Low-Power Design of CML Drivers for On-Chip Transmission-Lines

Transcription

1 Low-Power Design of CML Drivers for On-Chip Transmission-Lines Akira Tsuchiya Takeshi Kuboki Hidetoshi Onodera Dept. Communications and Computer Engineering, Kyoto University Yoshida-honmachi, Sakyo-ku, Kyoto , Japan {tsuchiya, kuboki, Abstract In this paper, we discuss a design technique to reduce the power dissipation of CML buffers for on-chip transmissionlines. CML buffers can operate in higher frequency than conventional static CMOS buffers. On the other hand, the power dissipation is larger than that of the static CMOS buffers. We reduce the power dissipation by using an impedance-unmatched driver. From the pole frequency analysis, our method can decrease the tail current without degrading the throughput of the CML buffer. Experimental results show that our method can reduce the power dissipation by 25%. V out V in V DD R D M M 2 I tail R D V in2 V out2 I. Introduction According to the continuous improvement of the LSI fabrication technologies, the performance of LSIs is rapidly accelerating. One of the big challenges in high-performance LSIs is the interconnect bottleneck problem. The on-chip interconnection is becoming a limitation of the whole chip performance because the performance of on-chip wires does not improve by technology scaling. At the same time, the multi-core architecture is a trend of microprocessor design []. In such LSIs, on-chip buses have strong impact on the chip performance and on-chip interconnection is required to achieve higher bandwidth [2]. Thus on-chip high-speed communication is one of the hot topics and several methods are developed [3 7]. In high-speed communication, the primal elements of the signaling system are the driver, the receiver and the interconnect. The on-chip interconnects can transmit 0Gbps or higher bit rate [8] and the bottleneck is the driver circuit. To realize highspeed circuits, current-mode-logic (CML) is one of the solutions [9 2]. However CML buffers have more design parameters comparing static CMOS inverter [3]. CML buffers can operate in higher frequency and have tolerance to the common mode noise, but CML buffers require the static current flow and the power dissipation is larger than that of the static CMOS buffers. Thus low-power design of CML buffers is discussed [4]. This paper proposes a low-power design of CML buffers for on-chip transmission-lines. Conventionally, impedancematched driver is a common practice for transmission-line drivers [3]. The proposed method use an impedanceunmatched driver to reduce the power dissipation. Impedanceunmatched driver cause reflection of the propagating wave, however the reflected wave rapidly attenuates on on-chip transmission-lines. Therefore the impedance-mismatch is not a serious problem. We use a resistance larger than the characteristic impedance as the pull-up resistance. By using the larger resistance, we can reduce the tail current. From the pole frequency analysis, we can reduce the tail current without degrading the Fig.. CML differential buffer. bandwidth of the buffer. We experimentally verify the proposed method. The contribution of this paper is a low-power design method that does not degrade the bandwidth. Section II explains fundamentals of CML buffers for on-chip transmission-lines. In Section III, we show a relationship between the pole frequency and the bandwidth of CML buffer. Section IV describes the proposed method and Section V shows experimental results. Section VI concludes this paper. II. Fundamentals of CML Buffer In this section, fundamentals of CML buffer are explained. First, the basic operation of CML buffers is explained. Next, the design guideline of Ref. [3] is introduced. A. Basic operation of CML buffer Figure shows a basic CML buffer. The CML buffer is based on the differential architecture. The main components of the CML buffer are two pull-up resistors R D, two nmos transistors for switching and a current source I tail. NMOS transistors control the current flow of each side of the differential pair according to the differential input. CML buffers can operate in high frequency because no pmos transistor is used and the nmos transistors are always in saturation region. Figure 2 shows a transfer characteristic of the CML buffer. As the differential input (V in V in2 ) varies, each output voltage varies from (V DD R D I tail ) to the supply voltage V DD. Thus the range of the differential output voltage (V out V out2 ) is from R D I tail to R D I tail. By assigning each condition to 0 and, the CML buffer can transmit differential signal.

