10/22/2016 Design of Analog Integrated Systems (ECE 615) Lecture 4 Nyquist Rate Digital-to-Analog Converters Ayman H. Ismail Integrated Circuits Laboratory Ain Shams University Cairo, Egypt ayman.hassan@eng.asu.edu.eg ECE615 – Lecture 4 Outline • Monotonicity • DAC types and architectures • Resistor based architectures – Resistor String DAC – Resistor String DAC with Digital Decoder – Resistor-String DAC Including Interpolation – Folded Resistor String DAC • Capacitor based architectures – Charge scaling DAC – Charge scaling DAC with Split Array – Charge-Redistribution Switched-Capacitor DAC Ayman H. Ismail ECE615 – Lecture 4 2 1 10/22/2016 Outline • Current source Based architectures – Thermometer current source DAC – current source DAC binary weighted – Segmented current source DAC • Segmented/Hybrid DAC • Components Mismatch – current sources – Resistors and capacitors Ayman H. Ismail ECE615 – Lecture 4 3 Outline • DAC design – Static DAC Errors (INL/DNL) • Thermometer • Binary weighted • Segmented – Dynamic errors Ayman H. Ismail ECE615 – Lecture 4 4 2 10/22/2016 Monotonicity • Recall from first lecture:– A monotonic DAC converter is one in which the output always increases as the input increases. In other words, the slope of the DAC transfer response is of only one sign. – If the maximum DNL error is less than 1 LSB, then a D/A converter is guaranteed to be monotonic..Similarly, a converter is guaranteed to be monotonic if the maximum INL is less than 0.5 LSB . – The equivalent term for ADC converters is missing codes. An A/D converter is guaranteed not to have any missing codes if the maximum DNL error is less than 1 LSB or if the maximum INL error is less than 0.5 LSB. Ayman H. Ismail ECE615 – Lecture 4 5 DAC Types and Architectures • DACs are composed of current, voltage, and/or charge based elements. Examples for these DACs categories are: – Current source based – Resistor string – Charge scaling , respectively • Architectures – Thermometer (unit element) – Binary weighted – Segmented • In general each of the different DAC categories can be implemented as thermometer, binary or segmented. Ayman H. Ismail ECE615 – Lecture 4 6 3 10/22/2016 Resistor String DAC • 2N Rs generate 2N equally spaced voltage values • Inherently monotonic • Disadvantages: – 2N resistors & ~2.2N switches – High settling time, τmax= 0.25 x 2NR.C (Speed power trade-off). • Actually settling time may be much higher due to switches delay. • Use digital decoding, 2N transistor junctions at buffer input, though. Ayman H. Ismail Unit element 3-bit DAC ECE615 – Lecture 4 7 Resistor String DAC with Digital Decoder • 2N Rs generate 2N equally spaced voltage values • Inherently monotonic • Disadvantages: – 2N resistors & ~2.2N switches – High settling time, τmax= 0.25 x 2NR.C (Speed power trade-off). • Actually settling time may be much higher due to switches delay. • Use digital decoding, 2N transistor junctions at buffer input, though. Ayman H. Ismail Unit element 3-bit DAC with digital decoder ECE615 – Lecture 4 8 4 10/22/2016 Resistor-String DAC Including Interpolation Unit element • Buffers may limit performance: – Random offset – Speed • 2.2N/2 resistors. Suitable for high resolutions • Since the second resistor string is used to decode only the lower-order bits, the matching requirements of the second resistor string are relaxed (more about matching requirements later) • The first resistor string must be accurate to N-bits. 6-bit DAC Ayman H. Ismail ECE615 – Lecture 4 9 Folded Resistor String DAC Unit element • folded resistor-string D/A reduces the amount of digital decoding and large capacitive loading. • Total number of transistor junctions at buffer input 2√2N. • Still 2N resistors are needed. 