Secure Handling Of Financial Data In Manufacturing

Experts at the Table: Semiconductor Engineering sat down to discuss the advantages associated with linking financial data with manufacturing data analytic platforms, real security challenges and the best uses for AI/ML methods, with Dieter Rathei, CEO of DR Yield; Jon Holt, senior director of product management at PDF Solutions, Alex Burlak, vice president of advanced analytics and test at proteanTecs; and Dirk de Vries, technical program manager and senior architect at Synopsys. What follows are excerpts of that conversation. To view part one of this discussion, click here.

(L-R): ProteanTecs’ Burlak, DR Yield’s Rathei, Synopsys’ de Vries, PDF Solutions’ Holt.

SE: Where do you think AI and machine learning are best applied in IC manufacturing?

Rathei: AI can be valuable if you have a large data set on which you can train on. We don’t have that yet for these kinds of financial decisions. Even if you have all the financial data and all the engineering data available, we don’t have a training set of what was the correct associated financial decision at each point in time. In order to make these qualified decisions, we would need to give engineers and fab operating personnel access to all the financial information first. We already said this is probably unlikely to happen very soon. The best approach to these problems would be to just think through the typical applications and use classical optimization algorithms for these financial decisions. I know it doesn’t sound very trendy at this time to not advocate for AI, but that’s where we are for this particular application.

SE: Can you use fake data?

de Vries: Maybe, but you need many examples if you want to train a neural network. And we’re talking about millions of examples, not hundreds or thousands. If we look at things that happen on the die level, you can do training. But if you’re looking at decisions that are in the hundreds or thousands, then typically that won’t pan out for training a network. I completely concur with what Dieter is saying. This is not the area where we think such a technique would be successfully applied.

SE: Is that likely to change with chiplets?

de Vries: If you look into other topics, like smart assembly when you do performance matching, selecting chiplets that go together – that’s a different story. There, you’ve got a lot more data to train on. But if we look at financial decisions in terms of manufacturing logistics, it’s tricky.

Holt: With some big ERP companies, we’ve been focusing on this challenge for a couple of years. All of these barriers to entry are valid. But there are training approaches being implemented at some major fabless and memory companies. Let’s take a step back. We have AI, machine learning (ML), deep learning, GenAI (i.e., large language models). With machine learning you build a model to do predictions, for example, of yield. One approach is using a GenAI and LLM to link these models together — not the data, but the output. For example you have a small language model on a small amount of data. Maybe you’re predicting the parts that go together well in the chiplet, or you’re identifying a yield issue in the fab. These are all separate models, but these agents can talk to each other. Also, LLM models have been created for optimizing the scheduler in financial data itself, but they don’t have access to the customer-specific IP, such as yield and performance data. There’s a barrier between those two. But the decisions made can be used in the LLM model to determine product prioritization, like how you balance the product orders in-line and how you order. There is some value that can be obtained right now. Is it efficient? No. But is it effective? It looks like it has some promise.

Burlak: The bigger problems require vast amounts of data to solve, which is not necessarily available except for more specific problems, or with specific KPIs that can be improved using machine learning or analytics. There is a lot of value that can be gained. For example, you can move up screening from final test to wafer sort. Using parametric data from on-chip agents and an ML prediction model, you can correlate final test yield versus wafer sort yield. This will allow you to shift-left decisions to screen parts earlier in the process. There is always some ROI in these models. You might be over-killing versus catching the true things that you would like to catch. Bringing these equations and the models, in terms of cost optimization, eventually you can achieve that. And then you can utilize that during production — for example, these models that were developed in the analytics platform to the manufacturing line — and then make decisions during wafer sort as to whether some parts will be screened now rather than assembled and screened later. So this kind of specific goal impacts financial data or utilizes financial data, and we have customers using this approach.

