• In recent years, there has been an increased focus on fan-out wafer level packaging. While fan-out wafer level packaging may be the right solution for some designs, it is not always the lowest cost solution. Flip […]

  • A little over a year ago, I wrote a Knowledge Portal entry about the cost of 3D ICs. Here I am again to tackle the issue of 2.5D and 3D cost reduction from a slightly different angle. This entry is based on what […]

    • Thanks for your comment! Those assembly yields are the kinds of yields we would expect to see at HVM. So no, they are not necessarily reflective of yields today. As for your second question, the total costs are in line with what we have seen of Fury X prices in the marketplace.

    • Only 1% test costs for a complex multi-die chip like AMD’s Fury (consisting of an interposer base die with on top four four-die HBM stacks and a large GPU die) seems rather unrealistic to me. The DfT features aboard the GPU are likely to take >5% silicon area. Also the HBM DRAM stacks have built-in DfT features. These DfT costs should, in absence of a separate “DfT costs” category be attributed to “Test” costs. Also the total actual test execution costs of this complex stacked-die product are bound to be significantly larger than 1%; see the talks by AMD, SK hynix (the HBM DRAM supplier) and their suppliers FormFactor and Teradyne at the IEEE 3D-TEST 2015 Workshop (http://3dtest.tttc-events.org).

      • We definitely agree that there is a significant amount of testing required for a design as complex as the AMD Fury. We also believe the testing cost is quite high today and will decrease over time. However, for this analysis we did not include any testing other than basic open short testing to find defects created during assembly. Our pie charts do not include wafer probe test, functional test, or burn in test. That is why our estimated test time is so small compared to the total test time to build and manufacture the complete product. That being said, we also believe that even the open short testing today is higher than our estimate, but will improve significantly as the product moves into HVM.

  • Through silicon via (TSV) technology is a key design element being incorporated into more and more advanced packaging designs today. TSVs offer distinct benefits in form factor and improved performance and can […]

  • ThumbnailLast year, I did an analysis that included the topic of wafer-to-wafer bonding. Specifically, it was a comparison of the three variations available when stacking wafers and/or die—wafer-to-wafer (W2W), […]

    • Excellent points, David. You’re right, I operated with a fairly narrow and strict set of assumptions here, and the result is a very generic model that doesn’t accurately reflect any single option out there, including the Tezzaron process. This analysis just serves to capture the potential cost and yield drivers for any bonding process. Given the right amount of time, information, and bandwidth, I think it would be very interesting to build accurate models of multiple processes, including yours, and then compare those.

    • 1.9D, yes, this is definitely true–I would hope it isn’t a huge surprise to most people that defect density matters! The point of this analysis had more to do with quantifying the ways in which defect density (and other factors) matter. Most folks do already know large die are more sensitive to defect density–but how large is large? And how sensitive are they really? Even though we’re using a generic model of the bonding process, we wanted to actually present a few numbers to attach to the concept.

  • As a cost modeling company, when we were asked to speak at 3D ASIP this past December, the initial topic choice seemed fairly obvious. We decided to tackle the question of whether the cost of 3D ICs will ever be […]

    • With TSV still being far out, when will TGV play a role?

      • Hi Dirk,
        Actually, TSVs are not far out and are being implemented in many applications today including MEMS, Image sensors, and now 3D stacked memory. The question is when/if they will be used in consumer products. TGVs are still in development, and the infrastructure needs to be more fully determined, so they still have a ways to go.

    • On the topic of TGVs, I’ll add that, from a cost modeling perspective, those aren’t as well understood as TSVs. We haven’t seen actual costs anywhere (maybe they are out there and we just don’t know where to look!). We keep hoping to do a glass-focused cost model with someone, but so far it hasn’t happened.

    • Thank you for the great input. Do you have a cost model for the TSV? I am also very interested in figuring out a cost model for TGV and perhaps our knowledge base could complement each other here.

      • Dirk, we do have a TSV cost model. I will follow up with you directly so as not to spam Francoise’s comment page. 🙂

        • Amy, I don’t consider it spam when it’s topical and beneficial to our readers! If you were trying to sell Prada shoes, that would be a different story. Thanks for reaching out, Dirk!

  • My Articles