Plugging Gaps In The IC Supply Chain

Consistent unique identifiers, from die to final end-system, open the door for new analytic workflows that can improve reliability and traceability.


Multiple touch points in manufacturing and packaging are exposing gaps in the data used to track different components, making it difficult to identify the source of issues that can affect yield and reliability, and opening the door to counterfeit or sub-standard parts.

This involves more than just assigning a simple identifying code to a chip. At different points in a device’s lifecycle, new identifiers are added. Problems arise because all of that identification data needs to be connected at all times, and with exponential growth in tracking data it can become unmanageable very quickly.

Unique identifiers are like fingerprints, except that for semiconductor devices, there could be dozens of them, starting as early as design and running through various manufacturing and packaging processes, and finally into the field. They can be attached to wafers, individual dies, packaged parts, and boards.

There is significant value in identifier data that extends well beyond a single device. For example, data can be collected for all chips manufactured on a single wafer, or for all wafers processed over a period of time. It can be used to track defects or latent defects before devices are shipped, or to monitor their behavior in the field. But it also has to be consistent and connected to maximize its value.

“The call for traceability starts with the fact that we’re dealing with items that are sometimes expensive, but not necessarily very expensive,” said Ophir Gaathon, CEO of DUST Identity. “They all look very much the same. But it’s not just the manufacturing information that has value. It is their life journey and pedigree. It’s everything that has to do with how they got into the end use that can influence how we’re thinking of the risk around those objects, or the performance of those systems.”

Full appreciation of the engineering challenges and potential data analytic scenarios requires a deep understanding of the various stages in the supply chain, where the identifiers are created, environments in which they could be read, and how they are paired with analytic workflows.

Identifiers for electronics traceability
Physical identifiers can tell a chip’s origins and manufacturing history, and they can be used to track performance. But not all identifiers work for all components and situations.

Embedding a unique identifier into a semiconductor device requires area that can only be accessed when powered on. Sometimes, power isn’t available, and for mixed-signal and analog devices, unique identifiers historically have been unavailable.

In general, there are two main usage scenarios. One traces a device’s manufacturing journey forward and backward in time. The other collects device performance data in an end-system. Data from both can create a complete picture. For example, at the die level data about the wafer ID and wafer position can be connected to other data during wafer fabrication and wafer test, especially data from metrology, inspection and other test steps, and all of it can be analyzed together.

Once components (e.g. substrates, dies, passives) arrive at an assembly facility, each component ideally will have an identifier, and the final assembled device can have a unique identifier. OSATs then can connect the die ID to the subsequent identifier for the final assembled package, using unique IDs from the board and final end-system.

SoC designers often use one-time programable memory (OTP) to log wafer ID, die x,y location, and additional metadata. With the growth of security applications and concerns about a trusted supply chain, SoCs often include a physically unclonable function (PUF) in the die. These also can be used for manufacturing analytic workflows.

PUFs are designed around naturally occurring device variability, and several PUF technology options exist. A common one is built around SRAM bit cells.

“SRAM PUF technology exists at birth of the chip at die level, and it can be utilized as soon as the chip is powered, said Geert-Jan Schrijen, CTO at Intrinsic ID. “Its earliest use is at the wafer test stage, where individual dies are powered on for the first time. SRAM PUFs can be used at any stage from that point onward — wafer, die, package, and board.”

More recently other on-die PUFs have been proposed based upon via resistance variation and transistor leakage currents.

“We provide a mixed-signal hard IP block (Quantum-Driven Identify, QDID) that creates a unique fingerprint for the host SoC,” said Chris Jones, director of applications at Crypto Quantique. “This can be read at any time for use in anti-counterfeit or cryptographic functions. The identifier includes an array of CMOS transistors that are integrated into the target SoC silicon design. The uniqueness of the identifier originates from the measured difference in leakage currents across the silicon oxide layer of a pair of standard CMOS transistors. At some point in the manufacturing flow it will be necessary to extract the identity of the device from each individual device.”

Depending up the hardware implementation, the creation of the unique identifier can be on-die or in a cloud-based solution. This has been supported by SRAM PUFs. A new approach is to combine multiple sources of manufacturing variability-based entropy. For example, paraphrased below:

“The proteanTecs Supply Chain Security (SCS) solution is based on generating a signature based on measured in-chip parameters and their dispersion and variability.”

External IDs can include wafer IDs inscribed on the wafer, and 2D barcodes on package substrates and packages. These are optically read. In the past few years, new external ID technologies have been developed that enable an unclonable identifier and reading of an unpowered device.

