Analysis and Design of High Performance and Low Power



Analysis and Design of High Performance and Low Power

Current Mode Logic CMOS

Phillip Chin, Junjie Su, and Xiaolan Zhong

University of California, Berkeley

Abstract

With the scaling down of CMOS transistors, many issues, once considered negligible, now have become a factor in design. Some of these problems are leakage current and power consumption. The solution this paper will address is using current mode logic as opposed to traditional voltages. In the early stages of this research, a number of gates have been implemented and are compared against their static CMOS counterparts. We also begin by analyzing 90nm MOSFETs and will continue to work down to lower channel lengths.

I. Introduction

The scaling down of CMOS transistors has posed new problems for designers. As the limits of size are being reached, new logic styles will be needed to continue the trend. Current mode logic (CML) was researched in the past, but may offer a solution for today’s technology. CML implements some analog components to compute logic. Typically, CML refers to a logic style where voltages values still determine the logic values of the inputs and outputs, but current is only used as an intermediate variable. The basic structure consists of differential pairs with dual outputs. The inputs steer the current down one branch or the other to compute logic [3].

This paper focuses on “true” CML where the input and output values are actually currents (throughout this paper, CML will refer to true CML). The basic method of computation is current mirrors, which pass values by mirroring currents through other transistors. CML offers a number of advantages such as a reduced number of transistors and a smaller area [1]. A one bit adder cell has been built with as little as 15 transistors using current mode logic [2]. Also, given a supply voltage, the performance of a CML gate can be optimized for delay, power, and the switching noise. Previously, with static CMOS, the only real parameter that could be changed was the size. With a control over the current, current as a parameter can be used to increase performance [1].

However, there are some drawbacks that will have to be overcome. The supply current of a CML gate is independent operating frequency; there will be static power dissipation. However, at higher frequencies less power is dissipated than the static CMOS counterparts [5]. This will lead a reduction in voltage supply and current. However, scaling down these parameters will lead to a greater impact of leakage current. In particular if the nanoAmpere level is the target, it will be about the same magnitude of the leakage current. Thus, the circuit become less robust and is more prone to errors.

II. CML Logic Gates

As stated before the inputs and outputs to CML gates are currents. A number of different implementations have been explored; depending on the gate, logic “1” could be presence of current with logic “0” being absence of current or visa versa.

A. Inverter

The CML Inverter is made using an NMOS current mirror, using a PMOS current source connected to the diode-connected NMOS. For this structure, the presence of current is logic “1”. When Iin is high, the current from the PMOS transistor will flow through that branch, leaving no current to go through the mirror. This will effectively turn off the mirror network and Iout will be low. However, depending on the current going through the PMOS, the leakage current of the mirror network, when it is turned off, can become a factor and produce errors in computation [1].

[pic]

Fig. 1. CML Inverter

B. NAND Gate

The inverter presented above can be modified slightly to produce a CML NAND gate. At the input node in the inverter, multiple inputs can be attached to that node. To account for this the PMOS must be resized such that the amount of current is multiplied by the number of inputs. So, if all the inputs are high, the current leaving will be the number of inputs times the original current (i.e. the amount supplied by the PMOS). Like the inverter the output will be low. However, if any of the inputs are low, there will be enough current to flow through the mirror such that the output NMOS is turned on to produce an output current. As long as there is a presence of current, the output is considered to be high [1].

[pic]

Fig. 2. CML NAND gate created from Inverter

A second NAND gate implementation is shown in Figure 3. The current sources in the bottom left represent the inputs to the NAND gate. A key feature to this design, is that M5, a PMOS (bad pull down device), is connected ground, required a threshold drop, so it can effectively be turned off when it needs to be. The key component that makes this work is that when all the inputs are high, all the current supplied by the upper left current supply and M6’s current will be drawn through the inputs. This will effectively turn off M5. The PMOS on the output branch must be sized to twice M6, because the maximum current flowing through M6 is 0.5I. Therefore the current at the outputs will equal to high and be drawn through the current supply at the output branch making Iout low. If any of the inputs are turned off, M5 will have to be on and M6 will be off. This will turn off the PMOS at the output branch, forcing the current to be drawn from Iout [4].

[pic]

Fig. 3. A second implementation of a CML NAND Gate

C. NOR Gate

The CML NOR gate is essentially the reverse of the second CML NAND Gate as can be seen in Figure 4. The upper left current supply supplies the opposite of the NAND gate. Like the NAND gate, an NMOS is connected to the supply voltage, requiring a threshold voltage drop across it, allowing it to be turned off when needed. When any of the inputs are high, M3’s current will rise to supply the current demanded by the inputs. This will keep M1 off and thus turning off the NMOS at the output branch. However, when all the inputs are off, the 0.5I will flow through M1 making Iout high [4].

