Chapter 19

The Complete Transceiver

Chapter 5 turned Bluetooth Low Energy into a budget, and ten chapters designed the blocks to it. This chapter puts them back together. It checks every block against its target, rebuilds the receive and transmit chains from the simulated numbers, runs the behavioural receiver of Chapter 5 with them, and then looks at what a block-by-block design cannot see: the PA pulling the VCO, spurs, leakage, the calibrations that hold it all together, and what happens when the radio shares a phone with an LTE transmitter. It ends with where to take the design next.

  • Compare designed blocks with a link budget, and judge a margin by how it was obtained: simulation, hand analysis or assumption.
  • Cascade gain, noise figure, IIP3 and IIP2 from the designed blocks, find the noise and distortion shares, and test how far a doubtful number may move before the system fails.
  • Run a behavioural receiver with the designed numbers and read the BLE compliance margins, including carrier offset, ADC resolution, blocking and intermodulation.
  • Check a transmitter against its output power, efficiency, emission limits and modulation accuracy, and estimate VCO pulling, reference spurs and LO leakage with simple models.
  • Explain the calibrations a radio of this kind needs, and the coexistence problem of a radio inside a smartphone.
Contents

Figure 19.1 is the course radio as the block chapters left it. Blue blocks were designed and simulated in SG13G2 (Chapters 9–16); grey dashed blocks were never designed in the course and enter this chapter as stated assumptions, exactly as Chapter 5 allocated them: the band filter and T/R switch, the complex channel filter and VGA, the ADCs and the PA’s driver. Every number in this chapter comes from tools/figs/ch19_transceiver.py, which reads the block chapters’ result files (tools/figs/ch09_numbers.json to ch16_numbers.json), imports Chapter 5’s receiver, Chapter 12’s loop and Chapter 14’s synthesizer budget, and runs no new transistor-level simulation.

Block diagram of the complete course transceiver with numbers. Receiver: antenna, a dashed grey block band filter plus T/R switch labelled minus 2 dB, NF 2 dB, assumed; the LNA, labelled Chapter 9, 17.6 dB, 5.64 mW, NF 1.17 dB, IIP3 minus 4.9 dBm; two mixers fed by a divide-by-two, labelled Chapter 10, mixer plus TIA, plus 8.05 dB, NF 11.88 dB, IIP3 plus 10.6 dBm, and Chapter 13, 2.01 mW; a dashed complex filter plus VGA, plus 30 dB, NF 20 dB, 1.4 MHz at plus 1 MHz; two dashed ADCs, 6 bits; a digital baseband block listing GFSK modem, offset estimate and the calibrations: I/Q trim, DTC gain, VCO band, two-point gain, PA power. Synthesizer: a VCO 4.63 to 5.22 gigahertz, Chapter 11, 0.77 mW, Q 14.9, minus 116.7 dBc per hertz at 1 megahertz, tuned by a fractional-N PLL with MASH 1-1-1 and DTC, Chapters 12 to 14, f u 400 kilohertz, 40 megahertz reference, minus 112.7 dBc per hertz at 1 megahertz at the LO; the VCO feeds the receive divide-by-two, crossing the Q path, and a transmit divide-by-two. Transmitter: divide-by-two, a dashed driver labelled assumed, and an a PA labelled Chapter 16, plus 10 dBm, 31.0 percent PAE, harmonics minus 75.0 and minus 71.2 dBc with second- and third-harmonic traps, whose output goes through a match back to the T/R switch.
Figure 19.1. The course radio as designed: every block carries the number its chapter handed over. Blue: designed and simulated in SG13G2 (tt, 27 °C); grey dashed: not designed in the course, taken from Table 5.3 or assumed here. The PA is the off-chip version with Chapter 16's second- and third-harmonic traps. The VCO covers 4.63–5.22 GHz in 16 bands (simulated, typical corner).

Further learning · Open-SERDES Tutorial

A radio is one kind of transceiver; a wireline serial link is another, built from the same SG13G2 HBTs and CMOS. BRICD’s open tutorial Designing a Serial Link from Scratch designs a 5 Gb/s SerDes in IHP SG13G2, and its Part III · Transceiver Circuits turns the link model into transistors: the serializer and output driver on the transmit side, and the termination, front-end and samplers on the receive side. Its later chapters on jitter, PLL clock generation and clock and data recovery continue Chapters 12–15 of this course in a wireline setting.

Table 5.3 gave each block of the receiver a gain, a noise figure (NF) and an input third-order intercept (IIP3), and the synthesizer a phase-noise target, so that the cascade would reach an NF of 6.06 dB, an IIP3 of −12.7 dBm and a sensitivity of −95.6 dBm, with 5.6 dB to spare against the course’s −90 dBm. The course radio’s specification added the transmitter: +10 dBm into 50 Ω, a power-added efficiency (PAE) of at least 30 %, and harmonics below −40 dBc. Table 19.1 puts the designed blocks next to those targets.

Table 19.1. The designed blocks against Table 5.3 and the course radio (tt, 27 °C unless stated). Margin in dB; for PAE, 10 log of the ratio. "Hand" and "assumed" mark numbers that no simulation in the course produced.

Block Item Target Designed Margin How obtained
LNA (Ch. 9) gain, band minimum ≥ 16 dB 17.43 dB +1.43 ngspice S-parameters
  NF, band maximum ≤ 2.5 dB 1.18 dB +1.32 ngspice noise
  IIP3 ≥ −6 dBm −4.9 dBm +1.1 Xyce harmonic balance
  IIP3, worst corner (ff, 85 °C) ≥ −6 dBm −5.7 dBm +0.3 Xyce harmonic balance
  S11, band maximum ≤ −10 dB −15.39 dB +5.39 ngspice S-parameters (−13.5 dB with Chapter 18’s EM pad)
Mixer + TIA (Ch. 10), transistor-level OTA, RF = 223 Ω conversion gain +8 dB +8.05 dB +0.05 Xyce transient, PSP switches and OTAs (behavioural op-amp at 220 Ω: +8.02 dB)
  SSB NF ≤ 12 dB 11.88 dB +0.12 periodic small-signal analysis with the OTA’s simulated admittance and noise, 0.5–1.5 MHz average (behavioural op-amp: 11.25 dB)
  IIP3 ≥ +4 dBm +10.6 dBm +6.6 Xyce two-tone with the OTAs
  IIP2, 1 % switch mismatch ≥ +35.3 dBm ≥ +41.5 dBm +6.2 Xyce two-tone with the OTAs, lowest value from −22 to −10 dBm (behavioural op-amp: +56.0 dBm)
VCO (Ch. 11) L(1 MHz) at the VCO ≤ −104.0 dBc/Hz −116.7 dBc/Hz +12.7 ISF and frozen noise of every source (equivalent F = 3.71)
  coverage of 4.802–4.962 GHz with margin 4.633–5.224 GHz (tt) 1.0 % (fast corner, lower edge) transients, three corners
Divider (Ch. 13) drive at 4.8 GHz ≥ 100 mV 335 mV +10.5 transient
I/Q path (Ch. 13, 10) IRR, 99 % of chips ≥ 30 dB 34.7 dB +4.7 Monte Carlo, Section 19.2 (baseband assumed)
Synthesizer (Ch. 12–14) L(1 MHz) at the LO ≤ −110 dBc/Hz −112.7 dBc/Hz +2.7 linear loop model with the simulated VCO and the modelled, calibrated DTC (impairments assumed)
  L(3 MHz) at the LO ≤ −119.9 dBc/Hz −126.9 dBc/Hz +7.0 the same
  reference spur (1 nA, 4 % mismatch) ≤ −59.3 dBc −112 dBc +52.7 Fourier series of the pump current
PA (Ch. 16), off-chip Q = 40, 2f0 and 3f0 traps saturated output ≥ +10 dBm +10.4 dBm +0.4 ngspice transient
  PAE at +10 dBm ≥ 30 % 31.0 % +0.14 ngspice transient
  2nd harmonic at +10 dBm ≤ −40 dBc −75.0 dBc +35.0 ngspice transient
  3rd harmonic at +10 dBm ≤ −40 dBc −71.2 dBc +31.2 ngspice transient
PA, on-chip spirals saturated output, peak PAE ≥ +10 dBm, ≥ 30 % +9.2 dBm, 23.8 % −0.8, −1.0 (fail) ngspice transient

Figure 19.2 shows the same margins at a glance. Of the 22 numbers, two fail and five are marginal (less than 1 dB). Both failures belong to the on-chip PA, which Chapter 16 already reported: with the Chapter 7 spirals its output network loses 29 % of the power. Off chip, with Q = 40 inductors and the harmonic traps that Chapter 16 added (3f0 at the collector, 2f0 at the output), the PA meets its specification, if only just: 31.0 % PAE against 30 %. All five marginal numbers are simulated, so the margins are small but real: the LNA’s IIP3 at the fast corner and 85 °C, the PA’s saturated power and PAE, and the mixer’s gain and noise figure, which Chapter 10 simulates with a transistor-level OTA in both TIAs and RF = 223 Ω. The OTA’s noise takes the mixer from the behavioural op-amp’s 11.25 dB to 11.88 dB, 0.12 dB inside its 12 dB budget. That is the number to watch, and Section 19.2 asks how far it may move.