“Our FemtoTag identifier is created using a laser source, which alters the target material,” explained Steve Walters, chief engineer at Aerocyonics. “It creates a unique marking that is readable with several different modalities. The tag is suitable for use on wafers, die, packages, boards and systems. The terahertz- readable tag can be written below the surface of a die or package, and is far-field readable without requiring direct line of sight. That makes it suitable for placement at lower levels (wafer, die, etc.), while still readable at higher levels of assembly.”

Another technology is based upon microscopic diamond dust particles. “The dust can be applied on the individual components or bare boards, either manually or using coupons,” said DUST Identity’s Gaathon. “There are multiple ways to incorporate dust into the electronic manufacturing workflow. For example, you can put it as part of certain polymers, or you can spray coat. It’s applicable to IC die and packaging components. We can validate our identifier using optical means. We look at the pattern and match that pattern in the same way that biometric works to match a fingerprint.”

Pairing identifiers with analytic workflows
Combining semiconductor device data across manufacturing facilities with in-field performance data provides data engineering teams never had access to before. ICs with telemetry circuit data supply highly nuanced analytic workflows. These are only possible with connected unique identifiers. The possible diverse analytic workflows include:

  • Feed-forward applications in manufacturing, in which engineering teams use data from a previous step to make decisions for a latter step.
  • Comprehension of probable manufacturing signatures that could explain a downstream yield excursion or reliability issue.
  • Using telemetry circuitry, engineers compare a part’s in-field performance with the fleet of those parts, providing spatial and temporal granularity for performance and environmental metrics (e.g., I/O data rate, transactions between functional blocks, temperature).

Over the past two decades, manufacturing data analytic workflows have been put in place to combine data from different sources. Test analytic workflows, in particular, focused on using data from wafer sort, assembly, and various final test insertions, and customized automation infrastructure was developed to enable it. These systems identify dies/units based on bin categories with simple threshold rules. For example, “place die with leakage ≥ X amps in good bin 2 designated for package type B.”

In the last 10-plus years capabilities shifted, facilitating analytic workflows and enabling more sophisticated decisions. The evolution in factory data automation systems and availability of vendor analytic platforms enabled this shift. Building on this infrastructure, the addition of unique identifiers permits sophisticated data-driven decisions for every die or unit, along with real-time decision making.

“Today, we have analytic tools and debug tools, focused on an insertion — for example all the STDF data from the final test insertion. Specifically, we make discoveries and learnings on that insertion,” said Eli Roth, smart manufacturing product manager at Teradyne. “What’s coming is how do you pull and use data from multiple insertions? How can you look at one final test insertion and the next final test insertion to make a comparison because of a step in the middle? Being able to identify those parts so that you can have that data readily available for real-time analytics becomes extremely valuable.”

The value of real-time analytics, and the realities of factory automation databases and data communication factory scenarios have motivated vendors and manufacturers to create the necessary infrastructure for real-time decision-making in a test cell.

That test cell consists of ATEs, handlers/probers, software, and additional computing resources. ATE plays a role in using data from previous test insertions to make test decisions. But it’s also critical to have a flexible and secure connection when connecting the test program-generated data with previous test data.

“Real-time data infrastructure (RTDI) has four components,” said Ken Butler, senior director of business development at Advantest America. “The edge box, along with a floor server are basically the traffic cops for all the data coming off the web and onto the floor. You don’t want your individual test cells talking to the web. Third is the registry, which sits out on the web and has the analytic platform and data. Finally, there is a software communication mechanism. Ours is ACS Nexus. It moves data between the host controller and edge computer, and from the host controller to the floor server and to the outside world, etc.”

Fig. 1: Connecting web analytics to manufacturing test cells on a test manufacturing floor. Source: Advantest

Fig. 1: Connecting web analytics to manufacturing test cells on a test manufacturing floor. Source: Advantest

Teradyne’s Roth described an application using real-time tester analytics to detect parameter drift after a burn-in step. Burn-in provides a thermal and electrical stress to IC devices for screening early life failures. Engineers typically use pass/fail results for screening these failures. Driven by demands for zero field failures, a parameter drift model is used to identify additional failures that would otherwise escape.

Unique identifiers play a central role in applying the machine learning derived model on a per-unit basis. The edge computing environment, coupled with a per-unit pass/fail decision works like a surgeon’s scalpel.