2 V DD V out V out2 N stages V DD - R D I tail R D I tail = V out V in - V in2 st 2 nd N th u u N- =X differential transmission-line V in, min Fig. 3. Tapered driver. Fig. 2. Transfer characteristics of CML differential buffer. B. Conventional design guideline of CML buffer The CML buffer shown in Fig. is the basic of differential amplifier [5] and a design guideline to use as a driver is already discussed [3]. From Fig., design parameters of a CML buffer are the pull-up resistance, the size of the nmos transistor and the tail current. Generally, the pull-up resistance R D is tuned to the characteristic impedance of the transmission-line to achieve impedance matching. The size of the nmos transistor is determined by the constraint of CML buffer operation. As shown in Fig. 2, a certain input voltage is needed to swing the output voltage from (V DD R D I tail ) to V DD. When all of the tail current flows through either nmos transistor, the output voltage swing reaches R D I tail. From the square law of nmos drain current, the relationship between the tail current and the minimum input voltage V in, min is expressed by I tail = 2 µc W ox L V2 in, min, () where µ is the mobility, C ox is the gate capacitance par unit area, W and L are the gate width and the gate length respectively. The mobility µ and the gate capacitance C ox are determined by the fabrication process and the gate length L is set to the minimum value in the fabrication process. Therefore the freedom in the nmos transistor is only the gate width W. To drive the next stage, the output voltage has to be larger than the minimum input voltage. V out V in, min. (2) From Eq. () and Eq. (2), The gate width W is determined by 2LI tail W. (3) µc ox Vout 2 Equation (2) is the lower limit of the output voltage. The upper limit of the output voltage depends on the threshold voltage V th. For high-speed operation, nmos transistors should operate in the saturation. From this constraint, the maximum output voltage is derived as [3] V out = R D I tail V th. (4) The last design parameter I tail is determined from the output voltage swing. The final stage of CML driver has to drive the transmission-line and the receiver. In on-chip transmission-line, loss of the transmission-line is not negligible. Therefore the lower bound of the output voltage V out is larger than the minimum input voltage of the receiver. By considering the atten- uation in the transmission-line, the output voltage V out is expressed by V out V in receiver, min, (5) exp( αl) where α and l are the attenuation constant and the length of the transmission-line respectively. According to this design guideline, the design parameters are fixed when the characteristic impedance of the interconnect and the required voltage swing is given. C. Tapered CML buffer On-chip CML driver has to drive a differential transmissionline. The characteristic impedance of on-chip differential lines is typically in the range from 50Ω to 200Ω. To drive such low impedance load, the tail current becomes large and the size of the switching transistor becomes large. Therefore tapered buffers are used and the number of stages and the taper factor are also design parameters. Figure 3 shows a tapered driver. In Fig. 3, the number of stages is N and the taper factor is u. In this discussion, the input stage of the tapered driver is written as the st stage. The last stage is the N-th stage. The transistor size and the tail current gradually scale up with the taper factor u. In opposite, the pull up resistance scales down with the taper factor u. The relationship between the k-th stage and the k + -st stage is expressed as R Dk+ = R Dk /u W k+ = uw k, (6) I tailk+ = ui tailk the subscript k + and k denote the (k + )-st stage and the k-th stage respectively. The parameter X is the ratio between the first stage and the last stage and is equal to u N. The latency of tapered CML buffer is discussed in Ref. [3] and Ref. [3] concludes that the delay of CML buffer has similarity with static CMOS buffer and the number of stages becomes optimal when the number of stages N satisfies N ln ( u N ). (7) In other word, the optimal taper factor is Napier s constant e. On the other hand, the bandwidth is also an important metric of CML driver. From the viewpoint of the bandwidth, the taper factor and the number of stages change the gain curve of the CML buffer. As the taper factor u becomes small, the pole frequency shifts toward high frequency because the load capacitance becomes small. However as the number of stages increases, the gain drops rapidly in high frequency region. An

3 20 5 stages 5 4 stages 0 V DD Voltage gain [db] 5 2 stages 0 3 stages -5-0 V out V out Frequency [GHz] V in I tail V in2 Fig. 4. Gain curve of tapered CML buffers. (X = 5) Fig. 5. Parasitic capacitances in CML buffer. example of gain curves is shown in Fig. 4. Figure 4 is the result of circuit simulation. The final stage of the buffer is designed to drive a differential transmission-line whose differential characteristic impedance is 00Ω. The ratio of the first stages and the final stage is fixed to 5 and we change the number of stages. As shown in Fig. 4, the gain of the 2-staged buffer starts to drop from the relatively low frequency. The gain curve of the 5-staged buffer remains to be flat, however, the gain drops rapidly in the higher frequency. Therefore on the bandwidth, it is not clear how many stages or taper factor maximize the bandwidth of CML buffer. III. Pole Frequency Analysis of Tapered CML Buffer In this section, we propose the pole frequency as an indicator of the bandwidth of CML buffers. As