Lollipop chart of margins in dB for 21 block specifications, from minus 10 to 25 dB. Teal lines for passes: LNA gain 1.4, NF 1.3, IIP3 1.1, S11 5.4; mixer IIP3 6.6, IIP2 6.2; VCO phase noise 12.7; divider drive 10.5; IRR 4.7, hollow; synthesizer 1 MHz 3.1 and 3 MHz 7.4, hollow; reference spur beyond 24 with an arrow; PA second harmonic 35.0, third harmonic 31.2. Gold dots just right of zero for the LNA IIP3 at the worst corner 0.3, the mixer gain 0.05, mixer NF 0.12, PA saturated output 0.4 and PA PAE off-chip 0.1. Red for the on-chip PA: saturated output minus 0.8 and PAE minus 1.0. To the right of each row the value and target are printed.
Figure 19.2. The receiver, synthesizer and off-chip PA pass; only the on-chip PA fails, and every marginal number is simulated. Margin of each designed block against its Table 5.3 or course-radio target (tt, 27 °C). Hollow markers: a hand analysis, an estimate or an assumption sits inside the number (the synthesizer's behavioural DTC, whose impairments are assumed; the I/Q Monte Carlo's baseband). Shaded: marginal, less than 1 dB.

What this means for your design

A margin is only as good as the way it was obtained. The synthesizer’s 2.7 dB at 1 MHz looks comfortable, and its VCO is now simulated with every noise source (Chapter 11), but its DTC is a behavioural model whose gain error, INL and jitter are assumed sizes (Chapter 14); the LNA’s 0.3 dB of IIP3 margin at the worst corner looks thin, but it is a simulated corner. When you sign off a block, record next to each number how it was obtained, and spend the next verification effort where an unsimulated number sits close to its limit.

19.2 The receive chain

The cascade with the designed numbers

The designed blocks do not match the budget one for one. The LNA has 1.6 dB more gain and 1.3 dB less noise than allocated; the mixer, with its transistor-level OTA, has 0.12 dB less noise and 6.6 dB more IIP3. The cascade moves accordingly. With the gains, NFs and IIP3s of Table 19.2, the Friis formula (Eq. (4.11)) and the cascaded-IIP3 formula of Chapter 4 give an NF of 4.62 dB and an IIP3 of −11.6 dBm, against 6.06 dB and −12.7 dBm in the budget. By Eq. (4.12), with SNRmin = 12.3 dB in 1 MHz, the sensitivity is −97.1 dBm: 7.1 dB better than the course’s −90 dBm and 27.1 dB better than BLE’s −70 dBm.

Which block should be improved first? The noise share of block \(k\) is its term in the Friis sum relative to the whole:

Key equation · Noise share of a block in a cascade

\[s_1 = \frac{F_1}{F}, \qquad s_k = \frac{F_k - 1}{F\,G_1G_2\cdots G_{k-1}} \quad (k > 1), \qquad \sum_k s_k = 1 \tag{19.1}\]

\(F\) is the cascade’s noise factor and \(G_k\) the available gains, as linear ratios. The same construction with \(G_1\cdots G_{k-1}/\mathrm{IIP3}_k\) in milliwatts gives each block’s share of \(1/\mathrm{IIP3}\).

Table 19.2. The receive cascade with the designed blocks against the Table 5.3 budget, which is Chapter 4's illustrative cascade (Table 4.2). Gains in dB, IIP3 in dBm; shares of the noise factor and of 1/IIP3 by Eq. (19.1). Band filter + switch and complex filter + VGA: assumed (Table 5.3 values).

Block Gain NF IIP3 Noise share IIP3 share Cascade NF after it Cascade IIP3 after it
Band filter + T/R switch (assumed) −2.0 2.0 +40 54.7 % 0 % 2.00 +40.0
LNA (Ch. 9) +17.61 1.17 −4.9 16.9 % 13.6 % 3.17 −2.9
I/Q mixer + TIA (Ch. 10, transistor-level OTA) +8.05 11.88 +10.6 13.7 % 22.1 % 3.93 −7.1
Complex filter + VGA (assumed) +30.0 20.0 +14 14.7 % 64.4 % 4.62 −11.6
Budget (Table 5.3) +52.0     39 / 31 / 15 / 16 % 0 / 13 / 53 / 34 % 6.06 −12.7

The answer has changed since Chapter 4. The budget’s LNA took 31 % of the noise; the designed LNA takes 17 %, and the passive loss in front of it now takes 55 %. Without the band filter and switch the receiver’s NF would be 2.62 dB. The LNA’s extra gain also changes the linearity: more of the signal reaches the later stages, which now set 86 % of 1/IIP3, most of it in the assumed baseband. The designed receiver is noise-rich and linearity-poor compared with its budget, a reasonable place to be: the 7.1 dB of sensitivity margin is worth more as power saved than as range.

Power is where the design is weakest as a product. The receive chain draws 24.9 mW before its baseband: 5.64 mW in the LNA, 2.73 mW in the mixer’s LO drivers (Chapter 10’s NAND–inverter chain), 13.6 mW in the two TIAs’ OTAs (simulated in Chapter 10, 6 % above the earlier 12.8 mW estimate), and 0.77, 2.01 and 0.18 mW in the VCO, the divide-by-two and the prescaler. The TIAs take more than half, for the reason Chapter 10 gave: a 50 Ω interface forces a small feedback resistor, whose noise the OTA must then beat. A low-noise transconductance amplifier driving the mixer in current mode, with the LNA’s surplus gain and the receiver’s surplus NF spent on lower TIA current, is the obvious next iteration.

How sure are these numbers? Figure 19.3b answers for the mixer and the assumed baseband. The mixer’s simulated 11.88 dB could rise to 21.4 dB before the system NF reaches its 8 dB specification, and to 26.5 dB before the sensitivity reaches −90 dBm; 3 dB worse raises the system NF only to 5.21 dB. With the behavioural op-amp of Chapter 10’s earlier simulations (+8.02 dB, 11.25 dB, +9.9 dBm) the system NF was 4.54 dB, the IIP3 −11.7 dBm and the sensitivity −97.2 dBm: the transistor-level OTA costs the receiver 0.08 dB of NF. The assumed baseband NF could rise from 20 to 29.5 dB. The LNA’s corners, taking the worst gain (17.06 dB), NF (1.40 dB) and IIP3 (−5.7 dBm) of Chapter 9’s corner table together, give 4.94 dB, −96.8 dBm and −11.2 dBm; the other blocks were not simulated over corners.

Two panels. (a) Cascade noise figure, left axis, and cascade IIP3, right axis, after each block: filter plus switch, LNA, mixer plus TIA, filter plus VGA. Designed NF rises 2.00, 3.17, 3.93, 4.62 dB; the Table 5.3 budget rises from 2 to 6.06 dB, above it. Designed IIP3 falls from plus 40 to minus 2.9, minus 7.1 and minus 11.6 dBm; the budget's from plus 40 to minus 4.0, minus 11.0 and minus 12.7. A dotted line marks the 8 dB NF specification. (b) Receiver NF against the NF of one block from 6 to 26 dB: varying the mixer plus TIA, the curve rises from 4.1 dB through the design point, 11.88 dB and 4.62 dB, and crosses 8 dB at 21.4 dB, marked; varying the filter plus VGA, a dashed curve reaches only 6.2 dB at 26 dB. Dotted and dash-dotted lines mark 8 dB and the NF for minus 90 dBm, 11.7 dB.
Figure 19.3. The designed receiver has an NF of 4.62 dB, 3.4 dB inside its specification; the mixer NF could be 9.6 dB worse before that matters. (a) Cascade NF and IIP3 at the output of each block, designed (filled) against the Table 5.3 budget (open), referred to the antenna. (b) System NF against the NF of the mixer + TIA or of the complex filter + VGA, all other blocks at their design values; markers: design points; diamond: the mixer NF at which the system reaches 8 dB.

Second-order distortion. A product generated in the mixer appears at the antenna reduced by the gain in front of the mixer:

\[\mathrm{IIP2}_{ant} = \mathrm{IIP2}_{mixer} - \left(G_{filter} + G_{LNA}\right)\ \ \mathrm{dB}. \tag{19.2}\]

With the transistor-level OTA and a deliberate 1 % switch mismatch, Chapter 10 found the mixer’s IIP2 no lower than +41.5 dBm; its IM2 product does not grow with a slope of 2, so this is a bound rather than an intercept. At the antenna it becomes +25.9 dBm, 4.6 dB above the +21.3 dBm that Chapter 5’s Wi-Fi interferer at −40 dBm requires. The behavioural op-amp’s +56.0 dBm had given +40.4 dBm and 19.1 dB.