4-bit DAC Ayman H. Ismail ECE615 – Lecture 4 10 5 10/22/2016 Charge Scaling DAC B-bit DAC Binary Weighted • In reset phase all cap’s are discharged. • In evaluation phase, depending on the digital word, the caps either remain connected to ground or are connected to Vref, and DAC operates as a capacitor divider. • DAC output sensitive to parasitics • Monotonicity depends on element matching (as explained later) Ayman H. Ismail ECE615 – Lecture 4 11 Charge Scaling DAC B-bit DAC Binary Weighted • OTA helps eliminate the parasitic capacitor effect (parasitic cap at virtual ground). • Total cap = 2BC • OTA issues – Offset/ flicker noise – Speed Ayman H. Ismail ECE615 – Lecture 4 12 6 10/22/2016 Charge Scaling DAC with Split Array 6-bit DAC Binary Weighted • Total cap (2.2n/2-1) C, instead of 2nC • For 6 bits 15C, instead of 64C, area saving • OTA Common mode has to extend to all values of VDAC out • Sensitive to parasitic cap Ayman H. Ismail ECE615 – Lecture 4 13 Charge-Redistribution SwitchedCapacitor DAC 4-bit DAC Binary Weighted • Offset/ flicker noise insensitive • Similar to capacitive reset switched-capacitor amplifier presented earlier, but with programmable gain (switched cap. Circuits – lecture 2) Ayman H. Ismail ECE615 – Lecture 4 14 7 10/22/2016 Thermometer Current Source DAC B-bit DAC Unit Element • Monotonicity does not depend on element matching • 2B-1 current sources & switches • Need large encoder with 2B-1 outputs. Therefore, impractical for large B (high resolution) Ayman H. Ismail ECE615 – Lecture 4 15 Current Source DAC Binary Weighted B-bit DAC Binary weighted • Monotonicity depends on element matching • B current sources & switches . However, 2B-1 unit elements as for unit element implementation. • No encoder needed Ayman H. Ismail ECE615 – Lecture 4 16 8 10/22/2016 Segmented DAC 1 • B-bits DAC • B= Bt + Bb • MSB bits (Bt) are implemented using unit elements (thermometer). Therefore, guaranteed monotonicity. • LSB implemented using binary weighted elements. Monotonicity depends matching. However, matching requirements are relaxed in this case (as discussed later) Ayman H. Ismail ECE615 – Lecture 4 17 Segmented DAC 2 • Segmented/Hyprid (different types used) • Example: 12bit DAC • 6-bit MSB DAC using R string • 6-bit LSB DAC using binary weighted charge redistribution • Segmented results in Complexity much lower than full R string • Full R string4096 resistors • Segmented 64 R + 7 Cs (64 unit caps) • Why use hyprid? Ayman H. Ismail Sub-DAC (6bit accurate) Main DAC (12-bit Accurate) ECE615 – Lecture 4 18 9 10/22/2016 Components Mismatch • Systematic errors – routing – Contact resistance – Edge effects in capacitor arrays – Process gradient – Finite current source output resistance • Random errors – Often Gaussian distribution (central limit theorem) • Systematic error should be minimized/ eliminated in the design phase. In this case, the circuit is limited by “random” errors. Ayman H. Ismail ECE615 – Lecture 4 19 Random Mismatch in Current Mirror • Mismatch in current mirrors can be described as • Offset voltage/current is function in σ∆Vth and σ ∆β/ β, which are given by where Avth and Aβ are technology dependent constants, W and L are the device dimensions • Hence, the larger the area of the device, the smaller the random error. The same applies to resistors and capacitors. Ayman H. Ismail ECE615 – Lecture 4 20 10 10/22/2016 Random Mismatch in R and C • For resistors R1 R2 • Similarly for capacitor • Note that mismatch defines the value of unit cap, but