mentioned in the previous section, the design parameters R D, W and I tail are determined by the characteristic impedance of wire and the required voltage swing. Therefore in this section, the design freedoms are the taper factor u and the number of cascade stages. A. Pole frequency of CML buffers The voltage swing of CML buffers is limited because it should be smaller than the threshold voltage. Thus we can expect that we can estimate the performance of CML buffers by the small signal analysis. The pole frequency of CML buffer shown in Fig. is determined from the resistor and the capacitor connected to the drain of the switching transistor [5]. The pole frequency ω p is expressed as ω p = R D, (8) if there is no parasitics. Without considering the parasitic capacitances, the load capacitance is the sum of the drain-backgate capacitance and the gate capacitance of the next stage. Then the pole frequency of tapered CML buffers is expressed as ω p = R DN ( CDBN + uc GN ), (9) where R DN, C DBN and C GN is the pull-up resistance, the drainbackgate capacitance and the gate capacitance of the final stage of tapered buffer respectively. random NRZ st 2 nd N stages N th u u N- =X Fig. 6. Experimental circuit for eye-diagram evaluation. Eye-diagram evaluation bridge termination (00 ) In reality, there are several parasitic capacitances as shown in Fig. 5. The overlapping or adjacent wires and the polysilicon resistor are the main causes of the parasitic capacitances. Therefore we have to take these parasitic capacitances into consideration. B. Relationship between pole frequency and eye-diagram To evaluate the performance of CML buffers, we evaluate the eye-diagram by circuit simulation. The experimental circuit is shown in Fig. 6. The CML buffer is designed in accordance with the design guideline explained in Section II.. We use a 80nm CMOS process whose supply voltage is.8v. The required voltage swing is set to 0.48V in differential. The ratio between the first stage and the final stage X is fixed to 5 and the number of stages is changed. The input is a random Non-Return-to-Zero sequence. The input pulse shape of trapezoidal and the rise/fall time is one tenth of the minimum pulse width. The output of the buffer is terminated by the resistor that represents the characteristic impedance of the transmission-line. We evaluate the eye-diagram at the output of the buffer. Figure 7 shows the eye opening voltage and the input frequency. As the input frequency becomes higher, the eye opening voltage degrades. In Fig. 7, the lowest pole frequency of each configuration is also shown. We calculate the pole frequency from the pull-up resistor R D, the drain-backgate capacitance, the gate capacitance and the parasitic capacitances shown in Fig. 5. From Fig. 5, the pole frequency indicates the frequency where the eye opening voltage starts to degrade. Therefore we propose the pole frequency as an indicator of the performance of CML buffers.

4 Eye opening voltage [V] stages (pole: 4.5GHz) 5 stages (pole: 2.6GHz) 4 stages 0. (pole: 8.9GHz) 2 stages (pole: 5.6GHz) Frequency [GHz] Fig. 7. Eye opening voltage and pole frequency. (X = 5) Pole frequency [GHz] 00 0 Parasitics of Tr Parasitics of wire are dominant are dominant Pull-up resistance [Ohm] 000 Fig. 8. Pull-up resistance and the pole frequency. IV. Low-Power Design of CML Buffer This section proposes a design technique to reduce the power dissipation of CML buffers without degrading the bandwidth. A. Power reduction by impedance-unmatched driver According to Section II, the tail current is determined by the pull-up resistance and the required voltage swing. Conventionally, the pull-up resistance is tuned to achieve impedance matching because impedance-matched drivers suppress the reflection of the electromagnetic wave and improve the signal integrity. However on on-chip transmission-lines, the attenuation is significant and the reflected wave attenuates rapidly. Therefore impedance mismatch does not cause serious problem in signal transmission and we can use impedance-unmatched drivers [6]. From Eq. (4), the product of the pull-up resistance R D and the tail current I tail have to be constant value because the output voltage swing V out is a given parameter. If we can increase the pull-up resistance R D, we can decrease the tail current I tail. On the other hand, tuning the pull-up resistance does not degrade the bandwidth of the CML driver. From Eq. (8), the pole frequency is expressed as the inversion of the product of the resistance R D and the capacitance. The load capacitance is composed by the gate capacitance, drain-backgate capacitance and parasitics as shown in Fig. 5. We can classify the components of the load capacitance as below; C Tr : Parasitic capacitance of the nmos transistor. This capacitance is proportional to the transistor size. C RD : Parasitic capacitance of the pull-up resistor. This capacitance is proportional to the resistance R D. C wire : Parasitic capacitance of metal wires. This capacitance is independent to the design parameters. In this paper, we assume that the pull-up resistance is realized by polysilicon. Thus the capacitance C RD is proportional to the resistance R D. To drive on-chip transmission-lines, the output impedance of the driver has to be comparable small to the characteristic impedance and the transistor size has to be large. Therefore in the CML driver for on-chip transmission-lines, the component C Tr is dominant and the load capacitance is approximately proportional to the size of the transistor. From Eq. (3), the transistor size is proportional to the tail current. As