Compression. The LNA compresses at −16.6 dBm at its input, −14.6 dBm at the antenna. The mixer goes first, and with the transistor-level OTA its TIA outputs, not its switches, set the limit: a single tone at the mixer’s input compresses its gain by 1 dB at −3.5 dBm (Chapter 10), which, through the 15.6 dB of gain in front of it, is −19.1 dBm at the antenna. Two tones of −4 dBm (−19.6 dBm at the antenna) compress it by 4.7 dB, where the behavioural op-amp lost only 1.1 dB. A third-order model of the front end with IIP3 = −7.1 dBm would put the 1 dB point at −16.7 dBm; the mixer compresses 2.4 dB earlier, because the OTA’s outputs clip rather than bend (its own IIP3 of +10.6 dBm would predict a 1 dB point near +1.0 dBm at its input). All three lie 15.9 dB or more above BLE’s −35 dBm blockers; the behavioural tests below confirm it.

The LO: reciprocal mixing and the DTC

The synthesizer of Chapter 14 replaces Chapter 5’s phase-noise mask in Figure 19.4a: Chapter 11’s simulated VCO in Chapter 12’s 400 kHz loop, with the MASH 1-1-1 and the modelled DTC of Chapter 14, its gain and INL calibrated. At the LO it reaches −112.7 dBc/Hz at 1 MHz and −126.9 dBc/Hz at 3 MHz. With the designed NF, the course case of Chapter 5 (a −40 dBm interferer 3 MHz away, the wanted signal at −87 dBm, 12.3 dB of SINR) leaves −99.7 dBm for reciprocal mixing in the channel, so Eq. (5.18) asks for −119.7 dBc/Hz at 3 MHz: the synthesizer meets it with 7.2 dB. The remaining requirements are easy: 11.6 dB of margin at 2 MHz, 25.8 dB for the −35 dBm blockers 27 MHz away, and 28–46 dB for BLE’s own tests at −67 dBm. Integrated from 10 kHz to 8 MHz, the phase error is 0.41° rms (Chapter 5’s mask: 1.11°), and the rms frequency error up to 0.5 MHz is 1.60 kHz. With Chapter 11’s earlier estimate of the VCO (noise factor F = 2–5, a 1/f³ corner assumed at 1 MHz) in place of its simulation, the 3 MHz margin would be 6.0–7.5 dB (the shaded band in Figure 19.4a); no conclusion here depends on it.

Two panels. (a) L of f at the 2.44 gigahertz LO from 10 kilohertz to 100 megahertz: the course synthesizer, a solid line, flat near minus 106 below 30 kilohertz, a gentle hump near minus 104 at 200 kilohertz, minus 112.7 at 1 megahertz, labelled, then falling to minus 142 at 10 megahertz and below minus 150 beyond 30 megahertz; a narrow light band around it marks the earlier VCO estimate F equal 2 to 5. Chapter 5's dotted mask lies above it at every offset. Orange filled triangles, requirements with the wanted signal at minus 87 dBm, at minus 109.7 at 2 megahertz, minus 119.7 at 3, minus 124.7 at 27 and 58 megahertz; grey open triangles for BLE's own tests at minus 89.3, minus 99.3 and minus 104.3. All triangles lie above the curve. (b) L at the LO against an assumed DTC residual from 0 to 100 percent, simulated VCO: at 1 megahertz, blue, nearly flat from minus 113.0 to minus 111.3 at 50 percent and minus 108.5 at 100; at 3 megahertz, orange, rising from minus 127.1 through minus 124.2 at 10 percent and crossing the minus 119.9 dotted specification near 21 percent, to minus 107.3 at 100 percent; thin dashed curves for F equal 5 lie about 1 dB higher at low residuals. Two stars mark the modelled, calibrated DTC of Chapter 14: on the 3 megahertz curve at 2.4 percent, minus 126.9, and on the 1 megahertz curve at 20 percent, minus 112.7. A dash-dotted line marks Section 14.4's 10 percent.
Figure 19.4. The synthesizer clears every reciprocal-mixing requirement; its 3 MHz margin rests on the DTC cancelling at least 78 % of the quantization error, which the calibrated DTC does with room to spare. (a) Phase noise at the LO of the Chapter 14 synthesizer (simulated VCO, MASH 1-1-1, modelled DTC with gain and INL calibrated) against Chapter 5's mask and the requirements of Eq. (5.18) with the designed NF of 4.62 dB; shaded: Chapter 11's earlier VCO estimate F = 2–5 with the same DTC. (b) L(1 MHz) and L(3 MHz) with an assumed DTC residual δ in place of the modelled DTC (Section 14.4's model), simulated VCO (solid) and F = 5 (dashed); stars: the modelled, calibrated DTC at its effective residual at each offset; dotted: the 1 and 3 MHz specifications.

Chapter 14 first assumed that the digital-to-time converter (DTC) cancels the modulator’s timing error to a 10 % residual, and then modelled the DTC and its calibration; the budget above uses the model. Figure 19.4b shows how much rests on the cancellation, with a residual δ in place of the model. The 1 MHz point hardly cares: it holds −110 dBc/Hz up to a residual of 73 %. The 3 MHz point does: it meets −119.9 dBc/Hz only for residuals below 21 %, and without the DTC (100 %) both points fail, at −108.5 and −107.3 dBc/Hz. The calibrated DTC is worth δ = 2.4 % at 3 MHz, but 20 % at 1 MHz, where its own jitter, rounding and leftover INL add a white floor that a scaled residual does not have (the stars). The DTC is therefore load-bearing at 3 MHz, and its gain and INL must be calibrated; Section 19.5 shows the gain loop.

I/Q accuracy

Chapter 5 allocated 0.2 dB of gain error and 2° of phase error to the whole I/Q path (33.6 dB of IRR). The largest designed contribution is the divide-by-two: Chapter 13 found its quadrature error to be 0.114° per millivolt of offset at its clock input, for 0.3 V of drive. The VCO delivers 0.335 V, so by Eq. (13.4) the sensitivity is 0.102°/mV. The offset comes from the threshold mismatch of the 6 µm × 0.13 µm clock pair, which by Pelgrom’s law [177], with the SG13G2 constant \(A_{VT}\) = 5.7 mV·µm extracted in The Art of Analog IC Design (Chapter 13), is

\[\sigma_{\theta} = \frac{d\theta}{dV_{os}}\,\frac{A_{VT}}{\sqrt{WL}} = 0.102\ ^\circ/\mathrm{mV}\times\frac{5.7\ \mathrm{mV\,\mu m}}{\sqrt{6 \times 0.13}\ \mathrm{\mu m}} = 0.102 \times 6.45\ \mathrm{mV} = 0.66^\circ. \tag{19.3}\]

The mixer with its OTAs adds 0.005 dB and 0.044° (Chapter 10’s 1 % switch mismatch, taken as one standard deviation). The mismatch between the I and Q TIAs was not simulated, and the complex filter and the VGA were not designed; we assume 0.1 dB and 0.3° (one standard deviation) for them together. A Monte Carlo of 200 000 chips (Figure 19.5) gives a total phase error of 0.73° rms. Its 3σ value, 2.18°, exceeds the 2° allocation, and 0.60 % of the chips are beyond 2°; yet none of the 200 000 falls below 30 dB of IRR, because the designed gain error is far smaller than the 0.2 dB that Chapter 5 allowed. The median IRR is 42.9 dB and 99 % of the chips reach 34.7 dB. The behavioural runs below use that 1 % chip: 0.14 dB and −1.92°.

Cumulative distribution of the image-rejection ratio over 200 000 simulated chips, on a log percentage axis from 0.01 to 100 percent against IRR from 10 to 80 dB. Divider offset only and divider plus mixer, nearly identical, reach 1 percent at 36.6 and 36.5 dB and 50 percent near 48 dB. Adding the assumed baseband moves the curve left: 1 percent at 34.7 dB, median 42.9 dB. After trimming to 0.5 millivolt and 0.02 dB steps the curve starts near 64 dB. Dotted lines at 15 dB, BLE's image test, and 30 dB, the course target; a dash-dotted line at 1 percent.
Figure 19.5. The divider's mismatch alone would exceed the 2° allocation at 3σ, but the IRR stays above 34.7 dB on 99 % of chips. Monte Carlo of the image-rejection ratio (Eq. (5.14)): divider clock-pair offset (Eq. 19.3), plus the mixer's errors, plus an assumed baseband (σ = 0.1 dB, 0.3°); and after the trim of Section 19.5.

The behavioural receiver with the designed numbers

Chapter 5 checked its budget with a complex-envelope receiver: random GFSK data, real filters, a discriminator, bit errors counted. tools/figs/_ch19_rx.py imports its pieces and adds what Chapter 5 left out, all referred to the antenna:

  • Noise for the designed NF, 4.62 dB, in absolute power: the wanted signal is specified in dBm.
  • Nonlinearity: a third-order, saturating model of the front end (band filter, LNA and mixer, IIP3 = −7.1 dBm) on the complex envelope, and one of the baseband (IIP3 = −9.7 dBm at the antenna) on each real I and Q rail, where it can also mix interferers across zero frequency.
  • The synthesizer’s phase noise, synthesized from the L(f) of Figure 19.4a (simulated VCO, modelled DTC), not a mask.
  • The I/Q errors of the 1 % chip, the assumed band filter (a third-order 0.5 dB Chebyshev, 2400–2483.5 MHz) acting on each interferer at its absolute frequency, a carrier frequency offset of up to ±150 kHz removed by a digital estimate, and a 6-bit ADC on each rail with an ideal AGC that puts the rms 12 dB below full scale.

