BiST Vs. In-Circuit Sensors

Hybrid solutions emerging as reliability concerns increase and coverage becomes more difficult.


Monitoring the health of a chip post-manufacturing, including how it is aging and performing over time, is becoming much more important as ICs make their way into safety-critical applications such as the central brain in automobiles.

Faced with longer lifespans and a growing body of functional safety rules, systems vendors need to be able to predict when a part will fail. But as sensing automotive IC failures becomes codified into standards, the newness of everything is hitting all at once — advanced-node designs, AI for predictive maintenance, zero defect tolerance and processing all the data needed to diagnose ICs’ health.

There are several key approaches being used for this:

  • Built-in self-test (BiST), a technology that has been around for a couple decades. While this technology has a proven track record, but it also takes up valuable space on a die.
  • In-circuit sensors, which take up less room on die and can monitor the chip non-stop during operation.
  • Increased testing, inspection and metrology to identify potential defects before chips are released into the market.

Each of these approaches has strong points and weaknesses, and generally all three are required to ensure quality over time.

BiST basics, non-stop alternatives
BiST comes in two key flavors — logic BiST (LBiST) and memory BiST (MBiST), which has a repair feature that LBiST doesn’t have. Both are integrated into the die.

BiST works by generating pseudo-random test patterns. It sends those patterns along scan chains to activate a response on the chip, comparing results of the tests to ideal behavior. Signals may be compressed to speed up the the test.

BiST usually activates in power on and power off. “If you’re running power-on self test when you started the vehicle initially, most of the vehicle manufacturers specify a time between key-on and when the vehicle has to be ready to run,” said Lee Harrison, automotive IC test solutions manager at Mentor, a Siemens Business. “So all the manufacturers are trying to squeeze as much testing in as possible during that very short time period.”

BiST is moving into in-system tests, as well, beyond the power-up/power-down approach that resembles an airplane pilot’s checklist. Mentor released an in-system observation scan BiST that can capture a test during every shift cycle. Likewise, Synopsys has its TestMAX XLBiST, which tolerates indeterminate digital states.

BiST is still not a non-stop test that runs continually. It runs at certain cycles. But sometimes it can spread out a test over time. Using an MBiST controller to squeeze tests in smaller chunks during operation is an option Mentor offers. “We have a slightly modified memory BiST controller, which can run memory BiST in-system without destroying the contents of the memory. It’s not as efficient as running a straightforward memory BiST, but we’re able to do over a longer period of time the same level of testing for that memory by breaking the test into very, very small parts, which has a minor impact on the overall functionality of the device. That’s memory BiST. Logic BiST is a lot more challenging, because the amount of data that’s involved.”

The idea of using in-chip sensors is to run continuous monitoring of the chip. A number of companies are offering real-time observations. Tiny sensors are being places on die to provide continuous data monitoring, with data management to go with it. PDF Solutions has a variety of in-die sensors relevant to manufacturing, test and quality. ProteanTecs also has on-chip monitoring with off-chip analytics that pulls more data from in-field, focusing on the physics of failure and providing insights. The company uses Agents, which follow Universal Chip Telemetry (UCT), and are embedded in the design of a chip. Companies such as Moortec have in-chip monitoring for cutting edge AI and machine learning chips.

Saving space on the die
BiST’s reputation as a real-estate hog on the die is changing. “BiST is not getting any smaller, but it’s not in the regions of 70%,” said Mentor’s Harrison. “For memory BiST, the technology we have enables you to utilize single memory BiST controller for many, many memories.”

Taking up less space on-die has motivated in-circuit monitors. Tiny sensors deployed in the empty area on a chip is a tactic that saves die space. “Our design-for-inspection system deploys tiny sensors, called DFI Filler Cells, into the empty/dummy areas of the chip (i.e., zero area penalty) for inline contactless measurement as soon as the first metal layer has been processed,” said Dennis Ciplickas, vice president of advanced solutions at PDF Solutions.

Reducing the amount of logic also helps a test live in a smaller space. “We do away with a lot of the logic of the BiST,” said Noam Brousard, vice president of product at proteanTecs. “We’ve shown already in real-life use cases that we add anywhere between 1%, 2%, 3% of extra area or so. It’s a very, very small addition to the chip.”

Gathering real-life data, in-system
BiST isn’t standing still, though. Test sequences are running faster these days. Mentor released its observation-scan, which has sped up the test process. But the real change may be where the data is processed.

