For the first time, researchers have provided actual crystallization and melting temperatures from inside the memory cell.
For the first time, researchers have provided actual crystallization and melting temperatures from inside the memory cell.
If (at present, still a big if in some quarters) phase change memory (PCM) has a commercial future, understanding and quantifying crystal growth rates in the context of actual memory cell structures is imperative. In a new and admirable paper (subscription required), a team from IBM Zurich has added something new and absolutely essential to our understanding of crystal growth rates in PCM structures.
The paper acknowledges that, 50 years after the PCM effect was discovered (with many unfulfilled promises and false dawns along the way), a detailed knowledge of the central property, the phase change mechanism, is still lacking. That lack of knowledge is the equivalent of trying to design and make conventional silicon-based devices without the knowledge of carrier mobility or the diffusion coefficients of dopants.
Finding the device operating conditions that provide maximum crystal growth rate and minimize the write (SET) time and energy is especially important. This is because SET is the slowest of the two write operations and has the greatest impact on the overall write/erase (SET/RESET) time budget.
At what might be considered the other end of the scale, for data retention there is the requirement to minimize crystal growth rates, especially at elevated temperatures. In that respect, most PCM structures have a built-in massive nucleating site in the form of an electrode of crystallized active material. This removes the need to waste write time waiting for nucleation to occur before crystallization. However, it cannot be ignored as a potential site for elevated temperature device failures. Some might argue that, for data retention, the outliers that cause the failures are not representative of normal device operation and are defect related.
Then there is a third item that might impact crystal growth rates: electro-migration. It may increase crystal growth rates by moving the required crystallizing material to the growing crystal interface faster than is possible by normal means. We raised this question with the IBM team, and they commented, "Our studies did not show sufficient evidence for any electrically induced crystal growth. The data could be explained based on standard thermally induced crystallization alone."
The IBM researchers used a PCM cell as a nano-sized laboratory and employed some very innovative techniques. They set about the task of providing the fundamental data that relates to crystal growth rates in an amorphous glass and in the PCM material as a super-cooled liquid, as well as the temperatures for maximum crystal growth rate and melting. It is surprising that the latter two temperatures inside a device were not really known until now, nor was there any way to obtain representative values./p>
IBM used the familiar PCM dome structure. From a small TiN bottom electrode (BE) contact, sometimes formed of a heater material, a dome of amorphous material is formed in crystallized active material that acts as the other electrode. The active material for this series of experiments was Ge2Sb2Te5 doped for extended write/erase endurance. The technology node was 90 nm. The memory cell structure in cross section is shown in Figure 1; the diameter of the BE was 50 nm. In an ideal world, during SET, or writing the memory cell to its low-resistance state, a crystal front grows from the hemispherical crystal surface of the TE and parallel to it up the temperature gradient toward the BE.
The really innovative part of this experimental work is the use of the cell as a nanolaboratory. The results obtained should be directly applicable to real memory cell structure design, without the doubts that are always raised about results from films and bulk material, where measurements are made outside the cell structure.
For crystal growth rates or growth velocity in the glassy state, the technique was to reset the device to form the same sized amorphous dome as shown in Figure 1. The dome's radius was 46 nm. Researchers then measured the change in threshold voltage as a function of time at a series of increasing ambient temperatures from 180°C to 270°C. The growing crystal interface reduces the threshold voltage of the material remaining in the amorphous state. This variable was used as the measure of the dome radius. The graphical inset in Figure 1 illustrates the initial form of the data obtained. Using this data, the activation energy for the growth rate over six decades of velocity as a function of temperature (1/K) was obtained. From what was clearly an Arrhenius relationship, an activation energy for the growth velocity of 3.01 eV was derived.
Those familiar with PCM devices will recognize that two important variables must be accounted for in such an experiment: the variation of threshold voltage with temperature and voltage drift caused by the relaxation of the amorphous structure. Equations describing those two variables are now well established and were used to provide an accurate determination of the growth velocity. In addition, because only a small amount of growth occurred, planar surface approximations could be used.
The melting temperature within the PCM cell was obtained by measuring the two-terminal resistance as a function of the applied input power and noting the input power when there was evidence, in the form of a discontinuity in resistance, that an amorphous plug or initiating molten hotspot had appeared above the BE in the crystallized material. Examples of the form of the curves for two temperatures from the range evaluated are illustrated in Figure 2 in blue and red.
This experiment was repeated over the temperature range of 100K to 400K. A plot of ambient temperature as a function of input power resulted in a straight line, the extrapolation of which to zero input power must be the melting temperature. Perhaps it takes a little time to appreciate the abstract idea of melting with zero input power. It is possible to obtain an estimate of the thermal resistance, and the concept becomes easier to understand with the help of the equation:
Tm = Rtherm (Pmelt)+ Tamb
When the ambient temperature Tamb equals Tm, the power in becomes zero (where Pmelt is the power required for the amorphous plug to appear, and Rtherm is the all-path thermal resistance from the top of the BE). The important piece of information obtained is the value of 877K (or 604°C) for the in-situ cell melting temperature Tm. The literature gives a range of values from 600°C to 615°C for the melting point of Ge2Sb2Te5 obtained by a number of different methods. The doping material used by IBM is not known, so direct comparison is not possible, but the values are close.