With Chapter 5’s inputs the new model reproduces Table 5.4’s tolerances within 1.1 dB, the resolution of a search in 2 and 3 dB steps, which checks the reimplementation.

Sensitivity and the carrier offset. Figure 19.6a shows the bit-error rate against the wanted power. With Table 5.3’s complex filter (4th-order Butterworth, 1 MHz noise bandwidth, cut-off 487 kHz) and no carrier offset, the BER reaches 10−3 at −96.9 dBm, within 0.2 dB of Eq. (4.12). But BLE allows a transmitter’s carrier to be 150 kHz off [184], and then the same receiver needs −91.7 dBm: the 1 MHz filter cuts into the shifted GFSK spectrum (99 % of its power within 1031 kHz), and the inter-symbol interference costs 5.2 dB even though the offset itself is removed digitally. Widening the filter to a 1.4 MHz noise bandwidth (cut-off 682 kHz) admits 1.5 dB more noise but recovers the shape: −97.2 dBm without offset, −96.3 and −96.2 dBm at +150 and −150 kHz. That is the filter this chapter recommends, and the one assumed from here on. Estimating the offset from each packet’s 8-bit preamble alone, instead of tracking it over many packets, costs another 0.7 dB (−96.5 dBm).

The ADC needs few bits: 6 bits at 12 dB of back-off lose nothing measurable (−97.2 against −97.0 dBm without an ADC, within the statistical scatter), 4 bits lose 1.4 dB. A limiting receiver with 1 or 2 bits needs a different back-off and is not modelled here.

Two panels. (a) Bit-error rate on a log axis from 1e-5 to 0.3 against the wanted signal from minus 101 to minus 86 dBm. With the 1.4 MHz filter and no offset the BER crosses 1e-3 at minus 97.2 dBm; with a plus 150 kilohertz offset at minus 96.3; with the 8-bit preamble estimate at minus 96.5; with the 1.0 MHz filter and no offset at minus 96.9; with the 1.0 MHz filter and a 150 kilohertz offset the dashed orange curve is far to the right and crosses 1e-3 only at minus 91.7 dBm. A dash-dotted line marks Eq. 4.12, minus 97.1 dBm. (b) Bars of margin to the BLE requirement in dB for nine C/I tests, four CW blocker tests and three intermodulation tests, for Chapter 5 inputs, the designed receiver with the 1.0 MHz filter and with the 1.4 MHz filter. All bars are positive; the smallest is 6.1 dB for the 1.4 MHz filter at plus 2 MHz. Blockers 3 and 4 MHz away give 16.1 and 15.7 dB; the far blockers at 27 and 58 MHz are hatched bars at 34 dB marked as lower bounds, with a dashed orange line at 15.9 dB, labelled mixer P1dB, for the mixer's 1 dB compression point; intermodulation gives about 19 dB.
Figure 19.6. With a 1.4 MHz channel filter the designed receiver passes every BLE test, the closest by 6.1 dB; Table 5.3's 1 MHz filter loses 5.2 dB of sensitivity to a 150 kHz carrier offset. Behavioural receiver with the designed numbers (NF 4.62 dB, front-end and baseband IIP3, the synthesizer phase noise of Figure 19.4a, the 1 % I/Q chip, 6-bit ADCs, assumed band filter). (a) BER against the wanted power at the antenna for the 1.0 MHz and 1.4 MHz channel filters, with and without a +150 kHz carrier offset, and with a per-packet preamble estimate; 105 bits per point. (b) Margin to each BLE requirement: interference tests (wanted −67 dBm, GFSK interferers) with Chapter 5's inputs and with the designed numbers for both filters; CW blockers outside the band (wanted −67 dBm, requirement −35 dBm) and intermodulation (wanted −64 dBm, requirement −50 dBm, the worse of ±n) for the 1.4 MHz filter.

Interference, blocking and intermodulation. With Chapter 5’s inputs the model passes everything, as Table 5.4 did. With the designed numbers and the 1.0 MHz filter, every interference test passes by at least 14.4 dB (Table 19.3, Figure 19.6b). The 1.4 MHz filter that the carrier offset demands pays for its width in selectivity: it attenuates the channel 2 MHz away by 39.0 dB instead of 50.8 dB, and the +2 MHz margin falls from 15.8 to 6.1 dB, the smallest of all; the ±1 MHz and image margins shrink to about 13 dB. Every test still passes. The image margin, 12.9 dB, is no longer simply the IRR less 15 dB (Chapter 5): the wider filter lets more of the image channel itself through. Neither filter is the obvious optimum, and a design between them, or a sharper filter of higher order, is the next iteration.

BLE’s blocking test (Core Specification, Vol 6, Part A, §4.3) places the wanted signal at 2426 MHz [184]. We also run the edge channels, 2402 and 2480 MHz, with a CW interferer just outside the band at 2399 and 2484 MHz: §4.2 counts any interferer outside 2400–2483.5 MHz as out-of-band blocking, so the same −35 dBm applies. These blockers, 3 and 4 MHz from the wanted channel, are tolerated up to −18.9 and −19.3 dBm, 16.1 and 15.7 dB above the requirement; at those levels the blocker’s reciprocal mixing puts about −86 and −90 dBm into the channel, and the baseband rails, which clip at −15.7 dBm referred to the antenna, are within 3–4 dB of clipping. With the synthesizer’s lower noise beyond 2 MHz, the baseband’s compression now shares this limit with reciprocal mixing.

Far from the channel, 27 and 58 MHz away, the model never failed up to −1 dBm; but its saturation keeps the phase of the signal, as a limiter does, and a compressed transistor stage does not, so the honest limit there is the mixer’s 1 dB compression point of Section 19.2: the OTA’s −3.5 dBm at the mixer’s input, −19.1 dBm at the antenna, 15.9 dB above the requirement. Two blockers at once compress it sooner: Chapter 10’s two tones of −4 dBm each (−19.6 dBm at the antenna) cost 4.7 dB of gain.

In BLE’s intermodulation test the two −50 dBm interferers are tolerated up to −30.6 to −30.8 dBm for every n = ±3, ±4 and ±5 (BLE asks for only one of n = 3, 4, 5 to pass, so testing all of them is stricter), 19.2–19.4 dB above the requirement and about 2.5 dB better than Eq. (5.16) predicts with the cascaded IIP3 (−33.1 dBm), because the third-order product of a CW and a GFSK interferer is GFSK-like and the discriminator tolerates it at a lower C/I than noise, the pitfall of Chapter 5 again.

Table 19.3. BLE receiver tests against the behavioural course receiver. C/I tests: the C/I at which the BER reaches 0.1 % (wanted −67 dBm, GFSK interferer; offsets from the wanted channel, LO 1 MHz below it). Blocking: the CW blocker power at 0.1 % BER (wanted −67 dBm; BLE's §4.3 test uses 2426 MHz, the edge-channel cases follow §4.2). Intermodulation: the power of each of the two interferers, CW at f0 + n MHz and GFSK at f0 + 2n MHz (wanted −64 dBm). Margin in dB.

Test Requirement Chapter 5 inputs Designed, 1.0 MHz filter Designed, 1.4 MHz filter
C/I, co-channel 21 dB 7.3 (13.7) 6.6 (14.4) 4.3 (16.7)
C/I, +1 MHz 15 dB −6.5 (21.5) −3.7 (18.7) 1.6 (13.4)
C/I, −1 MHz 15 dB −6.6 (21.6) −4.4 (19.4) 1.7 (13.3)
C/I, +2 MHz −17 dB −36.9 (19.9) −32.8 (15.8) −23.2 (6.2)
C/I, −2 MHz (image) −9 dB −25.6 (16.6) −26.6 (17.6) −21.9 (12.9)
C/I, +3 MHz −27 dB −46.2 (19.2) −45.6 (18.6) −37.6 (10.6)
C/I, −3 MHz (image − 1 MHz) −15 dB −37.8 (22.8) −38.1 (23.1) −34.5 (19.5)
C/I, +4 MHz −27 dB −49.1 (22.1) −51.7 (24.7) −46.8 (19.8)
C/I, −4 MHz −27 dB −49.4 (22.4) −51.8 (24.8) −46.8 (19.8)
Blocking, CW 2399 MHz, wanted 2402 MHz (edge channel; out of band by §4.2) −35 dBm     −18.9 dBm (16.1)
Blocking, CW 2484 MHz, wanted 2480 MHz (edge channel; out of band by §4.2) −35 dBm     −19.3 dBm (15.7)
Blocking, CW 2399 and 2484 MHz, wanted 2426 MHz (the §4.3 test) −35 dBm     no failure up to −1 dBm in the model; mixer P1dB (single tone) −19.1 dBm (15.9)
Intermodulation, n = ±3, ±4, ±5 (BLE requires one of n = 3, 4, 5) −50 dBm     −30.6 to −30.8 dBm (19.2–19.4)

19.3 The transmit chain

The transmitter is the synthesizer, a second divide-by-two, a driver and the PA (Figure 19.1). GFSK is applied at two points, the fractional divider and the VCO’s tuning input (Chapter 14), the PA runs in compression, and its output network is the Q = 40 off-chip version of Chapter 16. Table 19.4 collects the results.

