System-on-Chip. Electro-thermal effects in radio ICs. Overview. Tri-band GSM radio receiver (front-end)

Transcription

1 System-on-Chip Overview Pietro Andreani Dept. of Electrical and Information Technology Lund University, Sweden Material from Sensitivity Degradation in a Tri-Band GSM BiCMOS Direct-Conversion Receiver Caused by Transient Substrate Heating, by S. Mattisson, H. Hagberg, P. Andreani; IEEE Journal of Solid-State Circuits, Vol. 43, NO. 2, pp , Feb Front-end of tri-band GSM receiver The problem Mismatches in I/Q mixer core Electro-thermal modeling and simulations Simulations vs. measurements Layout options common centroid property Improved front-end design 2 Tri-band GSM radio receiver (front-end) The problem local oscillator low-noise amplifier quadrature generations mixer (double-balanced) DCS and PCS are made of identical blocks and work at almost the same frequencies (~ MHz) However, at high temperatures (75 o C), DCS shows a loss of sensitivity between 2dB and 4dB compared to PSC really bad! True puzzle!! Investigations show that ) the problem is tied to the DCS signal path; 2) a slow-settling (time constant of some 00 μs) DC component was detected at the differential base-band DCS/PCS outputs (largest in the DCS case); 3) the static performance of the receivers was not affected. System simulations confirmed the correlation between slow DC-offset variations and loss of sensitivity. low-pass filter 3 4

2 Slow-settling BB signal What could cause this slow BB transient? Prominent means of interaction between circuit blocks are ) substrate-induced electrical effects, 2) LO-leakage-induced offsets (through rectification), and 3) thermal gradients in the substrate However, only thermal effects are able to account for the presence of time constants as high as hundreds of microseconds seek a temperature-related cause for the slow-settling baseband DC offset GSM radios are non working continuously, but rather in a time-division fashion thermal transients appear at the beginning of each receive time slot Of the many places in a radio receiver where thermal effects may cause problems, the core of the double-balance mixer is a very strong candidate 5 6 Double-balanced mixer Core of double-balanced mixer LO+ LO+ RF+ LO- RF- mixer core Focus on e.g. leftmost pair A temperature difference ΔT between the two BJTs results in a ΔV BE difference of approx..5mv ΔT equivalent to driving the mixer core with a square wave S (from the LO) having a duty cycle 50% 7 8

3 Square wave from LO Down-converted currents T sqw 2π = ω 0 We assume: I t I0 is t 0t 2 I2 t I0 is t 0t 2 () = + () cos( ω ) () = () cos( ω ) signal bias t0 IF A =0, A 2 =, and = δ, we have: T 2 sqw ideal switching function DC and even-order distortion O() t = I0 + i () t cos ( ω0t) S() t ; 2 s O3() t = I0 i () t cos( ω0t) ( S2() t ); 2 s O2 t I0 i t 0t S t 2 s () = + () cos( ω ) ( ()) O4 t I0 i t 0t S2 t 2 s () = () cos( ω ) () S () t 2 = + cos( ω0 ) cos ( 0 )... 2 cos( 2 0 )... 2 t ω 3 t δ ω π t Current down-converted at baseband: 2 2 π π ( O + O ) 3 ( O + 2 O4) = i s 2I0( δ() t δ2() t ) = i s 2I Δ 0 δ () t 9 0 Is this the real cause? Electro-thermal simulations: what is required? A time-varying signal Δδ(t), amplified by twice the bias current, may account for the spurious baseband signal corrupting the desired signal However, is this mechanism the real one? Very difficult to say even estimate the impact of Δδ(t) power consumption levels in modern GSM receivers are very low (a few tens of mw at most) temperature gradient is therefore very small however, very demanding performances are required from the receivers An electro-thermal simulation is required: determines whether we are right about Δδ(t), and, in this case, can be used to find a better design (layout) It is highly desirable to use a standar circuit simulator such as Spice or Spectre; to do this, we need: Thermal model of the die Local interaction between thermal and electrical parameters Devices with local dynamic temperature dependence Instantaneous power dissipation in the devices 2

4 Equation for heat diffusion Equation for heat diffusion II (,,, ) u x y z t ρc = κ + + u x y z t p d x y z t t x x x u = temperature of the material ρ = = c= κ 6-3 density of the material ( gm for Si) (,,, ) (,,, ) - specific heat capacity of the material (=0.7 Jg K for Si) - o = thermal conductivity of the material ( = 20 Wm K at 75 C for Si) p = rate of heat production (i.e. dissipated power) density d discretization of the heat equation (in the x-direction): ( Δ ) ( ) ( ) ( +Δ ) u x x u x u x u x x ρcu& ( x) = κ + pd x ( ) ( ) ( ) κ ρcu& ( x) = u( x 2 ) u( x) u( x) u( x+δ x) + pd x power dissipated multiplying by the unit volume: in the unit volume ( ) ( ) ΔΔ y z ΔΔΔ x y zρcu& ( x) = κ u( x ) u( x) u( x) u( x+δ x) + p x We define the equalities below: C =ΔyΔzρc Rx = ΔΔ y z κ Cv& ( x) Rx i = p v( x x) v( x) v( x) v( x x) + i x v = u ( ) ( ) 3 4 Equation for heat diffusion III Augmented BJT model ( ) ( ) Cv& ( x) = v( x x) v( x) v( x) v( x x) i x R Δ +Δ + x The discretized heat equation can be solved numerically with an analog circuit simulator! (the discretization is of course the same in the y- and z-direction) 5 6

