edfas.org 35 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 26 NO. 2 interest. This is doable but is complex and not easily achieved without specialized models, tools, and know how. When probing advanced nodes, precise CAD-tostage alignment within digital logic areas is essential. Techniques include mapping clock cells using laser voltage imaging (LVI) to known CAD clock buffer locations[5] or by implementing alignment markers into some of the front end mask sets, explicitly added into the design to help the FA teams during silicon bring up. However, even when the CAD-to-stage alignment appears to be perfect, small changes in overall system drift can result in waveforms collected from a single target appear to be significantly different over multiple acquisition attempts. Therefore, as technology shrinks it can be increasingly difficult to successfully acquire pass versus fail data for comparison. An example of probing the same target on a 16 nm device multiple times that produces five different waveforms is shown in Fig. 4. A simple explanation for this is that by being only one pixel off of the target, the waveforms can vary dramatically, making repeatability extremely difficult, especially on smaller technology nodes. So, the questions in the mind of the analyst become: Am I really on the target I think I am? Can I trust this data if I don’t have a gate level simulation? How do I interpret the waveform when it has this much crosstalk? And if I am on the correct target, are the results showing that the waveform has corruption or am I just picking up crosstalk from another device positioned nearby? Differential LVP can help answer some of those questions. For example, probing pass and fail behavior simultaneously automatically guarantees equivalent good vs. bad probe target placement from the same location under the same system conditions. It removes any effects of temperature and stage drift since both data sets are collected simultaneously. Another bonus is that by subtracting a fail waveform from the accompanying pass waveform, all of the crosstalk is removed and there’s only the differences between the two. Essentially the subtracted fail from pass data (differential) reveals any pass vs fail mismatches. It also reveals when the difference between the two occurred within the test. Suddenly, there’s a new tool in the toolbox that allows to probe logic on soft failures and look for differences beyond the limitations of the laser spot size or mechanical stability of the probe system. Visible light probe at 785 nm can also be used as a way to make incremental but not breakthrough improvements which help reduce some waveform crosstalk challenges but do not eliminate the problem. Also, illumination wavelengths below 1064 nm require that samples are backside thinned to less than 5 μm. In many cases this can make the benefits of reduced crosstalk too risky versus potentially damaging one-of-a-kind qualification or customer returned samples. Damage can occur during sample preparation or from ultra-thin samples overheating during test. One thing to consider is that any techniques developed along the lines of differential LVP at 1064 nm, will also scale into the visible spectra, and could even be applied one day to re-emerging averaging based techniques such as e-beam probe. HOW DIFFERENTIAL LVP IS PERFORMED In a nutshell, it is necessary to modify an existing LVP tool to stream live LVP data from a target device and bin the data into separate channels, based on whether the device passed or failed the executed test loop. Proof-of-concept Fig. 4 When attempting to probe the same 16 nm-based target five times reveals five different outcomes based on drifting one pixel in X, Y, or Z.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==