Manufacturing Effects on Data Retention of Non-Volatile Memory ...

<strong>Manufacturing</strong> <strong>Effects</strong> onData Retention of Non-VolatileMemory DevicesDoes programming a Flash memory device before wave soldering orvapor phase reflow compromise the data retention? Can the heatcause premature data loss?IntroductionAt Data I/O Corporation, we manufacture a variety of high-speed programming equipmentthat is used in SMT manufacturing lines to program Flash memory devices and Flash-basedmicrocontrollers. Every now and then, a process engineer or program manager will ask usabout programming before soldering. They know that charge is stored on floating gates, andthey remember enough physics from their undergraduate days to make the connectionbetween heat and loss of charge which can cause a “bit-flip.” But they don’t know if this is asignificant problem; they just want reassurance that there is no effect on the quality of theirproduct.This is particularly important for many of the automotive applications were Flash-basedmicrocontrollers are in “mission critical” applications. The last thing an automotive companyor EMS vendor wants is a recall due to problems with data retention in one of the controlmodules.In some cases, we’ve seen engineers who are reluctant to use our equipment, even though thefinancial return is obvious. They continue to use slow and costly boundary scan or other In-System Programming (ISP) methods just because it allows them to program after soldering.It is not enough for us to just tell them it is not a problem; we have to prove it to them. Oncethey have “peace of mind” they are very happy to improve the efficiency of their SMT line byswitching to our high-speed gang-programming equipment. For example, our RoadRunnerprogrammer mounts onto the SMT line at the pick-n-place station (before soldering).So, let’s review why it is not a problem to program Non-Volatile Memory (NVM) devicesbefore exposing them to the heat of a typical SMT line.Reliability Physics 101Every semiconductor manufacturer that produces NVM has to verify that data retention for anew design is sufficient as well as establish a quality procedure for monitoring data retentionin production. Typically they will test for loss of data at elevated temperatures; then, using arelationship called the “T-Model,” they will calculate the equivalent lifetime for “normal”temperature ranges.Arrhenius rev B.doc 7/23/2004 Page 1 of 3

The T-Model is used because it takes into account both temperature and voltage effects. Thisis appropriate because the semiconductor manufacturer is trying to determine the dataretention for NVM used in an application where heat and power are applied at the same time.(An automotive application is a good example where a NVM device is used at elevatedtemperatures.)The T-Model is defined as follows:⎡Tt−TAf = exp ⎢⎣ Tou⎤⎥⎦where A f is the acceleration factor, T u is the temperature of the “normal” environment, and T tis the elevated temperature where failures are induced at an accelerated rate. (Note both ofthese temperatures are in degrees Kelvin.) To is the characteristic temperature for dataretention that embodies dielectric, field strength, and charge loss effects. For EPROMs usingan oxide-nitride-oxide process, one manufacturer empirically determined To to be 21°K forthe floating gate.To continue the example, let's say we are trying to determine the data retention for anautomotive application where a microcontroller (with embedded EPROM) sees a junctiontemperature of 125°C. For one of the tests, we might measure data retention for over 450hours at 250°C. (Obviously, this is at wafer burn-in; the plastic encapsulent use in SMTpackages cannot withstand this temperature.) Using the T-Model, we would calculate theacceleration factor as follows:Af⎡523− 398⎤= exp⎢= 384⎣ 21 ⎥⎦An acceleration factor of 384 means that if it lasted for 450 hours at 250°C, it will last 384times longer at 125°C, or about 20 years. This is certainly sufficient for any automotiveapplication.But when we think about our perennial question “Does the heat from reflow hurt the data?”,there is no voltage applied to the embedded Flash; the device is “off.” So how do we modelthis?Well, if temperature effects alone are what we are trying to model, then we can use thefamous Arrhenius relationship. This model describes the effect that temperature has on achemical process with a certain “activation energy.” Like the T-Model, it produces anacceleration factor for a system exposed to an elevated temperature. The relation is asfollows:Af⎡ E= exp ⎢⎣ ka⎛ 1⎜⎝ Tu1 ⎞⎤−⎟⎥Tt⎠⎦where A f is the acceleration factor, E a is the activation energy (in eV), k is Boltzman’sconstant (in eV), T u is the temperature of the “normal” environment, and T t is again theelevated temperature. (Again, these temperatures are in degrees Kelvin.)But in order to use the Arrhenius relationship, we need to empirically determine the“activation energy” by fitting a curve to some experimental results at different temperatures.The activation energy is a critical parameter for getting meaningful results. A typicalactivation energy is 0.8 eV for effects related to charge loss when no voltage is applied to thedevice.So, for another example, say there is a guided munitions application where an electronicmodule might sit in storage in the desert. Assuming the internal temperature of the shellreaches 125°C, the customer wants to know if this will compromise the data retention in theArrhenius rev B.doc 7/23/2004 Page 2 of 3

