rx SAW filters gets
+its own output going to its own dedicated Rita LNA input.  Note the lack of a
+baseband-controlled switch between DCS and PCS in the high band Rx path!
+
+For a long time prior to the 2019-12 breakthrough I was concerned about this
+lack of a baseband-controlled switch between DCS and PCS in the high band Rx
+path: this switch is present in every triband RFFE design I am familiar with,
+including the one we got from Openmoko, and I was concerned that feeding the
+high band Rx path to both DCS and PCS SAW filters without a switch would
+introduce a 3 dB penalty into these high band Rx paths.  But when I got my
+hands on the newly discovered Tango modem modules, looked at the GMagic numbers
+written into their FFS as part of the manufacturer's production calibration and
+then independently verified them with our CMU200, all fears were dispelled: the
+Rx performance of this M034 FEM in all 4 bands is exactly the same as our
+current Openmoko-based triband RFFE, with GMagic around 200 half-dB units.
+
+RFFE switches and control signals
+=================================
+
+Both our new quadband M034F and our more mature OM-based triband ASM have 3
+baseband-controlled switches inside: two Tx switches (one for LB Tx, one for HB
+Tx) and one Rx band steering switch.  The Rx switch is entirely different
+between the two RFFE designs:
+
+* Our OM-based triband ASM has an Rx path steering switch between DCS and PCS
+  just like every other triband RFFE design I have seen, and of course being
+  triband, it has no capability of switching between GSM850 and EGSM.
+
+* Epcos M034F has a baseband-controlled Rx path steering switch between GSM850
+  and EGSM (seems unavoidably necessary in this architecture), but it has NO
+  switch between DCS and PCS - instead both high band Rx paths are "always on".
+
+All 3 RFFE switches are controlled by Calypso TSPACT signals passed through PNP
+transistors that act as inverting buffers.  The assignments of which Calypso
+TSPACT controls which RFFE switch were made by people who came before us, and
+these assignments are different (incompatible) between Leonardo/Tango and
+Openmoko/FCDEV3B.
+
+Leonardo RFFE control signal assignments (used on Tango) are as follows:
+
+TSPACT1 = Tx high bands
+TSPACT2 = Tx low bands
+TSPACT4 = Rx GSM850 band
+
+Openmoko's version (used on FCDEV3B) is as follows:
+
+TSPACT1 = Rx PCS band
+TSPACT2 = Tx high bands
+TSPACT4 = Tx low bands
+
+In the now-seemingly-unlikely event that we end up building a new FreeCalypso
+GSM device (not Tango) with M034F quadband FEM, we have two sensible choices
+for TSPACT control signal wiring.  One option would be to copy Leonardo and
+Tango, and the other option would be to use the following novel assignment:
+
+TSPACT2 = Tx high bands
+TSPACT4 = Tx low bands
+TSPACT5 = Rx GSM850 band
+
+The latter option can be made firmware-compatible with our current OM-based
+triband RFFE, by having our firmware drive the signals as follows:
+
+TSPACT1 = Rx PCS band
+TSPACT2 = Tx high bands
+TSPACT4 = Tx low bands
+TSPACT5 = Rx GSM850 band
+
+Right now this consideration is purely academic because there presently exists
+no business case for building new FreeCalypso hardware, given the existence of
+Tango modules.
+
+RF magic glue: Rx paths
+=======================
+
+RF signal paths coming out of the 3 Rx SAW filters in our current OM-based
+design or out of M034F Rx pins are differential pairs; Rita LNA inputs to which
+these signal paths go are also differential.  But these RF signal paths are not
+simple pairs of PCB traces going from one chip to the other, instead every
+existing design features "RF magic glue" matching networks in these signal
+paths, made up of inductors and/or capacitors.  These "magic glue" LC networks
+exist in a seemingly endless number of permutations: every historical design I
+got a chance to examine exhibits a slightly different LC network topology among
+the 3 Rx signal paths that are present.  Sometimes the same LC network topology
+is replicated for all 3 Rx signal paths, othertimes the same board will feature
+different LC network topologies for different Rx band signal paths.