“In the past that data has been very, very localized to the IC,” said Harrison. “But what we’re seeing is all vehicle manufacturers are collecting and using that data, not just from, from single vehicles, but from vehicles all over the world to help build a bigger picture of the kind of health of the system. You see that with vehicle manufacturers such as Tesla. They’re able to roll out updates that will improve the performance of the vehicle, it could improve the life cycle of the vehicle and those type of things. Buyers are learning from the data that they’re collecting that they may be able to make adjustments to the overall system, which extends the lifetime of the vehicle.”

That’s a good starting point for the next phase, which is a much more granular dive into what is happening during operation.

“Most recently we are also deploying electrically active sensors in the product die with our CV Core System,” said PDF Solutions’ Ciplickas. “These CV Core IP blocks characterize electrical and material characteristics related to product performance and reliability, and provide response data both during manufacturing, assembly and test as well as during system operation for in-life monitoring. In contrast to embedded circuit monitors, which are designed detect performance drift or anomalous behavior, our sensors give insight into the physics of failure. Simply detecting that a chip has failed doesn’t give insight as to why it failed, and data from bitmapping, EFA/PFA, RMA’s and 8D processes is precious but extremely limited. Similarly, knowledge that circuit performance is drifting on one chip is difficult to generalize into predictive action beyond that chip. On-chip sensors that track systematic trends in structural and electrical characteristics give into the physics of the trend and our experience from manufacturing and test is that models based on physics are have more predictive power.”

But how to handle that amount of data still needs to be worked out. “At the moment, that’s the part that’s very much in its infancy, because till now the testing has been very localized,” said Mentor’s Harrison. “One of the things Mentor is exploring, now as part of Siemens, is that Siemens has some very neat software aimed toward product lifecycle management. It provides a great mechanism to collect that data. And you can build dashboards on top of that to display and monitor key pieces of data, but as yet there’s no kind of globalized standard on how we collect that data and use it. Everything is still in the early stages, and vehicle manufacturers are coming up with their own solutions and ways of doing things. And then you’ve got the risk there. We’ve been talking to a number of automotive OEMs, which are very concerned that when they collect that system data they don’t also collect the personal data that’s associated with individual drivers and their own personal settings. You need to make sure from a security aspect that we’re just collecting vehicle data and it’s not associated to any kind of driver-specific personal data. The OEMs are being very, very cautious about that at the moment.”

AI, yea or nay
Using AI to sort through data is a big part of using embedded in-chip monitoring.

“We pull all of this data together, including available manufacturing and test data,” said PDF Solution’s Ciplickas. “Early readout from our DFI (design for inspection) system enables fast learning cycles in the fab. The design of experiments in the DFI filler cells helps to rapidly zero in on issues and root causes. CV Core sensors monitor and diagnose changes in chip behavior from on-wafer to in-package to in-system, and finally in-field. We use machine learning across this data set to find subtle relationships that let us apply AI to modulate test flows to optimize the cost of quality, detect trouble directly at the tool level during manufacturing and assembly, and disposition lots, wafers and die before shipment to avoid quality escapes into the field.”

AI has proven less useful for BiST. Harrison said Mentor experimented with it from BiST data, but found it didn’t improve BiST enough to have a use case.

Test effectiveness
BiST is like automatic test pattern generation, except that ATPG is a high-quality test that happens on a big piece of test equipment with a lot of memory during post-silicon test. “What we do in-system with the logic test and the logic BiST is very similar,” said Harrison. “You still use the same scan structure, so that’s the same testing structures inside the design. But instead of streaming the pattern data from a piece of test equipment, we have a small piece of logic in-system that generates those patterns. It’s not going to be as efficient and as detailed as the ATPG because with ATPG, we can create the patterns exactly to detect specific faults so we can be very targeted. With logic BiST we rely on the randomness of the patterns to protect as many faults as we possibly can.”

That randomness can get test coverage to the low-90% range versus ATPG’s 99%. That’s still a high-quality level for in-system tests. But rounding it out with sensor and monitoring systems produces even higher coverage.

“It addresses the concerns for most of our customers in terms of getting to where they need to be for their automotive certification,” said Harrison. “You want BiST because you don’t just want to look at what’s operating now.

Related Content

Built-in self-test (BiST)

Automatic Test Pattern Generation (ATPG)

Sensing Automotive IC Failures

BiST Grows Up In Automotive

Scan Diagnosis

Leave a Reply

(Note: This name will be displayed publicly)