embedded Flash. In this case, we would use data from a test where we measured dataretention at wafer burn-in with no power applied to the circuit (other than brief intervals toread the data and make sure it was still OK). So let’s assume we measured good retention forover 1000 hours at 250°C. (This is typical of today’s NVM.) Using the Arrheniusrelationship, we would calculate the acceleration factor as follows:A f⎡ 0.8 ⎛ 1 1 ⎞⎤exp⎢= 2635⎜ − ⎟⎥⎣8.6171x 10 ⎝ 398 523⎠⎦=−So, with an acceleration factor of 263, the data retention in the application (at 125°C) will bemore than 263 x 1000 = 263,000 hours or over 30 years. (I’m sure the US Army would behappy with this result.)This same process can be applied to data retention effects from heat exposure duringsoldering. Peak temperatures of 235°C are typical, but the durations are much shorter. Let’suse the Arrhenius relationship to calculate how much of the “life” we are using up by heatinga NVM device after programming.SMT Reflow ExampleSay you are using vapor phase reflow with a peak temperature of 235°C at a duration of 60seconds. The thermal profile also includes ramping up to 125°C for a few minutes, but wehave already shown that there is negligible effect at this temperature.The worse case assumption would be that the thermal mass of the NVM device is so smallthat the die temperature is tracking the ambient temperature. Using the same data point as inthe previous example, we can calculate the acceleration factor between non-powered dataretention tests and the SMT vapor phase reflow temperature as follows:A f⎡ 0.8 ⎛ 1 1 ⎞⎤exp⎢= 1.695⎜ − ⎟⎥⎣8.6171x 10 ⎝ 508 523 ⎠⎦=−So now we are saying that there is very little difference between charge loss at 250°Ccompared to 235°C; only a factor of about 1.7. So if we spend 1 minute at 235°C, what havewe done to the average data retention time? To find out, just divide 1 minute by 1000 hourstimes 1.7. This turns out to be about 10 ppm or 0.001% of the calculated life.So, referring back to the guided munitions application, the effect of programming before SMTreflow reduces the data retention lifetime from 30 years to 29.9997 years. So who cares?On most data sheets for SMT devices, there is a list of Absolute Maximum Ratings. Includedin these parameters is usually the maximum heat exposure this part can withstand duringsoldering and still be reliable. But this maximum rating is mostly due to the thermomechanicaleffects. For example, at extreme temperatures the encapsalent will flow and tearoff wire bonds. But as we have shown, data retention is not a problem for this brief thermalexposure.Note that for a thorough estimate of reliability of a NVM device, we would also have toinclude thermo-mechanical effects and temperature-humidity effects.ConclusionWe have greatly simplified the reliability physics for NVM devices by just considering chargeleakage at elevated temperatures in an unpowered device. But this is the appropriate modelbased on the question we are always asked. And now we have the answer – there is negligibleeffect on data retention by soldering NVM devices after they have been programmed.Arrhenius rev B.doc 7/23/2004 Page 3 of 3