In general, factory automation systems need to track a unique identifier and connect it to its manufacturing identity (equipment genealogy, operation timestamps), to fully enable analytics workflows. “While the product is in manufacturing, good crisp communication of was/before and is/after should be sufficient to form a data link for traceability or genealogy,” said Mike McIntyre, director of software product management for Onto Innovation. “Within the factory communication confines there should be zero potential for outside interference or tampering with unit traceability.”

Other uses
Identifiers extend well beyond a product in manufacturing. There are compliance requirements to track the raw materials used to make those devices, as well.

“Identifiers are key to implementing the equipment in the right way, and to connecting the dots between that equipment and product, but there’s an entire ecosystem involved, said Alan Porter, vice president of electronics industry at Siemens EDA. “It’s not just the manufacturing floor itself. You want to include gate-to-gate, for instance, the movement of the raw materials to the floor. Also, you need the right data and IT technology to capture and push the information (data and metadata) in compliance with required government regulations, as well as a factory’s own metrics like sustainability.”

Traceability is needed in the assembly processes to meet yield, quality and reliability objectives.

“Packaging of FCBGA parts is a complex multi-step process, and the intent of unit-level traceability (ULT) is to be able to track every step of the process,” said Gerard John, senior director of chiplets/FCBGA business unit at Amkor Technology. “This starts with the substrates. Substrates act as an interface to transform the fine interconnect pitch on the dies to something that can easily be mounted to a system board. Substrates supply power delivery from the system to the die, and provide ‘high-fidelity’ transfer of signals between the system and the die. Therefore, tracking them is key to uncovering device-to-device variations caused by different substrate batches or locations on the manufacturing panel. For example, fails associated with panel location may have something to do with the heat profile, or the top right corner has more metal exposed. Either of these root causes can result in a poor die-to-substrate contact.”

Unique identifiers facilitate engineering analysis that joins data across manufacturing facilities, from wafer inspection data to test data, wafer test data to assembling die into different packages. The growth of multi-chip product solutions requires matching performance between different IC components, and binning based on final performance. Called ‘smart pairing,’ this could be achieved for clock frequency and power metrics. And with the proliferation of telemetry circuits, it’s not hard to imagine additional metrics for matching, such as I/O circuit data rates.

Bridging in-field device data to other devices, as well as a device’s own manufacturing history, absolutely requires unique identifiers. With this nascent application, several scenarios that promise benefits to engineering teams. This includes comparing a part’s in-field performance with other parts, linking it to a part’s manufacturing history, and an improved understanding of the mission profile for power, transaction loads, and thermal profile, among other things

Flexibility is key. Identifiers need to be readable in the various end-system environments. For example, sometimes engineers determine the need for a unique identifier later in the system design cycle. And while most on-die identifiers need to be planned during the design cycle, an engineer is in luck if the IC device has an embedded SRAM.

“Identifiers can be retrofitted on existing devices that have uninitialized SRAM available,” says Intrinsic ID’s Schrijen. “This can be implemented by embedding fuzzy-extractor software into devices for deriving keys or using the fingerprint matching algorithms in the cloud/on a server. No hardware changes required since the PUF element is already present.”

The same software-based algorithm and server compute can be supported by existing parametric telemetry circuitry. “Since our Supply Chain Security (SCS) solution is based on our general-purpose monitoring subsystem, any customer using it gets the ability to implement SCS by just adding a software application on our platform,” said Nir Sever, senior director of business development at proteanTecs.

In addition, external identifiers can be read with line-of-sight optics, but that’s not always possible. “In-field usage of passive identifiers can be highly restricted due to readability of tags once deployed in a system,” said Aerocyonics’ Walters. “If tag readability is negated once integrated into an assembly or system, the identifier technology has limited application for in-field use. Additionally, if an identifier technology requires a device to be functional to read its contents, then the identifier is not well suited for addressing product reliability needs.”

Since the industry’s inception, data analysis has been associated with specific manufacturing steps. Now, either through internally developed or commercial solutions, that data can be connected within a factory and leveraged to improve granularity and security, from materials to final product, and into the field.

It will take time for this capability to roll out, of course. So far, its use has been limited to categorizing dies or units by wafer IDs, lot IDs, or test bin categories. But unique IC identifiers open the door to much more nuanced data-driven decisions, as well as the ability to bridge data between manufacturing facilities.

With the expected growth in heterogenous products this capability becomes essential. In-field data and connecting end-system data with a component’s lifecycle data are the next horizon, and the potential for improved reliability and traceability is significant.

Related story
Fingerprinting Chips For Traceability
Diverse identifier technologies enable fingerprinting for all device types.

Leave a Reply

(Note: This name will be displayed publicly)