5 Electro-thermal model of mixer+substrate Discrete model of silicon die Grid uniform in the x and y direction, with = Δy = 20μm Grid geometric in the vertical direction: with Δz = 20μm 2 0,,2... Die is 300μm thick 4 vertical grid lines In all, the grid contains nodes the thermal circuit contains some devices (mostly R and C) connected to some 2000 nodes 7 8 A complication Measurements vs. simulations DCS We want to run a time-efficient simulation It would be desirable to include the actual RF LO drive in the simulation, but this is hardly time efficient Solution: mixer-core transistors are biased at mid-point any V BE offset generates a base-band offset The ratio of unswitched-bjt offset to switched-bjt offset was determined to 7.5 accordingly, the impact of the mixergenerated offset was scaled down by the same factor 7.5 However, the rather large power dissipation of the LO buffers was taken into account 9 20

6 Measurements vs. simulations PCS Simulations details and outcome Data amplified a factor 0 for ease of detection Comparison data taken after 00μs, when the frequency synthesizer has safely settled (and up to 800μs) Simulations match measurements with an error smaller than 0%! (apart from the Q channel in the PCS receiver, where the absolute error is however very small) Simulations can even capture the peak in the DCS Q-channel waveform with the correct time constant Conclusive proof that the sensitivity reduction is caused by thermal drift in the mixer cores 2 22 DCS vs. PCS DCS is more sensitive than PCS because of the different locations of the mixers (and LO buffers) with respect to the DCS/PCS LNAs Placement of mixer core temp. gradient temp. gradient from DCS LNA V BE < V BE2 and V BE3 < V BE4 δ positive and δ 2 negative Δδ = δ - δ 2 large (it would be much better if δ and δ 2 had the same sign!) More thorough analysis takes into account placement of LO buffers as well 23 24

7 Temperatures in mixer core Placement of mixer core II 85m o K (!) -D common-centroid layout V BE < V BE2 and V BE3 > V BE4 δ and δ 2 positive Δδ = 0 in presence of linear temperature gradient! Contrary to what one would guess at first sight on account of the LNA alone, it is the I-channel mixer core to experience the largest temperature variation, although the Q-channel mixer core is closer to the DCS LNA However, the Q-channel mixer core is placed more symmetrically, if both DCS LNA and LO buffers are considered (a more favorable overall placement explains the lower PCS sensitivity, and the GSM insensitivity) 25 However, temperature gradients are in general NOT linear commoncentroid can reduce their impact, but not totally cancel them Nevertheless, apart from brute-force separation of sensitive circuits, our best option is to exploit symmetry 2D layout rejects temperature gradients along its two axes of symmetry 26 Temperature and gradient Remarks about self heating in mixer core Self heating has been disregarded so far Self heating is very rapid (see again the DCS Q-channel transient, where the signal initially falls very quickly because of self heating) its impact is practically constant during the signal burst can be cancelled in the digital baseband Single hot-spot: normalized temperature (left) and normalized temperature gradient vs. distance gradient is clearly non linear, especially close to the hot spot more serious temperature-related problems expected with the ever increasing shrinkage of ICs Apart from this, if self heating is considered, it is the original layout to be D common centroid, not the modified one Also in this case, the optimal layout is the 2D common centroid, which rejects self heating as well 27 28

8 2D vs. D - simulations Why more problems at higher temperatures? LNAs are biased for a constant transconductance vs. temperature higher current at higher temperatures higher power dissipation higher thermal gradient 2D common-centroid mixer core provides a further x0 reduction of the thermally-induced baseband transient The thermal conductivity κ drops at higher temperatures less efficient heat diffusion in the substrate this has not been taken into account in the simulations for simplicity a temperaturedependent conductivity can be modeled with a non-linear resistor However, the 2D common-centroid layout has more parasitic capacitance Improved designs D mixer cores DCS with -D common-centroid mixer core D common centroid layout only needed a rewiring of the metal layers as opposite to a 2D solution, which would require a wholly new layout Simulations and measurements do not fit so well for the I- channels now however, the size of the offsets is now much smaller than before large relative errors, but small absolute errors! The measured transient are now almost a factor 0 smaller no more sensitivity loss at high temperatures! receiver could be released for production 3 32

9 PCS with -D common-centroid mixer core Final remark Before the electro-thermal simulation environment was available, it was thought that the main cause of baseband offset was the thermal gradient across the resistances at the output of the I/Q mixers A new common-centroid rewiring of these resistances was fabricated, but the improvement was only modest Electro-thermal simulations run shortly afterwards confirmed that these resistances played a minor role in the generation of the baseband offset 33 34