Cut Apart to Grow: why are we splitting semiconductor chips?
TL;DR
- Shrinking transistors is getting harder, while AI training compute has far outpaced hardware price-performance [1]. Chiplets (split pieces of a chip) help keep computing power growing by splitting a large die into smaller ones that are tested before assembly, so a defect wastes a small piece instead of a whole chip. Chiplets also allow different functions to be built separately and combined in one package. The pieces can be joined side by side (2.5D) or stacked (3D), each with its own advantages and disadvantages. If the product calls for both, a chiplet design may use both. Chiplets are worthwhile if the savings and benefits are worth the added costs for splitting (Section 4).
Why would anyone cut a perfectly good chip into pieces, then pay to glue it back together?
The computing used to train the biggest AI models has grown about 4× per year since 2010, while GPU price-performance has improved only about 1.3× per year (excluding tensor-core jumps) [1]. That gap has been closed mostly by spending more, not by improved price-performance [1], and the chips themselves are running into limits.
The steady growth in computing power enabled the development of whole modern technologies and engineering fields, like AI, that we now take for granted.
Surprisingly, cutting the semiconductor chip into multiple pieces seems to be one possible solution. To see why, we need to understand a few things first.
Gordon Moore predicted that the number of transistors (switches that store and process 0s and 1s) on a chip would double roughly every two years (updated from every year) by shrinking in size. That pattern is called Moore’s Law [2][3]. Shrinking transistors fits more chips on each wafer (the circular silicon disc chips are printed on), which lowers the cost per chip [4], and lets each chip hold more transistors. A transistor has a source and a drain (the two ends of the transistor), a channel between them through which the current passes, and a gate that switches that current on and off.
But Moore’s Law is now slowing down, because shrinking the transistors further has become far more challenging. To put the size of transistors into perspective, the width of a strand of human hair is about 50,000 nanometers [5]. Researchers at Berkeley Lab’s Molecular Foundry have built a transistor with a working 1 nanometer gate! (lab result using carbon nanotubes and molybdenum disulfide, not a product) [5].
Below roughly 5 nanometers [5] (This is physical size, not a label. The “5 nm” or “3 nm” names used for chip generations are now just labels and no longer match any measured part.), conventional silicon transistors are predicted to stop scaling, because they start hitting physical limits that make it much harder to reliably control the flow of electrons from source to drain. Make a transistor too small and electrons start slipping through the barrier on their own, leaking current, an effect called quantum tunneling [6].
Changing the structure of the transistors from flat (planar) designs to FinFETs (3-D “tri-gate” transistors, where the gate surrounds the channel on three sides [7]) and GAAFETs (the gate wraps all the way around [8]) helps keep scaling going, but further structural changes are still needed [8].
Advanced packaging and chiplet integration are now helping keep computing power growing. The semiconductor industry is transitioning from monolithic single-die chips (one piece of silicon holding the whole chip) to modular chiplet architectures (multiple pieces of silicon joined together to form a chip) integrated side-by-side (out, 2.5D) or stacked vertically (up, 3D).
Splitting a large monolithic die into smaller pieces (called chiplets) may help reduce the cost per functional die (each individual block of silicon cut is called a die) and improve overall system performance and computing power.
But splitting isn’t free. So when is splitting a chip worth it? To answer that, we need to compare the yield savings, the added costs, and see how the choice between 2.5D and 3D changes things.
1. The wall
A computer chip is a thin slab of silicon carrying billions of transistors. These allow the computer chip to perform the “smart” processes that we use the computers for. More transistors let a chip do more at once, which is how Moore’s Law translated into better performance. How fast a chip actually gets also depends on its design and other factors [7].
Chips aren’t made one at a time. Many copies are printed in a grid on a wafer (size varies but the industry standard is slightly larger than a standard dinner-plate), then cut apart into dies.
The printing uses light in a process called photolithography. Light is shaped by a stencil of the circuit (called a mask, or reticle) to be printed, and optical systems shrink the pattern onto the wafer. The shorter the wavelength of the light used, the smaller the features it can print [9]. ASML’s EUV (extreme ultraviolet) machines use light of wavelength 13.5 nanometers [9].
On the standard EUV scanner, the largest single shot exposure (patch) is 26 × 33 mm, about 858 mm² [10]. The newest high-NA (high numerical aperture) machines print finer detail with better optical resolution, but their lenses magnify 4× in one direction and 8× in the other (anamorphic optics) while the mask stays the same size, which results in the largest single shot exposure field now being 26 × 16.5 mm = 429 mm² [10], exactly half. A chip bigger than that needs two exposures stitched together (patterns from separate exposures joined into one die), which adds complexity [10].