does not define the value of resistor Ayman H. Ismail ECE615 – Lecture 4 21 The Gaussian Distribution Ayman H. Ismail ECE615 – Lecture 4 22 11 10/22/2016 Yield Ayman H. Ismail ECE615 – Lecture 4 23 Yield Ayman H. Ismail ECE615 – Lecture 4 24 12 10/22/2016 Thermometer DAC Static Errors (DNL) • The LSB can be expressed as • DNL can be expressed as • Therefore Ayman H. Ismail ECE615 – Lecture 4 25 Thermometer DAC Static Errors (INL) Ayman H. Ismail ECE615 – Lecture 4 26 13 10/22/2016 Thermometer DAC Static Errors (INL) • Max error • Therefore Ayman H. Ismail ECE615 – Lecture 4 27 Binary Weighted DAC Static Errors (INL and DNL) • INL same as for the thermometer DAC case • DNL depends on transition • Worst case at mid-scale Ayman H. Ismail ECE615 – Lecture 4 28 14 10/22/2016 Segmented DAC Static Errors (INL and DNL) • INL same as for the thermometer DAC case • DNL depends on transition • Worst case occurs when LSB DAC turns off and one more MSB DAC element turns on • Same DNL as a binary weighted DAC with Bb+1 bits Ayman H. Ismail ECE615 – Lecture 4 29 Unit Element, Binary Weighted and Segmented Comparison Ayman H. Ismail ECE615 – Lecture 4 30 15 10/22/2016 For 12-bit DAC • For B=12, σu=1% Ayman H. Ismail ECE615 – Lecture 4 31 Example 1 • Assume a thermometer 6-bit DAC with σu = 1% and it is required to have 99.73% of all converters to meet the spec. What is the DNL spec that can be set? • For 99.73% yield, need X = 3 • σDNL = σu = 1% • DNL=3 σDNL = 3% • DNL specification for a yield of 99.73% is ±0.03 LSB • However, a converter will meet specs only if all codes meet DNL spec, i.e. DNL(i) < DNLspec for all i • A converter with more codes is less likely to have all codes meet the specification Ayman H. Ismail ECE615 – Lecture 4 32 16 10/22/2016 Example 1 • Assuming that all DNL(i) values are independent, then • • • • P(all codes meet spec) = P(single code meets spec)N P(all codes meet spec)1/N = P(single code meets spec) N=63 (6 bits) P(single code meets spec) =0.99731/63 = 0.99995708 • Therefore for all codes to meet the spec with probability of 99.73%. Each code, must meet the spec with prob. 0.99731/63. • X = sqrt(2)*erfinv(0.99731/63) = 4.09 • Refined result. For 99.73% yield, DNL spec should be ±0.0409 LSB Ayman H. Ismail ECE615 – Lecture 4 33 Example 2 • It is required to design an 8-bit a binary weighted charge scaling capacitive DAC. Assume capacitance matching coefficient of 1%.μm. (Note : If INL and DNL are not specified, assume target specifications INL=DNL=±0.5 LSB with 99.73% yield) • • • • • For 99.73% yield INL= 3σINL Therefore, σINL = 0.5/3= 0.167 LSB Similarly, σDNL = 0.5/3= 0.167 LSB For binary weighted Ayman H. Ismail ECE615 – Lecture 4 34 17 10/22/2016 Example 2 • From DNL spec σu= 0.167 /16 = 0.0625 =1.04% • From INL spec σu=2x 0.167/16 =2.08% • • • • • σu is limited by DNL σu=Ac/√Area, Ac= 1 %.μm. Cu area = (1/6.25)2 μm2= 0.924μm2 Assuming capacitance density= 20fF/ μm2 Cu = 18.5fF. Is this value OK? In general, what can be done if you get a very small Cu? • Total Capacitance= 28 x 18.5= 4736fF • Note that having a non zero INL and DNL results in an SNDR lower than the ideal value predicted by 6.02+1.76 dB expression Ayman H. Ismail ECE615 – Lecture 4 35 Example 3 • It is required to design a 4-bit a binary weighted charge scaling capacitive DAC with 8-bits of accuracy. Assume capacitance matching coefficient of 1%.