Holt: We see that, as well. We call it dynamic product costing, but it’s feeding the manufacturing data. You know what it takes if you’re retesting, or the cycle time, or the number of tests or holds or cycle time in line, and that impacts the cost of that product. That data is fed to the financial system, which is then used to determine the price of the part. So that’s actually operating in the other direction.

de Vries: With the case that Alex described, there’s some subtle reasoning to consider. If you look at an MCM or chiplet, fundamentally an individual die can take down an entire part. If you have some kind of a sensor for face recognition in a phone, that’s a complex MCM with maybe eight or nine dies inside. One failure can take down the entire unit. The financial impact is very large. And we see there’s a bigger opportunity to improve. When there’s a one-to-one correspondence — one die, one packaged part, typically our customers’ test engineers do a decent job. Where they combine the data, the optimization requires a little bit more work. However, we do notice that in these complex MCM parts, there are a lot more opportunities because the traditional analysis techniques optimize everything. It’s an interesting situation where there’s more opportunity for improvement. That’s especially true in the MCM, 3D-IC, and chiplet space where these financial considerations are extremely relevant.

SE: How is data security best handled?

Holt: One approach is a black-box, where the output of the model is taken as an input to a larger model. That’s one way, and it makes it challenging. If the output is not as expected, and you want to drill down to determine a root cause or diagnostic, then you face the challenges of data protection. However, the usefulness can still be obtained in using that model. It would be great if data protection wasn’t such an issue — especially as we’re going to 3D-ICs, as Dirk mentioned, where you have  60 to 70 ICs built onto a substrate. As Alex mentioned, you want to optimize the part matching. And as Dieter said, when you’re manufacturing these advanced assembly packages you want to perform things like inspection. You need to know what the layout is to do an optical inspection, or at least the top-layer layout of every component. You want to be able to test these components, and you want to be able to match the performance. Data sharing is a real issue in advanced packaging manufacturing. When you were manufacturing a single chip, all that data was available to you. When the data comes from multiple suppliers, a partial answer is the black-box approach.

Rathei: IDMs have all the data available within their organization. That’s why these kind of applications appear to be less difficult there. But even at IDMs there are different levels of confidentiality. Not everybody in the organization gets the information they would need to make all these decisions. One aspect we should talk about is how this financial data can have big impacts on test flow operations. How does this work with OSATs? Because if you want to optimize your whole process flow, you also need to consider the part that may be outsourced to a vendor like an OSAT. It depends on how the contracts are made there. In most cases, you just pay the OSAT for a wafer being tested or certain parts being tested. And it’s not the responsibility of the OSAT to optimize the yield. But on the other hand, we want all these kinds of decisions that we talked about. This will become a very complex data integration issue. How do we deal with this kind of data sharing between advanced OSAT operations and factories in this kind of environment?

de Vries: That’s a very difficult question. Black box is indeed a common technique. It basically processes the data and then indicates to the operator that perhaps these two bins are worthwhile retesting in terms of recovery rate and test time. That’s an example of a black box approach, where the financial data is hidden. One other approach, for instance, when we’re looking at a dashboard, would be to normalize for the wafer price. Then you can express a delta with a yield target in terms of arbitrary units. They are the same for the different products, so you can make a decision that directs your efforts. But there’s no tangible dollar amount. That normalization strategy is based on the pareto — in a sense, the breakdown of where the most revenue is lost is stable, but only on a relative scale.

Rathei: One problem that I see with this approach is how easily it could be reverse engineered. If an engineer sees the normalized yield loss on a product and they also have access to the actual yield loss, they can quite easily decode the normalization.

de Vries: The normalization of the yield loss in the percent itself really doesn’t work. There are a lot of smart engineers, and they can reverse engineer stuff. But if you look at pricing information, and convert it into a dollar amount, then basically it becomes a multiplier. This is not easy to reverse engineer. If you say that, for the entire operation, revenue loss is a certain level, that’s one way you can break it down by product. Then you need to make a lot of assumptions to get behind the real financials. An interesting example is when I worked for a collaboration of multiple large IDMs. They thought about how to conceal our yields by normalizing yield. But then you cannot perform any relevant analytics. In the end, the yields were shared, but the financials were kept separate.