Think of a loom that can only weave cloth up to a fixed width. To make a wider sheet, you must weave two strips separately and then join them, and the join has to line up perfectly.
The first reason to split a chip into chiplets has nothing to do with the functioning: there is a limit to how big one chip can be without stitching. The second reason is defects.
2. How one defect becomes a costly failure
Chips are built at a microscopic scale, and at that scale things can go wrong. A speck of dust lands on the wafer, a pattern prints slightly off, or a layer of material has a flaw. These can create defects, and they land at random spots. When a defect hits a critical part of a die, it can kill that die and such defects are called killer defects. [11]
Imagine cutting a large sheet of cloth from a roll with random flaws scattered through it. If you cut one big sheet, a single flaw anywhere on that sheet ruins all of it. But if you cut four small pieces and inspect each one, a flaw only ruins one piece and you can bin just that piece. Cutting small looks like the obvious choice until you remember you have to sew the pieces back together. Every seam costs time and money and a seam can be a weak spot.
Semiconductor chips face the same trade-off. Silicon wafers have tiny random defects so one defect can kill a big chip, while a chiplet design loses only the small piece it lands on. But joining the pieces back has its own costs and limits which we will discuss later in this article.
The fraction of dies that work properly (or don’t have killer defects) is called the yield. Poisson yield model is a common statistical model to calculate the yield.
Poisson Yield Model:
[11] [12]
where D₀ is the average number of defects per unit area and A is the die area.
This simplified model, gives the probability that a die contains zero fatal defects, or killer defects. It falls steeply as the area grows.
Here is a worked example with an assumed defect density of D₀ = 0.1 defects per cm², or 0.001 defects per mm². I chose this value only for illustration, so it is not a measured figure.
- One big 858 mm² die: yield ≈ e^(−858 * 0.001) ≈ 42%
- One die a quarter that size, 214.5 mm²: yield ≈ e^(-214.5 * 0.001) ≈ 81%
4 out of 10 big dies working vs 8 out of 10 small dies working. A wafer costs about the same to process whether the dies work or not, so the cost of every failed die has to be compensated by the dies that survive. That’s what makes dies with killer defects costly.
Poisson Yield Model assumes fatal defects occur randomly and independently. However, real manufacturing situations show that defects may cluster in some wafer regions. Let’s picture the cloth analogy again. This time the same number of flaws arrive, but in clumps. More sheets come out clean than expected because the flaws pile up on a few rather than being evenly distributed. The Poisson model ignores this, so it is too pessimistic. [12]
The Negative Binomial Model accounts for this using a clustering parameter
Negative Binomial Model:
[11] [13]
Where α is the clustering parameter. (a high α value means little clustering while a low α means heavy clustering.) A 2008 lecture, crediting a standard textbook, gives an empirically determined typical range of 1.5 to 2.0 for memory and microprocessors [11], while a 2022 cost model uses 10 for recent logic chips, as a more realistic setting than older defaults [13]. I use α = 2 for the examples throughout this article. This is a conservative choice: a higher α would lower the big die’s yield and make splitting look even better.
Taking an example again with the same D₀ = 0.001 defects per mm² as before, and an assumed α = 2 (moderate clustering). Both values are illustrative, not measured.
- Big die, 858 mm² (8.58 cm²): A·D₀ = 0.858. Then 1 + 0.858/2 = 1.429, and 1.429^(−2) ≈ 49%.
- Small die, 214.5 mm²: A·D₀ = 0.2145. Then 1 + 0.2145/2 = 1.10725, and 1.10725^(−2) ≈ 82%.
What the table shows:
- The models barely disagree on small dies. Clustering matters mostly for big ones.
- With realistic clustering, we can still observe the steep drop in yield as area increases, so the area penalty holds for whichever model you use.
3. What splitting buys you
Cutting one big die into smaller ones buys you something, but only if you do one more thing.
The smaller dies survive far more often (82% against 49% in our example). But you need all four to work. The saving comes from inspecting each piece first before joining them. If you join four untested dies together, the chance that all four work is 0.82⁴ ≈ 45% which is slightly worse than the big die by itself. Each small die is tested on its own, and only the good dies are packaged together. That is what turns the 82% into a real saving. Cutting the cloth saves nothing if you sew them together without checking.
One of the measures to compare the approaches is to see how much wafer area is spent per working chip.
A worked example using the same assumptions as Section 2:
(D₀ = 0.001 per mm², α = 2):
- One 858 mm² die: 858 ÷ 0.49 ≈ 1,751 mm² of silicon per working chip.
- Four 214.5 mm² dies, tested first: 858 ÷ 0.82 ≈ 1,046 mm² per working set.
- Splitting uses about 60% of the silicon (1,046 ÷ 1,751 ≈ 0.60).