Output power and efficiency. The PA, with Chapter 16’s harmonic traps, delivers +10 dBm at 31.0 % PAE (33.7 % collector efficiency) from 14.9 mA at 2.0 V, 29.8 mW, with −2.2 dBm of drive. It saturates at +10.4 dBm, so it has 0.4 dB of headroom. That headroom is the problem of the whole chain. If, as in Figure 19.1, the transmitter shares the antenna with the receiver through the T/R switch and the band filter, Table 5.3’s 2 dB of loss applies to it too, and the antenna receives +8.0 dBm; delivering +10 dBm there would need +12.0 dBm from a PA that saturates 1.6 dB lower. The 0.60 mW of drive must also come from somewhere: the divide-by-two’s 1 kΩ CML output cannot deliver it into the PA’s 50 Ω input, so a driver stage, not designed in the course, sits in between. Counting the PA, VCO, transmit divider and prescaler, 32.8 mW, the transmitter turns 30.5 % of its DC power into +10 dBm, before the driver and the digital circuits.

Harmonics and spurs. Regulations measure emissions at the antenna. In Europe the limit from 1 to 12.75 GHz is −30 dBm [189]. In the United States both harmonics fall in restricted bands (4.5–5.15 and 7.25–7.75 GHz) [190], where the limit is a field of 500 µV/m at 3 m, an equivalent isotropically radiated power of −41.2 dBm. Chapter 16’s harmonic traps bring the second harmonic down to −65.0 dBm and the third to −61.2 dBm at the antenna, 35.0 and 31.2 dB inside ETSI’s limit and 23.8 and 20.0 dB inside the FCC’s (Figure 19.7b). Without traps the two were −52.6 and −21.8 dBm, and the third failed both limits; a 3f0 trap alone fixed the third but raised the second to 0.5 dB above the FCC limit, which the 2f0 trap at the output then removed at no cost in PAE. No extra filter is needed; one with 0.3 dB of loss would lower the PAE to 28.9 %. The synthesizer’s spurs are harmless here: the fractional spur at 2 MHz (Chapter 14’s worst case, 4 % charge-pump mismatch) is −57.0 dBm and the reference spur −102 dBm.

Three panels. (a) Transmitted PSD in dBm per megahertz against offset from minus 6 to 6 megahertz for plus 10 dBm: GFSK alone falls below minus 100 by 3 megahertz; with the LO phase noise the skirts fall more slowly, from about minus 50 dBm per megahertz at 2 megahertz to minus 66 at 6 megahertz. Red steps mark the in-band emission limit: minus 20 dBm at 1.5 to 2.5 megahertz and minus 30 beyond. (b) Bars of level at the antenna: second harmonic minus 65.0 dBm and third harmonic minus 61.2 dBm, both well below the ETSI minus 30 dashed line and the FCC minus 41.2 dotted line; fractional spur at 2 megahertz minus 57.0; reference spur at 40 megahertz minus 102.0. (c) Deviation in kilohertz against the direct-path gain error from minus 40 to 40 percent: delta f1 average falls from 275 to 225 kilohertz, delta f2 average rises from 163 to 277 kilohertz; a shaded band 225 to 275 kilohertz, a dotted line at 185 kilohertz, and a grey band for the plus or minus 30 percent K VCO spread.
Figure 19.7. The transmitter's spectrum and harmonics are clean; its two-point modulation needs a calibrated gain. (a) Spectrum of the course transmitter at +10 dBm: GFSK (BT = 0.5, h = 0.5) through the synthesizer's phase noise (Figure 19.4a), Welch estimate; red: BLE's in-band emission limits for a 1 MHz band. (b) Harmonics (Chapter 16, off-chip network with the 2f0 and 3f0 traps) and synthesizer spurs against ETSI EN 300 328 and FCC §15.209 in restricted bands. (c) BLE modulation characteristics of the two-point modulator against the gain error γ of the VCO path (Eq. (14.7), Chapter 12's loop): Δf1avg for 11110000, Δf2avg for 1010; shaded: KVCO ± 30 % uncalibrated.

In-band emission. The LO’s phase noise spreads the carrier into the neighbouring channels. Through the designed synthesizer, the 1 MHz band at ±2 MHz receives up to −49.3 dBm and the bands at ±3 MHz −56.5 dBm (Figure 19.7a), 29.3 and 26.5 dB inside BLE’s −20 and −30 dBm limits [184]. At 2 MHz the GFSK spectrum itself (−55.7 dBm) and the phase noise (−50.8 dBm) contribute together; from 3 MHz on the phase noise alone sets the level.

Modulation accuracy. BLE specifies the transmitted deviation rather than an EVM. The Core Specification asks for a modulation index of 0.45–0.55, a minimum deviation of 185 kHz, and a deviation for 1010 patterns of at least 80 % of that for 00001111 patterns [184]; the RF test specification measures these as Δf1avg (over 11110000 patterns, 225–275 kHz), Δf2max (above 185 kHz) and Δf2avg/Δf1avg (at least 0.8), the names used here. The phase noise hardly matters: 0.41° rms and 1.60 kHz of rms frequency error against a 250 kHz deviation. The two-point modulator does. With matched paths the model gives Δf1avg = 250.0 kHz and Δf2avg = 219.8 kHz, a ratio of 0.879, set by GFSK’s own inter-symbol interference. Through the divider alone, without the VCO path, the 400 kHz loop passes the 500 kHz pattern so poorly that Δf2avg falls to 76.6 kHz. With the VCO path’s gain off by γ (Eq. (14.7)), Figure 19.7c shows Δf1avg barely moving and Δf2avg following γ almost linearly: the ratio drops below 0.8 for γ < −10 %, and Δf2 below 185 kHz near −25 %. The test tolerates a VCO path that is too strong far better than one that is too weak, but KVCO’s spread after band calibration, −21 % to +27 % around its design value (Chapter 11), covers both: the gain must be calibrated (Section 19.5).

Table 19.4. The course transmitter (2.44 GHz, tt, 27 °C; PA with the Q = 40 off-chip network, VPA = 2.0 V).

Item Requirement Course transmitter Margin
Output power at the PA +10 dBm +10 dBm (saturates at +10.4 dBm) 0.4 dB of headroom
Output power at the antenna, through 2 dB of filter and switch — +8.0 dBm PA would need +12.0 dBm
PAE at +10 dBm ≥ 30 % 31.0 % (collector efficiency 33.7 %) 1.0 point
DC power: PA, VCO, divide-by-two, prescaler — 29.8 + 0.77 + 2.01 + 0.18 = 32.8 mW driver not designed
2nd harmonic ≤ −30 dBm (ETSI), ≤ −41.2 dBm (FCC) −65.0 dBm 35.0 dB, 23.8 dB
3rd harmonic ≤ −30 dBm (ETSI), ≤ −41.2 dBm (FCC) −61.2 dBm 31.2 dB, 20.0 dB
In-band emission at ±2 MHz, ≥ ±3 MHz ≤ −20 dBm, ≤ −30 dBm −49.3 dBm, −56.5 dBm 29.3 dB, 26.5 dB
Δf1avg, Δf2avg/Δf1avg (matched two-point paths) 225–275 kHz, ≥ 0.8 250.0 kHz, 0.879 needs γ ≥ −10 %
LO phase error, 10 kHz–8 MHz — 0.41° rms  

19.4 Spurs, coupling and self-interference

A block simulated alone cannot see the others. Three couplings matter most in this radio.

The PA’s second harmonic and the VCO

Chapter 5 placed the VCO at twice the carrier so that the PA’s output would not pull it (Chapter 5, Worked example 5.3). The second harmonic of the PA does fall on the VCO. How much coupling can the radio take? Adler’s equation (5.15) [71, 82] gives the lock range in terms of the injected and the oscillating currents. If a fraction of the PA’s harmonic power reaches the tank, the currents scale as the square roots of the powers that each delivers to the tank’s parallel resistance:

Key equation · Lock range for an injected power

\[f_L = \frac{f_0}{2Q}\sqrt{\frac{P_{inj}}{P_{osc}}}, \qquad P_{osc} = \frac{V_{osc}^2}{2R_p}, \qquad P_{inj} = P_{2f_c} - \mathrm{isolation} \tag{19.4}\]

Within ±fL of its free-running frequency the VCO locks to the injection; a coherent injection shifts its frequency by up to fL.

The course VCO has Q = 14.9, Rp = 906 Ω and 0.784 V of differential amplitude: Posc = −4.7 dBm and f0/2Q = 163.8 MHz. Its injection is the PA’s second harmonic, −65.0 dBm at the output with Chapter 16’s harmonic traps. With 40 dB of isolation between the PA’s output and the tank (Chapter 5’s “easily reached” coupling, our assumption), Eq. (19.4) gives fL = 1.58 kHz at the VCO. Had the VCO run at the carrier, it would receive the +10 dBm fundamental, 75.0 dB more, and fL would be 4.45 MHz, ten times the loop bandwidth (Figure 19.8a). BLE’s 50 kHz drift limit [184] applies to the carrier, and the divide-by-two halves the course VCO’s frequency error on its way there (0.79 kHz at 40 dB). To stay below it, the VCO at the carrier would need 79 dB of isolation; at twice the carrier, 4.0 dB.