[pic]

Fig. 4. CML NOR Gate

III. Performance Comparison

As a preliminary test, the static CMOS and CML NOR gates were compared in performance. As was noted earlier, for higher frequencies, the CML actually performs better [5]. The Static CMOS suffers greatly, which would suggest for high frequency applications, CML would be the best choice.

[pic]

Fig. 5. Delay-Power Product Comparison between Static CMOS and CML NOR gate.

As mentioned earlier the robustness of CML gates are still in question. This is due to the increasing influence of the leakage current. In the case of the NOR gate, the logic swing is not perfect. As can be seen from figure 6, the input switches at a rate of 5 GHz. The high value of current is 10 μA. There are a few noticeable spikes, but it swings from to the desired values. However, if less than 1 μA, the circuit becomes difficult to bias and the output is centered at a new current value.

Many difficulties were experienced when running these tests. In particular, tests were run on SPICE, but it seems not to cater well to current. For example, nodes were inaccessible for currents, where as voltage could be measured easily. Also, it was very difficult putting a load at the output of the gate. For voltage mode logic, a capacitor sufficed, because it held the voltage. Unfortunately, most common devices will not work for the current case. In our case we used a diode connected MOSFET, but it did not work perfectly and there was always current flowing through it, even at times when it should not have been. This made it difficult to implement the circuit, even though a partially working one was built.

[pic]

Fig. 6. CML NOR gate out with input switching at 5GHz.

An inverter chain was also built and tested for robustness. Due to the difficulties of SPICE, the currents from each stage could not be measured, but the voltages were measured. As the signal progressed through the inverter chain, the voltage signal degraded. This may suggest that CML circuits do not work well when put in a long chain. It also seems safe to assume that the current signal must have been affected as well, which would suggest poor robustness.

IV. Conclusions and Proposals

Based on our simulation, if CML is used in a high frequency, high performance task, it has an advantage over static CMOS. However, the robustness of the circuit still remains an issue. It is greatly affected by leakage current and bias voltage. In these cases static CMOS these problems for the most part do not exist. These problems also made the CML gate difficult to implement. From the tests that were run, it appears that CML circuits may perform poorly if put in long chain.

In these preliminary tests, we used ideal current sources whenever they were drawn that way in the figures. These will have to later be replaced by MOS current sources. We believe this will actually help, because MOS current sources themselves have leakage current, whereas ideal current sources do not. Biasing is another concern with the circuits. As lower voltage and currents are applied, the precision of the bias voltage becomes ever more important. It will clearly have a great effect as the smaller transistors are used.

In the future, we plan to run tests with the 60nm and 45nm models. The results in this regime are normally less predictable, but we hope that we can develop logic that can be properly computable. As mentioned above with leakage currents, we hope to possibly take advantage of leakage currents and use it to our benefit. We will combine the implemented building blocks into larger modules and chains and eventually into a larger structure such as an adder. It will be first made in layout, with the netlist extracted from it.

V. References

[1] Ismail Enis Ungan and Murat Askar, A Wired-AND Current-Mode Circuit Technique in CMOS for Low-Voltage, High-Speed and Mixed-Signal VLSIC, Analog Integrated Circuits and Signal Processing, pp. 59-70, November 18, 1996.

[2] K. Navi, A. Kazeminejad, and D. Etiemble, Performance of CMOS Current Mode Full Adders, Proceedings of The Twenty-fourth International Symposium on Multiple-Valued Logic, pp.27-34, May 27, 1994

[3] Jason M. Musicer, Jan Rabaey, MOS Current Mode Logic for Low Power, Low Noise CORDIC Computation in Mixed-Signal Environments, Proceedings of ISLPED, 2000, pp.102-107, July 26-27, 2000

[4] Jan Rabaey’s EE241 slides

[5] Vasanth Kakani, Delay Analysis and Optimal Biasing for High Speed Low Power Current Mode Logic Circuits, 2004 IEEE International Symposium on Circuits and Systems. IEEE Part Vol. 2, pp. II-869-872, May 23-26, 2004.

-----------------------

Delay*Power vs. Operating Frequency of NOR Gate

0

200

400

600

800

1000

1200

1400

1600

1800

2000

0

1

2

3

4

Operating Frequency (GHZ)

Delay*Power (ps)

Static CMOS

CML

................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download