However, this is the best case. It ignores the extra area for connections between the pieces, packaging cost, and assembly losses. Let’s see how those costs add back in Section 4.
The gains also shrink as you cut more. Two pieces would save 28% of the silicon and four save 40% (from our example), but eight would save about 46% and sixteen would save 48%. The extra costs of splitting may outweigh the savings of going from eight to sixteen (2 percentage points).
After area overhead and assembly losses, going from eight pieces to sixteen makes things worse, not better, so there is a sweet spot rather than a “more is better” rule.
Splitting also lets manufacturers build each kind of die on the process that suits it, and reuse the same pieces across products. In AMD’s Zen 2 EPYC, CPU cores sit on several compute dies, while a separate I/O (Input/Output) die handles the communication of the circuit or chiplet with other parts of the system. Both dies are manufactured on different processes. [13]
So the question becomes: how big is the bill?
4. The Bill
Splitting a chip is not free. Just like seams between the pieces of clothes cost thread and time, assembly of chiplets does too. Moreover, the assembly can fail just like the seam can fail too.
For splitting, the bill has five basic charges:
- Area and Packaging Overhead for Die-to-Die (D2D) PHY Circuits and Interconnects:
Each chiplet requires dedicated circuitry and physical interfaces to communicate with neighboring chiplets, demanding additional silicon area. This silicon area overhead represents a significant cost. Furthermore, the physical connection substrates (foundational support layers) themselves add substantial material and manufacturing costs to the overall system. [13] - Packaging [13]:
The pieces need something to sit on: there are three common packaging approaches:
1. MCM (multi-chip module, or SIP, system-in-package): the pieces sit side by side on one shared base, called a substrate.
2. InFO (integrated fan-out): A thin layer of fine wiring, a redistribution layer (RDL), links the pieces.
3. Interposer: the pieces sit on a silicon interposer, a slice of silicon carrying fine wiring. The interposer is made like a chip, so it can have its own defects too. - Assembly:
If joining fails, the good pieces on that assembly may have to be scrapped. This is significant when dies are expensive and packaging yield is poor. More pieces means more joins, and more chances to fail. [13] - Fixed costs:
NRE (non-recurring engineering) is the one-time cost of design, verification and masks. Every distinct chiplet needs its own masks and design work. RE (recurring engineering) is the per-chip cost of wafers, packaging and testing. [13] - Testing:
KGD, or Known Good Die, is, as the name suggests, a bare die that has been fully tested and passed specific tests to meet electrical and performance specifications before packaging. Each piece must pass testing before it is joined, and the finished assembly is tested afterwards. Testing matters because it prevents a bad die slipping through and wasting good pieces joined to it. [14] [15]
Testing can happen at several points in several ways: Wafer Sort (Bare Die Probe), Pre-Bond Metrology, Chip-on-Wafer (CoW) module probe, System-Level Test (SLT) [16] [17]. Each stage has its own mechanism and cost.
Perfect KGD testing is not practical because of the physical test discrepancies between the testing situations and the final packaged performance situations as die-level tests can’t always reach every test vector [15].
So, chipmakers build Design-for-Testability (DFT) techniques such as Built-In-Self-Test (BIST) sequencers (automated test routines), and redundancy/repair logic (reroutes signals around detected micro-defects where the design allows it) directly into the silicon [15].
Let’s revisit our running example with the updated charges (D₀ = 0.001 per mm², α = 2, four chiplets):
Assuming 10% extra silicon for links [13], each piece grows to 235.95 mm², and the yield becomes 80%, so silicon per working set is 943 ÷ 0.80 ≈ 1,180 mm², or 67% of the single monolithic die (increased from 60%). Assuming assembly succeeds 99% of the time for each of the four chiplets, so for the assembled package: 0.99⁴ ≈ 96%, silicon per working set becomes 1,180 ÷ 0.96 ≈ 1,229 mm², or 70% of the monolithic die.
5. Was splitting even worth it in the first place?
The answer depends on what “worth it” means. Splitting, as we saw, can reduce the silicon wasted on defective dies, but that doesn’t necessarily mean the finished chip costs less. The fair comparison is when it lowers the cost of each working finished chip.
From my analysis, chiplets pay off when the cost avoided due to discarding fewer defective dies exceeds the additional costs of linkage, packaging, testing, assembly, and design.
As a rule: silicon saved per working chip × wafer cost per mm² is greater than extra packaging, testing and assembly cost per chip + design and verification cost per chip.
In our running example (using illustrative values), splitting saves 1,751 − 1,229 = 522 mm² of silicon per working chip. At an assumed $0.24 per mm² (illustrative, roughly a $17,000, 300 mm diameter wafer, before edge losses), that is about $125 per chip. So the extra packaging, test costs, design costs must total less than about $125 per chip, or the estimated economic advantage disappears.