explained in Section II, the product of R D and I tail is determined by the output voltage swing. Thus the pole frequency can be rewritten as ω p = R D R D W R D I tail = const. (0) Therefore if the load capacitance is proportional to the transistor size, the pole frequency does not change by tuning the pull-up resistance R D. As a result, the impedance-unmatched driver can reduce the tail current without degrading the bandwidth. We experimentally verify the impedance-unmatched driver in the next section. B. Relationship between the pole frequency and the pull-up resistance If the load capacitance is proportional to the gate width, we can decrease the tail current by increasing the pull-up resistance. However the load capacitance includes the parasitic capacitance of wire and the pull-up resistance. The wire parasitics is not depends on the gate width. If we realize the pull-up resistor by polysilicon, the parasitic capacitance of the resistance is proportional to the resistance value. When the gate width is large and the gate capacitance and the drain-backgate capacitance is dominant, we can reduce the tail current without degrading the bandwidth of the CML buffer. As the size of switching transistor decreases, the parasitic capacitance of wires and the resistance becomes significant. In this case, the pole frequency shifts to the lower frequency as increasing the pull-up resistance. Figure 8 shows the relationship between the pull-up resistance and the pole frequency. The parasitic capacitances are extracted from a CML buffer designed in a 90nm CMOS process. When the resistance is small, the switching transistor is large and the gate capacitance and the drain-backgate capacitance is dominant. In this region, the pole frequency is almost the constant even the pull-up resistance is changes. As the resistance becomes larger, the parasitic capacitance of wire and that of the resistance becomes dominant. Thus the pole frequency becomes lower when the pull-up resistance is large. Therefore the proposed method is limited by the parasitic capacitance of wires and that of the resistance.

5 Random input (differential) u u N-2 cascade driver X=u N- Eye-diagram evaluation terminator eye opening voltage [V] N=4 Fig. 9. Experimental circuit. 0.2 RD=50 Ohm RD=00 Ohm V. Experimental Results In this section, we show some experimental results of the impedance-unmatched driver. First, we explain the simulation setup. Next, the bandwidth and the power dissipation of impedance-unmatched drivers are evaluated by circuit simulation. A. Simulation setup The experimental circuit is shown in Fig. 9. We evaluate the bandwidth of the cascaded driver. The CML buffer is scaled by a constant taper factor u and the number of stages is N. The ratio X is the ratio between the first stage and the last stage and is expressed by X = u N. The input of the driver is a random Non- Return-to-Zero pulse sequence. To evaluate the performance of the driver, we assume that the transmission-line is lossless and the length is zero. Thus the output of the driver is directly connected to the bridge termination. A 90nm CMOS process is assumed as the fabrication process and the supply voltage is.0v. We evaluate the eye-diagram at the output of the driver. B. Comparison between impedance-matched and impedanceunmatched driver We fix the ratio X to 9 and vary the number of stages N. We assume that the differential characteristic impedance of the transmission-line is 00Ω. Therefore the resistance of the terminator is 00Ω. If we tune the driver to achieve impedance matching, the pull-up resistance of the final stage is 50Ω. We set the pull-up resistance 00Ω. Figure 0 shows the simulation result. The x-axis is the frequency of the input pulse and the y-axis is the eye-opening voltage. The number of stages is 4 and 5. The taper factor u is 2.08 and.73 respectively. From Fig. 0, the eye-opening voltage is almost the same even if the pull-up resistance R D is changed. The comparison of the total tail current is shown in Table I. From Fig. 0 and Table I, impedance-unmatched driver reduces the total tail current by 25% without degrading the bandwidth. VI. Conclusion A design technique to reduce the power dissipation of CML driver is proposed. Conventionally, impedance-matched driver is a common practice of the driver design. We use a impedanceunmatched driver whose pull-up resistance is larger than that of impedance-matched driver. By using larger pull-up resistance, we can reduce the tail current of CML buffers. From the pole eye opening voltage [V] frequency [GHz] RD=50 Ohm RD=00 Ohm N= frequency [GHz] Fig. 0. Input frequency and the eye-opening (X = 9, N = 4, 5). TABLE I Total tail current of the driver. #stages matched-driver proposed (N) (R D = 50Ω) (R D = 00Ω) ratio 4 3.9mA 0.4mA mA 2.4mA mA 4.4mA mA 8.5mA : (ratio)=(proposed)/(matched-driver) frequency analysis, our method can reduce the power dissipation without degrading the bandwidth. Experimental results show that the impedance-unmatched driver reduce the tail current by 25% and its bandwidth is almost the same as that of the conventional impedance-matched driver. References [] D. Pham, S.Asano, M. Bolliger, M. Day, H. Hofstee, C. Johns, J. Kahle, A. Kameyama, J. Keaty, Y. Masubuchi, M. Riley, D. Shippy, D. Stasiak, M.Wang, J. Warnock, S. Weitzel, D. Wendel, T. Yamazaki, and K. Yazawa, The Design and Implementation of a First-Generation CELL processor, IEEE International Solid-State Circuits Conference Digest of Technical Papers, Feb [2] Tao Lin and Lawrence T. Pileggi, Throughput-Driven IC Communication Fabric Synthesis, Proc. ICCAD, pp , 2002.