+
+Openmoko's Rx path "magic glue" networks look neat and tidy, with the same
+topology in all 3 Rx band signal paths.  This topology consists of a series
+inductor in each trace (6 of them in total for 3 differential pairs) and shunt
+capacitor footprints both before and after these series inductors - but all of
+these shunt capacitor footprints are left unpopulated.  Inductor values are
+naturally a little different for each frequency band.
+
+Prior to the discovery of Tango modules, when I was thinking about designing
+and building new FreeCalypso hw with our long-desired M034F quadband FEM, I had
+the big dilemma of deciding what topology should be implemented for the "RF
+magic glue" matching networks.  OM's version looks very neat and tidy and it is
+known to work well with OM's triband RFFE, but would it still work well if that
+triband RFFE were to be replaced with M034F?  The answer is very uncertain,
+especially considering that the output impedance specifications given in our
+only available M034F.pdf datasheet look very weird.
+
+The discovery of Tango has saved the day in this department.  The "magic glue"
+Rx path matching networks implemented on this module are exactly the same as
+what is depicted on our available Leonardo schematics (the ones we've had since
+2011), and the fact that these modules once were a successful mass-produced
+commercial product gives a huge confidence boost to what was previously just a
+schematic drawing of completely unknown working quality.  These Leonardo Rx path
+matching networks score very low on the scale of understandability (and they
+feature entirely different topology for each of the 3 Rx band signal paths),
+but the fact that we have a working reference in the form of Tango is all that
+matters at the end of the day...
+
+RF magic glue: Tx paths
+=======================
+
+RF Tx signal paths running from the PA to the FEM or ASM also pass through
+matching networks made up of inductors and/or capacitors.  The two versions of
+most interest to FreeCalypso are once again Leonardo and Openmoko.  The version
+of Leonardo that has been mass-produced as Tango has the same T network topology
+in both LB and HB Tx paths:
+
+PA_OUT-----C1--+--C2-----FEM_IN
+               |
+               C3
+               |
+              GND
+
+LB: C1=C2=47pF, C3 is unpopulated
+HB: C1=C2=12pF, C3 is unpopulated
+
+It needs to be noted that with the cap-to-ground leg unpopulated, the T network
+is degenerate and is equivalent to a single series capacitor.  It needs to be
+understood that TI were not in the business of selling finished GSM phones or
+modems - instead their primary business was selling chips, and their secondary
+business was providing support for these chips, support that included reference
+and development boards.  Because Leonardo was a reference board rather than a
+finished product, it appears that TI designed this board with a provision for a
+possible matching network - the comment on the schematic drawing says "PA
+matching possible" - but apparently no matching was needed, hence a degenerate
+T network was populated.  It is certainly a little strange that the makers of
+Tango kept this degenerate T network and did not replace it with a single
+series capacitor, but what do we know...
+
+Openmoko used pi networks instead:
+
+LB: PA_OUT-----+--10nH--+-----ASM_IN
+               |        |
+              1pF       NC
+               |        |
+              GND      GND
+
+HB: PA_OUT-----+--33pF--+-----ASM_IN
+               |        |
+               NC       NC
+               |        |
+              GND      GND
+
+The HB pi network is degenerate, reduced to a single series capacitor, and the
+LB pi network is almost degenerate given how small that 1 pF cap to ground is.
+But pi networks have this advantage: a degenerate pi network is reduced to just
+one component, whereas a degenerate T network requires two components to be
+populated.
+
+It is also worth noting that both "from" and "to" components connected by these
+matching networks are different between Leonardo and OM: Leonardo and Tango use
+RF3133 PA, OM's version has RF3166; Leonardo FEM is M034F, OM's version has a
+Darfon ASM instead.
+
+In the now-seemingly-unlikely event that we end up building a new FreeCalypso
+GSM device (not Tango) with M034F quadband FEM, which PA are we going to use,
+and what "RF magic glue" matching network topology for PA-to-FEM RF Tx paths?
+The Mother's current preference is to use RF3166 and OM-style pi networks,
+preferably with a consultation with some better-than-me RF expert - but all
+such plans are extremely nebulous, written with forks on water.