Maximum growth rate and temperature
From current knowledge, at low temperatures, crystal growth rate increases with increasing temperature, and crystal growth rate at the melting temperature must be zero. It would therefore be expected that there would be a maximum value somewhere between those two extremes.
For their next ingenious experiment, the IBM researchers set themselves the task of finding the temperature for maximum crystal growth rate, designated as Tcryst, and the actual growth rate within a cell structure.
For this measurement, a stepped pulse of two parts was used (see Figure 3). The first part of the pulse had amplitude sufficient to melt or reset the crystallized material. The pulse was then stepped down to a crystallizing pulse to partially set the device. The amplitude of the second part of the pulse provided a means of controlling the temperature of the amorphous-crystal interface at values close to the melting temperature, where some crystal growth would occur at the interface. By varying the power of the crystallizing pulse, it is possible to obtain an interface temperature that results in maximum crystal growth rate at temperatures where the memory material is in its super-cooled liquid state. Resistance was used as the measure of crystal growth, with the smallest resistance value indicating the highest crystal growth rate. A representation of a typical curve of resistance as a function of step power for one temperature is shown in the graphical inset of Figure 3; the minimum in the curve indicates the step power for maximum crystal growth rate.
The next step was to obtain the minimum input power for Pcryst over an ambient temperature range of 150K to 400K. A graphical plot of ambient temperature as a function of power for maximum growth rate results in a straight line, and its extrapolated intercept with the zero power axis provides the crystallization temperature. The result was a crystallization temperature for maximum growth rate Tcryst that averaged to 750K or 477°C (see Figure 4).
The final step was to find the overall crystal growth rate as a function of temperature across the complete range of temperatures. This involved an extension of the experiment to find Pcryst. The goal was to find vg, the growth velocity as a function of temperature (T). This required an accurate estimate of ua, the radius of the dome, which was obtained by measuring current voltage characteristics and the resistance after each stepped pulse and then calculating the values using known transport equations. The calculations are complex and are detailed in the paper and associated supplementary information.
The result is a maximum growth rate at Tcryst of 0.55 meter/s at 750K. When the data from all four experiments is combined, the result is a curve (see Figure 4) showing growth velocity as a function of temperature. Importantly, this provides growth velocity values well beyond any measured before, with the added value that the data is representative of what is happening in the PCM cell. For crystallization, the curve in Figure 3 can be divided into three distinct parts: crystallization of the active material in the glassy state, its time-dependent glassy condition, and its post-glass condition as a super-cooled liquid.
The IBM researchers can legitimately claim they have broken a few records and have extended crystal growth rate measurements at temperatures and in material states (super-cooled liquid) for PCM materials well beyond a level ever measured before. For the first time, they have provided actual crystallization and melting temperatures from inside the memory cell. They also provided a means to measure key PCM memory cell design parameters, such as temperature for maximum crystal growth rate.
For example, it is now possible to use the maximum crystal growth rate at 750K of 0.55 m/s for Ge2Sb2Te5 to predict the minimum write (SET) time for a memory device with a planar inter-electrode spacing of 10 nm to be 18.18 ns, providing that is by set pulse shaping the growing crystal-amorphous interface could be maintained at 750K.
In an email interview, we raised a question regarding the validity of this calculation with the IBM team. The response was, "Yes, that would be correct if we could ensure that the temperature equals Tcryst at the interface. However, please note that this is slower than undoped Ge2Sb2Te5. Doping is needed to achieve good endurance at the expense of slightly slower SET speed."
Moving forward, it would be wise if all promising PCM materials were subject to this new measurement regime, especially when composition changes are made to enhance other PCM characteristics.
This new work also suggests that it is now necessary to review some of the accelerated elevated temperature data retention measurements, and that great care must be taken in interpreting the results of such tests. (See PCM data retention and the impact of crystal electrodes (Part 1) and PCM data retention and the impact of crystal electrodes (Part 2).)
Following all their careful and meticulous work, the IBM authors chose to use the word "estimation" in describing their results. We asked why this was necessary.
We have tried our best to perform all the experiments with utmost care and to perform a thorough error analysis. Hence, the growth velocities in the glass, the melting temperature, and the temperature of maximum growth can very well be called measurements. Given the error bars due to extrapolation and the wide range of growth velocities over the whole temperature range, the fit of the overall growth velocity curve is rather sensitive to the parameters used which also depend on the assumptions we had to make, thus we prefer the term "estimation" to indicate that the curve is more representative of a band (albeit very small) of possible curves best estimated by our fit.
In their paper, the IBM authors offered a final sobering comment on PCM and the associated switching materials. They observed, though the high speed of crystallization is a blessing, it also leads to the biggest curse of the technology: the structural relaxation toward the equilibrium that was observed in this new work and the literature as drift in growth rate, threshold voltage, and resistance.