6 [3] Naresh Shanbhag, Pushing the Limits of Interconnect Performance: A Communication-Centric Approach, Tutorial of IEEE Workshop on Signal Propagation on Interconnects, May [4] Anup P. Jose, George Patounakis, and K. L. Shepard, Near-Speed-of-Light On-Chip Interconnects Using Pulsed Current-Mode Signaling, Symposium on VLSI Circuits Digest of Technical Papers, pp.08, Jun [5] Pingshan Wang, Gen Pei, and Edwin Chih-Chuan Kan, Pulsed Wave Interconnect, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol.2, no.5, pp , May [4] Massimo Alioto and Gaetano Palumbo, Design Techniques for Low-Power Cascaded CML Gates, Proceedings of IEEE International Symposium on Circuits and Systems, pp , May [5] Behzad Razavi, Design of Analog CMOS Integrated Circuits, McGraw-Hill, Aug [6] Akira Tsuchiya, Masanori Hashimoto, and Hidetoshi Onodera, Driver Sizing for High-Performance Interconnects Considering Transmission-Line Effects, Proceedings of SASIMI200, pp , Oct 200. [6] Masanori Hashimoto, Akira Tsuchiya, and Hidetoshi Onodera, On-chip Global Signaling by Wave Pipelining, Proceedings of IEEE Topical Meeting on Electrical Performance of Electronic Packaging, pp.3 34, Oct [7] Muhammad Khellah, James Tschanz, Yibin Ye, Siva Narendra, and Vivek De, Static Pulsed Bus for On-Chip Interconnects, Symposium on VLSI Circuits Digest of Technical Papers, pp.78 79, [8] Akira Tsuchiya, Yuuya Gotoh, Masanori Hashimoto, and Hidetoshi Onodera, Performance Limitation of On-chip Global Interconnects for High-speed Signaling, proceedings of IEEE Custom Integrated Circuits Conference, pp , Oct [9] Masayuki Mizuno, Masakazu Yamashina, Koichiro Furuta, Hiroyuki Igura, Hitoshi Abiko, Kazuhiro Okabe, Atsuki Ono, and Hachiro Yamada, A GHz MOS Adaptive Pipeline Technique Using MOS Current-Mode Logic, IEEE Journal of Solid-State Circuits, vol.3, no.6, pp , Jun 996. [0] Hiok-Tiaq Ng and David J. Allstot, CMOS Current Steering Logic for Low-Voltage Mixed-Signal Integrated Circuits, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol.5, no.3, pp , Sept 997. [] Akira Tanabe, Masato Umetani, Ikuo Fujiwara, Takayuki Ogura, Kotaro Kataoka, Masao Okihara, Hiroshi Sakuraba, Tetsuo Endoh, and Fujio Masuoka, 0.8-µm CMOS 0- Gb/s Multiplexer/Demultiplexer ICs Using Current Mode Logic with Tolerance to Threshold Voltage Fluctuation, IEEE Journal of Solid-State Circuits, vol.36, no.6, pp , Jun 200. [2] Mohamad W. Allam and Mohamed I. Elmasry, Dynamic Current Mode Logic (DyCML): A New Low-Power High- Performance Logic Style, IEEE Journal of Solid-State Circuits, vol.36, no.3, pp , Mar 200. [3] Payam Heydari and Ravindran Mohanavelu, Design of Ultrahigh-Speed Low-Voltage CMOS CML Buffers and Latches, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol.2, no.0, pp , Oct 2004.