The harmonic is also coherent: it is derived from the VCO itself, through the dividers and the PA, so it injects at exactly the VCO’s frequency with a fixed phase. That shifts the VCO’s free-running frequency by at most fL, which the type-II loop removes. The risk is the moment the injection changes: when the PA ramps up before a packet, the harmonic grows and its phase moves with the PA’s AM-PM (−33.5° at +10 dBm). We model the worst case, a VCO-referred frequency step of fL that grows with the square of a 2 µs raised-cosine ramp, and run it through Chapter 12’s closed loop (Figure 19.8b). With 40 dB of isolation the LO’s frequency moves by at most 0.24 kHz and settles in 4.1 µs; with 20 dB, 2.42 kHz and 6.2 µs. Both are small against BLE’s ±50 kHz drift and its 400 Hz/µs drift-rate limit once the packet starts [184]. The harmonic traps matter here too: with the 3f0 trap alone, whose second harmonic was 24 dB higher, the 20 dB case reached 39.7 kHz.

Two panels. (a) Lock range referred to the carrier on a log axis from 10 hertz to 1 gigahertz against isolation from 0 to 100 dB: for a VCO at the carrier receiving the PA fundamental at plus 10 dBm, an orange line falls from about 400 megahertz at 0 dB to 4 kilohertz at 100 dB; for the course VCO at twice the carrier receiving the second harmonic at minus 65.0 dBm, its lock range halved by the divider, a blue line 75.0 dB lower falls from about 80 kilohertz at 0 dB to 10 hertz near 78 dB. Horizontal lines mark the BLE drift limit of 50 kilohertz and the 400 kilohertz loop bandwidth; a dashed vertical line marks 40 dB. (b) LO frequency error in kilohertz against time from 0 to 12 microseconds after the start of a 2 microsecond PA ramp, shaded: with 20 dB of isolation the error rises to 2.42 kilohertz, swings negative and settles within about 6 microseconds; with 40 dB, peak 0.24 kilohertz; with 60 dB, 0.02 kilohertz.
Figure 19.8. Running the VCO at twice the carrier buys 75.0 dB of isolation against PA pulling; what remains is a kick well below a kilohertz during the PA ramp at 40 dB of isolation, which the loop removes in microseconds. (a) Adler lock range, Eq. (19.4), referred to the carrier, against the isolation between the PA output and the VCO tank (course VCO: Q = 14.9, Posc = −4.7 dBm), for the PA's fundamental on a VCO at fc and its second harmonic (Chapter 16) on the course VCO at 2fc, halved by the divide-by-two. (b) LO frequency error through Chapter 12's loop (fu = 400 kHz) for a coherent injection that grows with a 2 µs PA ramp, in quadrature with the oscillation (the worst phase), for three isolations.

Common pitfall · Taking the isolation at the wrong node

The PA’s 2f0 traps keep the second harmonic off the output pin, −75.0 dBc, but inside the PA the class-AB collector current has a second harmonic only about 15 dB below the fundamental (Chapter 16’s conduction-angle analysis gives −15.3 dBc at 240°), and it circulates through the trap, the bond wires and the ground. That loop, not the antenna port, is what couples magnetically to the VCO’s spiral. The 40 dB in this section is an assumption about the whole path; the number that decides it comes from an electromagnetic simulation of the layout, the job of Chapter 18.

Reference and fractional spurs

Chapter 12’s estimate, −112 dBc for 1 nA of leakage and 4 % pump mismatch, is 52.7 dB inside the course requirement of −59.3 dBc, so coupling will set the real spur, as it did in the taped-out SG13G2 PLL (−40.2 dBc reported [106]). Turning Eq. (12.12) around gives the budget for coupling. A ripple of amplitude \(V_r\) at fref on the tuning line makes a spur \(S\) (dBc at the VCO, 6 dB above the LO) if

\[V_r = \frac{2f_{ref}}{K_{VCO}}\,10^{S/20}. \tag{19.5}\]

For the course requirement this is 1.32 mV at 40 MHz with KVCO = 131 MHz/V, 1.04 mV at 167 MHz/V, the top of the calibrated range; for BLE’s emission limit alone, 12.2 mV. A millivolt of 40 MHz on a high-impedance node that runs past the reference buffer and the divider is easy to collect, which is why the third pole of Chapter 12 and a guarded, short tuning line matter. The fractional spur at 2 MHz, −67.0 dBc in Chapter 14’s worst case, needs to stay below −49.7 dBc for a −50 dBm interferer 2 MHz away from a wanted signal at −87 dBm: 17.3 dB to spare.

LO leakage and DC offset

With matched switches the four 25 % LO phases cancel at the mixer’s RF node (Chapter 10). Take the gate-drain capacitance as 30 % of the 79.6 fF switch gate and a 1 % mismatch between phases (both assumptions): a residue of 0.24 fF, driven by the 0.54 V fundamental of each 25 % pulse, injects 1.98 µA at 2.44 GHz. Into the 18 Ω of the RF node (the LNA’s 50 Ω output in parallel with the mixer’s 28 Ω input) that is 35.8 µV, −78.9 dBm at the mixer. Through the LNA’s reverse isolation (S12 = −38.1 dB) and the filter, −119.0 dBm reaches the antenna, below the thermal noise in 1 MHz. The same leak, reflected and self-mixed, makes a DC offset of microvolts.

The DC offset that does matter comes from the TIAs’ OTAs (σ = 0.11 mV each, Chapter 10), whose offset Chapter 10’s noise gain of 1.71 multiplies. In a low-IF receiver it sits 1 MHz below the channel, and the complex filter attenuates zero frequency by 25.3 dB with the 1 MHz filter and 13.7 dB with the recommended 1.4 MHz filter, before the ADC. A DC servo or a high-pass corner of a few tens of kilohertz, which Chapter 5 found costs nothing at a 1 MHz IF, keeps the remainder from eating the ADC’s range.

19.5 Calibration and digital assistance

Several of the margins above exist only after calibration. Table 19.5 lists them; Figure 19.9 shows two.

Table 19.5. Calibrations of the course radio: what, why and how well (numbers from this chapter's models).

What Why How Result
VCO band 16 bands of 32.6 MHz; process and temperature move the tank count LO cycles for 10 reference periods (0.25 µs), binary search over 4 bits (Section 11.8 simulates it) about 5 µs per channel; bands overlap 3.5 times
DTC gain residual must stay below 21 % (Figure 19.4b) sign-sign LMS on the bang-bang phase-detector output 30 % → < 10 % in 12.9 µs, < 2 % in 18 µs (μ = 4 × 10−4)
Two-point gain Δf2avg/Δf1avg ≥ 0.8 needs γ ≥ −10 %; KVCO spreads −21 % to +27 % after band calibration measure KVCO: step the fraction, read the tuning voltage (or LMS on the phase error) to a few percent
I/Q phase and gain IRR ≥ 30 dB with margin factory: test tone, trim the divider’s clock-input offset (0.051° per 0.5 mV step) and the VGA gain (0.02 dB) 1 % point of IRR from 34.7 to 64.6 dB
DC offset TIA op-amp offsets eat ADC range DC servo or high-pass of tens of kHz no BER cost at 1 MHz IF (Chapter 5)
Carrier offset up to ±150 kHz at the transmitter (BLE’s limit), plus the receiver’s own LO error estimate on the preamble, track over packets 0.6 dB from the preamble alone
PA power optional in BLE (LE Power Control, steps of at most 8 dB); saves energy on short links lower the drive, or scale the PA supply at 0 dBm: 10.7 % or 20.4 % collector efficiency

DTC gain. The DTC delays each divider edge by its predicted timing error, a swing of 1.6 ns for the MASH 1-1-1. If its gain is off by δ, the phase detector sees δ times the error, correlated with the prediction. A bang-bang detector’s sign, multiplied by the sign of the prediction, therefore points the way to the right gain, and a sign-sign LMS loop [176] steps the gain by μ each reference cycle. We ran it on the bit-exact MASH 1-1-1 sequence of the 2442 MHz channel from Chapter 14, with 0.5 ps of random timing noise standing for the loop’s own phase noise and a starting error of 30 % (Figure 19.9a). With μ = 4 × 10−4 the error falls below 10 % in 12.9 µs and below 2 % in 18 µs, and wanders by 0.08 % rms afterwards; with μ = 10−4, 51.4 and 71.9 µs, 0.04 %. At δ = 2 % the 3 MHz phase noise improves from −124.3 dBc/Hz (the 10 % residual) to −127.0 dBc/Hz. The model ignores the DTC’s nonlinearity. Chapter 14 models it: a 10 ps bow and 2 ps segment steps leave a −105.9 dBc/Hz floor and a −67.5 dBc spur at 2 MHz even with the gain calibrated, so the INL needs its own calibration; with a 16-segment pre-distortion [176] the synthesizer meets both specifications, and the DTC is worth an effective 2.4 % residual at 3 MHz. That calibrated DTC is the one this chapter’s synthesizer uses (Section 19.2).