μm. An example of such DAC is a DAC used as the Main DAC of a segmented 8-bit DAC. Note that the sub DAC needs to be accurate to the 4-bit level only. • (Note: If INL and DNL are not specified, assume target specifications INL=DNL=0.5 LSB) • • • • Referring to previous example C’u=16 x Cu=296fF Total cap= 16 x C’u=4736fF How does the area of the 4-bit, 8-bit accuracy DAC compare to that of an 8-bit DAC? • Then why use segmentation in the first place? Ayman H. Ismail ECE615 – Lecture 4 36 18 10/22/2016 Dynamic DAC Errors • Finite settling time and slewing – Finite RC time constant – Signal dependent slewing non-linearity • Feed-through – Coupling from switch signals to DAC output – Clock feed-through • Glitches due to timing errors – Different current sources not switching simultaneously . • Reducing dynamic errors, usually, requires higher power dissipation Ayman H. Ismail ECE615 – Lecture 4 37 Dynamic Errors: Glitches – I1 represents the MSB current, and I2 represents the sum of N-1 LSB currents. – If MSB current turns off slightly early, a glitch of zero current occurs. A glitch to the maximum current value occurs if both currents are temporarily ON Ayman H. Ismail ECE615 – Lecture 4 38 19 10/22/2016 References • Analog Integrated Circuit Design, David Johns and Ken Martin, Jon Wiley & sons, 2012. • Data Converters, Franco Maloberti, Springer 2007. • Haideh Khorramabadi, ECS 247(Analog-Digital Interface Integrated Circuits) Handouts, University of California, Berkeley, 2005 Ayman H. Ismail ECE615 – Lecture 4 39 Additional slide ECE615 – Lecture 4 20 10/22/2016 Binary Weighted Resistor DAC Ayman H. Ismail ECE615 – Lecture 4 41 R-2R Based Resistor DAC Ayman H. Ismail ECE615 – Lecture 4 42 21 10/22/2016 Dynamic Element Matching • Dynamic element matching (DEM)algorithms were introduced in 1974 by Klaas Klaassen [K.B. Klaasen, “Digitally controlled absolute voltage division,” IEEE Trans.Instrumentation and Measurement, vol. 24, no. 2, pp. 106-112, June 1975] who used DEM to obtain a constant division ratio from a voltage divider consisting of a mismatched resistors • DEM average out errors due to component mismatch and therefore achieves accuracy beyond intrinsic device matching A. H. Ismail ECE615 – Lecture 4 43 Dynamic Element Matching Algorithms • DEM is implemented by using different instances of circuit elements (transistors, resistors, caps) used to build the system each time the system the output analog signal is generated • Since different instances introduce different errors. The errors introduced by instances is averaged out • DEM is used to linearize DAC’s • Dynamic Element Matching Algorithms:– Data weighted averaging – Individual level averaging – Vector based mismatch shaping – Tree-structure element selection A. H. Ismail ECE615 – Lecture 4 44 22 10/22/2016 DATA-Weighted Averaging • The data weighted averaging DEM algorithm rotates circuit elements, insuring that each element is used the same number of times. • The circuit elements are selected sequentially from the array starting with the next available unused element. • Data weighted averaging generates tones at the output if the DAC input is not busy (DC input or low frequency input) • It can be shown that DWA is equivalent to shaping mismatch error (noise) by a first order high pass filter A. H. Ismail ECE615 – Lecture 4 45 Individual Level Averaging • Individual level averaging (ILA) is basically a DWA carried out for each code. • Individual level averaging rotates or flips circuit elements in a periodic fashion, but a separate rotation state is maintained for each digital level. • The advantage of the individual level averaging algorithm over DWA is that it ILA is less likely to generate tones even for DC or periodic inputs • However, compared to DWA, ILA converges more slowly to zero average condition A. H. Ismail ECE615 – Lecture 4 46 23