Page 37 - ChipScale_Sep-Oct_2020-digital

Summary
Heterogeneous packaging is going to
be the main driver for today’s and future
semiconductor packaging in a variety
of applications from system in package
(SiP) to chiplet packaging. The main
process challenge for these multi-chip
expensive packages remains how to keep
the package cost as low as possible so
we can reference Moore’s Law as being
economically valid, even 50-plus years
after the invention of integrated circuits.
Final package inspection is the last
process step to determine heterogeneous
package/product yield. Therefore, having
the most capable solution for final
package 2D/3D inspection is inevitable in
order to enable heterogenous packaging
and drive the industry forward.
Our key vision capabilities such as
large field of view (FOV), stitching, and
Figure 6: Deep-learning second-inspection process results. high throughput, have been specialized

defect modes are scratch, FM (particle), management system. Our coherent by our heterogenous package inspection.
mold scratch, mold crack, incomplete integration of machine vision and deep- In addition, integration of our deep-
mold, Cu exposed, and chip-out, learning engine (Figure 5), is considered learning technology and machine vision
respectively. as edge computing (processing) prior to a into one process has resulted in yield
customer’s yield management process. improvement and significant cost savings
Deploying deep learning on edge In a “regular” cloud-computing for our customers: top heterogeneous
vs. cloud computing system, all data processing is done by an chip makers.
Last, but not the least feature of our AI AI server. In this configuration, the units
system is edge computing, demonstrated per hour (UPH) is impacted because of References
in Figure 7. Some semiconductor the data transferring time between the 1. P. Wesling, et al., “Heterogeneous
manufacturers perform AI deployed inspection machine and the AI server. integration roadmap,” Ch. 8: Single-
on cloud computing by their yield If the server is down, all machines are chip and multi-chip integration,” IEEE
down too. INTEKPLUS 2019 Edition.
edge computing means 2. S. Chitchian, et al., “Combined image
the inspection engine has processing algorithms for improved
self-computing power for optical coherence tomography of the
image capturing and data prostate nerves,” J. Biomedical Optics
processing. UPH is not 15, 2010.
impacted because there is 3. S. Chitchian, et al., “Segmentation of
no need to send inspection optical coherence tomography images
images to a server. The for differentiation of the cavernous
cloud server performs AI nerves from the prostate gland,” J.
learning and model training Biomedical Optics 14, 2009.
so there is no need for a 4. S. Chitchian, et al., “Denoising
high-performance server. during optical coherence tomography
In addition, if the server of the prostate nerves via wavelet
is down, all machines can shrinkage using dual-tree complex
continue running. wavelet transform,” J. Biomedical
Figure 7: Deep learning deployed on edge computing vs. cloud computing. Optics 14, 2009.

Biography
Shahab Chitchian is Chief Strategy Officer and R&D Director at INTEKPLUS Corporation Ltd., Daejeon,
South Korea. Previously, he was Senior Staff Process Engineer at Samsung Electronics, Semiconductor Test
and System Package division where he worked on advanced FOWLP development and high-bandwidth memory
packaging. He also was a Senior Process Engineer at Intel Co., Assembly and Test Technology Development
(ATTD), worked on EMIB package development. He received his MSc and PhD in Electrical and Optical
Engineering, from the U. of North Carolina at Charlotte. Email shahab@intekplus.com

35
Chip Scale Review September • October • 2020 [ChipScaleReview.com] 35