Two panels. (a) DTC gain error in percent against time from 0 to 100 microseconds: starting at 30 percent, the mu equal 4e-4 curve drops linearly to about 0 by 15 microseconds and the mu equal 1e-4 curve by about 75 microseconds; both then stay within a fraction of a percent of zero. A teal band from minus 10 to 10 percent is labelled Chapter 14's 10 percent assumption. (b) Collector efficiency against output power from minus 20 to 10 dBm: both curves start at 33.7 percent at 10 dBm, marked Chapter 16; with the drive reduced at a fixed 2.0 volt supply the efficiency falls to 10.7 percent at 0 dBm and 3.4 percent at minus 10 dBm; with the supply scaled it falls more slowly, to 20.4 and 9.1 percent.
Figure 19.9. A one-bit LMS loop calibrates the DTC far below Chapter 14's 10 % within microseconds; scaling the PA supply doubles its efficiency at reduced power. (a) Sign-sign LMS of the DTC gain on the MASH 1-1-1 timing error of the 2442 MHz channel (40 MHz reference, 0.5 ps of random timing noise), from a 30 % error, for two step sizes. (b) Collector efficiency of the course PA against output power, from Chapter 16's 33.7 % at +10 dBm: lower drive at a fixed 2.0 V (efficiency ∝ output amplitude), or a supply scaled with the output amplitude above a 0.6 V knee (a simple class-B model).

PA power control. BLE does not require it. The Core Specification bounds only the output power at the maximum setting, measured at the antenna connector, which sets the power class [184]: Class 1.5 covers more than +4 up to +10 dBm, so the course PA’s +10 dBm is the top of that class, and the +8.0 dBm that reaches the antenna through the shared filter and switch lies inside it. Power control is optional, in steps of at most 8 dB under LE Power Control. Still, a link close to its peer should not spend +10 dBm. Lowering the drive keeps the 2.0 V supply and the quiescent current, and a class-B-like stage loses efficiency in proportion to its output amplitude: 33.7 % at +10 dBm becomes 10.7 % at 0 dBm and 3.4 % at −10 dBm (Figure 19.9b). Lowering the supply instead, with a DC-DC converter, keeps the stage near saturation; only the 0.6 V knee of the cascode, now a larger part of the supply, costs efficiency: 20.4 % at 0 dBm (1.04 V) and 9.1 % at −10 dBm (0.74 V). This is average power tracking, which the smartphone of Section 19.6 uses for its cellular PAs.

What this means for your design

Calibration is part of the architecture, not a repair. Each loop above needs a measurement path (a counter, the phase detector’s sign, a test tone, the ADC), a place to act (a capacitor bank, a DTC gain register, an offset DAC, a supply), and time in the radio’s schedule. Plan them with the blocks: the divider’s 0.1° per millivolt that causes the I/Q error is also the knob that trims it.

19.6 Case study: a smartphone RF front-end

The course radio is a single-standard, single-antenna radio. A smartphone carries a dozen of them. Pretl’s course notes [1] end with a walk through the published schematics of the Fairphone (Gen. 6), and the comparison is instructive. The summary below paraphrases that appendix; the part numbers and functions come from it, and through it from the schematic labels.

Architecture. The phone covers 2G to 5G below 7.125 GHz, Wi-Fi 6E with 2 × 2 MIMO, Bluetooth, satellite navigation (GNSS) in two bands and NFC. The application processor and the modem share one SoC; one RF transceiver IC handles all cellular bands and GNSS and exchanges I/Q samples with the modem over a fast serial link; a separate combo chip handles Wi-Fi and Bluetooth with its PAs on chip. The cellular transceiver has four receive groups, for 4 × 4 downlink MIMO, two transmit chains, a feedback receiver input and two GNSS inputs, and no IF filters: it is a direct-conversion transceiver, the zero-IF architecture of Chapter 5. Its LOs cover more than six to one, from 617 MHz to 3.8 GHz, with divide-by-two stages that also make the quadrature, as in the course radio; by Pretl’s estimate that takes three VCOs per synthesizer.

Front end. Everything the course radio leaves to “band filter + T/R switch” is a module there. Eleven antenna feeds serve MIMO, concurrent radios and sounding; diplexers and triplexers share antennas between bands (one antenna carries a cellular downlink band, GNSS L1 and 2.4 GHz Wi-Fi); antenna tuners match the low-band antennas; most frequency-division bands need their own acoustic duplexer, whose transmit-to-receive gap of 10–20 MHz asks for a quality factor near 100. The LNAs are separate modules ahead of the transceiver, because the losses in front of them (diplexer, switch, filter) would otherwise add directly to the NF. The multimode PA modules each have their own DC-DC converter for average power tracking, and a directional coupler feeds a sample of the output back to the transceiver for power control, I/Q and LO-leakage calibration and predistortion: the calibration paths of Section 19.5, made explicit. Dual uplink in two bands at once creates intermodulation products that land in the phone’s own receive bands; keeping them below the noise would need an IIP3 near +85 dBm, so the standard permits a defined loss of sensitivity instead.

The course radio in such a phone. Bluetooth shares the 2.4 GHz ports and antennas with Wi-Fi. Suppose the course receiver sat behind the same kind of front end (Figure 19.10): a diplexer, an antenna switch and a filter, which we assume lose 0.4, 0.7 and 1.5 dB (2.6 dB in total, instead of Table 5.3’s 2 dB). The NF rises from 4.62 to 5.22 dB and the sensitivity to −96.5 dBm, still 6.5 dB inside the course’s −90 dBm.

Block diagram of two radios in a phone. Top: a shared antenna feeds a diplexer of 0.4 dB, which also serves GNSS L1 and a cellular downlink; its 2.4 gigahertz branch goes through an antenna switch of 0.7 dB, also serving Wi-Fi, a coexistence filter of 1.5 dB in orange, the LNA and the course receiver, with NF 5.22 dB and a sensitivity of minus 96.5 dBm; a brace marks 2.6 dB in front of the LNA, where Table 5.3 allowed 2 dB. Bottom: a cellular transceiver drives a PA, supplied by a DC-DC converter for average power tracking, through a duplexer or filter to the cellular antenna. A note lists the LTE uplinks next to the ISM band: band 40, 2300 to 2400 megahertz, plus 23 dBm; band 41, 2496 to 2690 megahertz, plus 26 dBm; band 7, 2500 to 2570 megahertz, plus 23 dBm. A dashed orange arrow from the cellular antenna to the shared antenna says: 15 dB antenna isolation, plus 8 dBm for band 40 and plus 11 dBm for band 41 at the Bluetooth port.
Figure 19.10. Inside a phone, the course receiver sits behind 2.6 dB of shared front end, and 15 dB from an LTE transmitter. The assumed losses and isolation of Section 19.6; the coexistence filter (orange) is the block that Figure 19.11 sizes.

The real issue is the phone’s own transmitters. LTE band 40 (2300–2400 MHz) ends at the ISM band’s lower edge, and band 41 (2496–2690 MHz) and band 7’s uplink (2500–2570 MHz) start 12.5 MHz above its upper edge; the Fairphone supports all three, band 41 at the higher power class. Assume 15 dB of isolation between the cellular and Bluetooth antennas. A band-40 uplink at +23 dBm, centred 12 MHz below the 2402 MHz channel, then arrives at the Bluetooth port at +8 dBm; band 41 at +26 dBm, 26 MHz above 2480 MHz, at +11 dBm. How much can the course receiver take? Reciprocal mixing sets the limit: a blocker that spreads, through the LO’s phase noise, as much power into the channel as the thermal noise desensitizes the receiver by 3 dB:

\[P_{B,max} = \left(-174 + \mathrm{NF} + 10\log_{10}B\right) - \overline{\mathcal{L}} - 10\log_{10}B, \qquad R \ge P_{TX} - I_{ant} - P_{B,max}, \tag{19.6}\]

where \(\overline{\mathcal{L}}\) is the LO’s phase noise averaged over the offsets the blocker occupies (a 20 MHz LTE channel here), \(R\) the rejection a filter must provide and \(I_{ant}\) the antenna isolation. For band 40, 2–22 MHz from the 2402 MHz channel, the designed synthesizer averages −134.5 dBc/Hz and PB,max = −34.3 dBm; for band 41, 16–36 MHz from 2480 MHz, it averages −149.5 dBc/Hz, so reciprocal mixing would allow −19.2 dBm. Further out, compression comes first. The OTA mixer compresses by 1 dB at −3.5 dBm at its input (Section 19.2), which is −18.5 dBm at the phone’s antenna port, 2.4 dB before the cubic estimate from the front end’s IIP3 (−16.1 dBm). A blocker 3 dB below the compression point costs the wanted signal about 1 dB of gain (Section 4.7), so −21.5 dBm is the limit for bands 41 and 7. Figure 19.11 shows the result: band 40 needs 42.3 dB of rejection at 2390 MHz, band 41 32.5 dB at 2506 MHz and band 7 29.5 dB at 2510 MHz, while the assumed three-pole band filter provides 4.5, 10.9 and 12.7 dB. A coexistence filter with steep acoustic skirts must find the remaining 37.8, 21.6 and 16.8 dB, within about 10 MHz of the band edge. No receiver design removes the other half of the problem: the LTE transmitter’s own noise inside the ISM band, which must be below −122.8 dBc/MHz (relative to +23 dBm) to stay 6 dB under the course receiver’s noise floor. That falls to the cellular side’s filters, and, where they cannot reach it, to the two radios taking turns in time.

Power at the Bluetooth antenna port in dBm against frequency from 2280 to 2620 megahertz. The ISM band 2400 to 2483.5 megahertz is shaded. A blue curve, the largest blocker the course receiver tolerates for 3 dB desensitization, sits at minus 21.5 dBm, set by the mixer's compression, far from the band and above the band, and dips only close to the band edges, to about minus 34 dBm at 2390 megahertz and minus 40 dBm within a few megahertz of the band. Orange filled triangles mark the LTE uplinks at the port with 15 dB isolation: band 40 at 2390 megahertz, plus 8 dBm, and band 41 at 2506 megahertz, plus 11 dBm; open triangles the same after the assumed band filter, plus 3.5 and 0.1 dBm. Arrows from the open triangles down to the blue curve are labelled B40: a further 37.8 dB needed, B41: a further 21.6 dB needed. Orange bars along the bottom mark LTE band 40 and band 41.
Figure 19.11. Next to an LTE transmitter in the same phone, the course receiver needs 22–38 dB more rejection than a simple band filter gives, within 10–25 MHz of the ISM band. LTE band-40 and band-41 uplinks (+23 and +26 dBm, nearest 20 MHz channels) at the Bluetooth antenna port for an assumed 15 dB of antenna isolation, before and after the assumed third-order band filter, against the largest 20 MHz-wide blocker the course receiver tolerates for 3 dB of desensitization (Eq. (19.6) with the designed synthesizer's phase noise averaged over the blocker, limited 3 dB below the OTA mixer's simulated 1 dB compression point), referred to the nearest BLE channel (2402 or 2480 MHz).

What this means for your design

In a phone, a receiver’s hardest numbers come from its neighbours, not from its own standard. BLE’s blocking test asks for −35 dBm; the phone’s band-40 transmitter delivers +8 dBm. The course synthesizer’s far-out phase noise and the front end’s compression set the receiver’s half of the answer, and a filter, the antenna layout and the radios’ scheduling set the rest. Chapter 5’s budget would be written differently for such a product: start from the coexistence cases, and let BLE’s tests be the easy ones.

19.7 Where to go next

Lay it out and tape it out. The course has carried the design to schematics with EM-fitted passives; Chapter 18 shows the next steps, layout, design-rule checks and electromagnetic extraction, on the millimetre-wave PA’s matching networks and the pads, and the same steps apply to every block here. IHP runs open multi-project-wafer (MPW) shuttles for open-source designs in SG13G2 [191, 211]. The April 2025 run alone carried a 160 GHz LNA, 180 GHz PAs, a 5 GHz upconversion mixer, VCOs, a GPS-band LNA, RISC-V systems and Tiny Tapeout shuttles [211, 5]. Small dies with GSG pads for the LNA, the VCO with its divider, and the PA are realistic first chips; the full transceiver needs the calibration logic of Section 19.5 as well.

Measure. Each block’s measurement checks one assumption of this chapter. The LNA’s S-parameters, after the open–short de-embedding of Chapter 3, and its noise figure by the Y-factor method check Chapter 9 and the pad capacitance; the mixer’s conversion gain, noise figure and compression check Chapter 10’s simulation with the transistor-level OTA; a phase-noise measurement of the VCO and the closed loop checks Chapter 11’s simulated phase noise and Chapter 14’s DTC model; a load-pull of the PA checks Chapter 16’s network loss; and a bit-error-rate test with a BLE tester closes the loop on Figure 19.6.

Close the thin spots. The weakest numbers are known: the mixer’s OTA with real capacitors and resistors, a layout and its corners (Chapter 10 simulates it at the typical corner only, and its noise figure with behavioural switches), a transformer or a third-harmonic termination for the PA (Chapter 16), the corners of the mixer and the synthesizer, and a transistor-level complex filter, VGA and ADC from the sibling course, The Art of Analog IC Design [4]. Each would turn an assumption of this chapter into a simulation.

Try it · The transceiver script

Every number in this chapter comes from tools/figs/ch19_transceiver.py and its receiver model tools/figs/_ch19_rx.py: python3 tools/figs/ch19_transceiver.py redraws everything in about a minute from cached behavioural runs, and rx --resim reruns the receiver (about ten minutes). Three experiments: (1) In ASSUME, lower the band filter and switch loss from 2 to 1 dB (filter_switch) and run cascade blocks: how much of the 1 dB reaches the sensitivity, and why more than half? (2) In rx_part, change the 1.4 MHz filter (1.4e6) to 1.2 MHz and run rx --resim: how much of the +150 kHz sensitivity do you keep, and how much of the +2 MHz interference margin do you win back? (3) In ASSUME, set ramp_us to 0.5 and run coupling: how large is the kick for 20 dB of isolation in Figure 19.8b, and does it still settle before a packet could start?

  • Checked against Table 5.3, the receiver, synthesizer and off-chip PA (with Chapter 16’s harmonic traps) pass; only the on-chip PA fails, on output power and PAE, and the mixer’s gain and NF, now simulated with a transistor-level OTA in its TIAs (Chapter 10), sit just inside their targets, by 0.05 and 0.12 dB.
  • The designed receive chain reaches NF 4.62 dB, IIP3 −11.6 dBm, an IIP2 of at least +25.9 dBm at the antenna and a sensitivity of −97.1 dBm; the band filter and switch carry 55 % of the noise, and the mixer’s NF could be 9.6 dB worse before the 8 dB specification is reached. The mixer’s OTA compresses first: 1 dB at −19.1 dBm at the antenna.
  • In the behavioural receiver, Table 5.3’s 1 MHz channel filter loses 5.2 dB to BLE’s ±150 kHz carrier offset; a 1.4 MHz filter keeps the sensitivity at −96.2 dBm or better. With it every BLE test passes: interference by at least 6.1 dB (at +2 MHz), blocking by 15.7–16.1 dB (the far blockers by 15.9 dB, the mixer’s 1 dB compression point), intermodulation by 19 dB.
  • The divider’s mismatch gives 0.73° rms of I/Q phase error, more than 2° at 3σ, yet 99 % of chips exceed 34.7 dB of IRR; the synthesizer, with Chapter 11’s simulated VCO and Chapter 14’s calibrated DTC, has 2.7 dB of margin at 1 MHz and 7.0 dB at 3 MHz; a cancellation residual above 21 % would fail 3 MHz, and a sign-sign LMS calibrates the DTC’s gain in under 20 µs.
  • The transmitter delivers +10 dBm at 31.0 % PAE but only +8.0 dBm through the shared filter and switch; with Chapter 16’s harmonic traps both harmonics clear the ETSI and FCC limits by 20 dB or more. Its spectrum (−49.3 dBm at 2 MHz) and modulation (Δf2avg/Δf1avg = 0.879) pass once the two-point gain is calibrated to γ ≥ −10 %.
  • Running the VCO at 2fc cuts PA pulling by 75.0 dB; reference spurs will be set by coupling (1.3 mV of 40 MHz ripple on the tuning line is enough to fail), and in a phone the LTE transmitters next to the ISM band need a coexistence filter of 17–38 dB beyond a simple band filter.

Further reading

  • BRICD, Designing a Serial Link from Scratch, Part III · Transceiver Circuits: the transmitter and receiver circuits of a 5 Gb/s serial link in SG13G2, the wireline counterpart of this chapter’s radio.
  • H. Pretl, Radio-Frequency Integrated Circuits [1]: course notes from JKU Linz whose last appendix walks through the RF front end of the Fairphone (Gen. 6), the source of Section 19.6.
  • F.-W. Kuo et al. [174] and S. Binsfeld Ferreira et al. [175]: a complete Bluetooth Low Energy transceiver with a 2.75 mW receiver and a switchable on-chip TX/RX matching network, and the system design behind its receiver; a measure of how far the course radio’s receive chain, at least 24.9 mW before its baseband, is from the state of the art.
  • B. Razavi, RF Microelectronics [6]: transceiver architectures, the link budget and its interactions, and injection pulling, at much greater depth.
  • S. Pamarti, L. Jansson and I. Galton [149]: a 2.4 GHz fractional-N synthesizer with 1 Mb/s in-loop modulation, the alternative to the course’s two-point modulator.
  • S. Levantino, G. Marzin and C. Samori [176]: adaptive calibration of the gain and the nonlinearity of a DTC in a fractional-N PLL.

Back to top

SkillSurf · RFIC Design with Open-Source Tools, draft v0.1 (November 2026). Every source is listed on the References page.

This course was written with the help of AI tools and may contain mistakes. If you find one, please tell us through SkillSurf or email skillsurfedu@gmail.com, and we will correct it.

This site uses Just the Docs, a documentation theme for Jekyll.