The same model can be used to see how the saving disappears when splitting an already small die. Imagine we take a monolithic die of the size 214.5 mm², with a yield of about 82% (from our example), it needs about 262 mm² silicon per working chip. But split it into four with the same 10% area overhead (now 59 mm², yield computes to approximately 94%), and 99% assembly success rate per piece (about 96% for all four), it now needs about 261 mm² of silicon per working chip. That is nearly no saving at all, even before adding the other costs. Compare that with the 30% for the big die (858 mm²).
Splitting becomes more attractive when the monolithic die is large and defects make each failed die expensive, while chiplets can be tested and used efficiently. It can also help when different functions benefit from different designs and manufacturing processes. High volume production can make the design and verification costs easier to recover. [13]
Monolithic dies become more attractive when the original design is small or already yields well, when package connections are expensive, or when dies must exchange data so frequently that the added links affect the power or performance of the chip.
So there is no universal die size at which splitting becomes worth it. The threshold depends on defect economics, packaging and test costs, production volume, and the functions of the chip itself. The worked example shows why splitting can save silicon material but, to decide whether a real design saves money, those other costs and constraints have to enter the conversation too.
6. So, out or up?
The first way to assemble the chiplets back to form a chip is to place them side by side on one shared base (out). This is called 2.5D [18]. The wiring between the pieces can come from a full slice of silicon, called an interposer (which can itself have defects) [13], or from small silicon bridges [17].
Building out allows the design to combine compute, memory, or dies performing different processes across the package. It also lets one package hold more silicon than one full exposure shot prints [19]. The trade-off is that the wiring base adds cost and package area.
The second way, three-dimensional stacking (up), puts one die on top of another and connects them vertically [18].
The advantage of 3D stacking is how many connections fit in the same area. Standard microbumps (tiny solder bumps that link a chip to the interposer or base beneath it) sit about 40–50 µm apart, while with hybrid bonding, where the dies are bonded directly, that can be brought down to below 10 µm [16]. In principle, Building up can bring parts of a system closer together, shorten connections between layers and pack more circuitry in a smaller footprint. That is useful when two layers need to communicate frequently, or when a memory stack needs a very wide connection to a processor. Stacking does not automatically make every connection faster; it still depends on how the whole system is designed.
But placing layers close together creates a different problem: heat. For a die sitting flat on its substrate, heat spreads out and has a direct route to escape through the heat sink, but in a 3D stack, heat has to flow from the bottom layers to the top layers to escape to the heat sink [18][20]. Stacking also makes assembly difficult as the surfaces must be polished extremely flat [18], and testing a partially assembled multi-die system is a new kind of test challenge [16].
Well, in practice, a package can use both [17]. TSMC’s CoWoS-S, for example, places logic chiplets and stacks of high-bandwidth memory (HBM: memory dies stacked with a very wide connection to the processor) on a large silicon interposer [19]. Stacked dies are joined by TSVs (through-silicon vias, tiny vertical wires through the silicon) [18]. So the system is built “up” for the memory stack and “out” across the interposer.
The choice between them depends on what needs to connect and why they need to connect, and whether the gains justify the added cost and complexity.
7. So, is it worth it?
Advanced packaging and chiplet integration help computing power keep growing beyond the limits of a monolithic die. It pays off only when the yield savings and other benefits are worth the additional costs of links, packaging, testing, assembly, design and challenges. In my illustrative example, four pieces needed about 70% of the silicon of one big die after link overhead. That 30% saving disappears if the total bill costs more. So chips grow past the limits of a monolithic die using advanced packaging and chiplet integration, but with trade-offs.
8. What surprised me and what I loved learning
The cost I’m least sure of is design and verification. I left it out of my numbers, treating everything beyond silicon as a budget (about $125 per chip) that packaging, testing, assembly and design together would have to stay under, and [13] describes the one-time engineering costs of multi-chip designs as very high. Next, I would study these costs in a published chiplet product’s cost breakdown, to see how large they really are next to silicon. I would then use those numbers to build a simulation model of when splitting pays off.
Two things surprised me the most. First, how those one-time costs behave: they are spread over every chip sold, so they shrink per chip as volume grows, which is why splitting can make sense for a product made in really high volume and not for one made in low volume [13]. Second, I initially expected more pieces to keep helping, but after overhead and assembly losses, 16 pieces saved less than 8. Splitting has a sweet spot, and past it you are losing more than you’re saving.
9. Limitations/What is still unresolved
- The cost models are simplified and the yield models use illustrative values.
- The inputs for the clustering parameters are uncertain: A 2008 lecture gives a typical range of 1.5 to 2 [11], while a 2022 cost model calibrated for recent logic uses 10 [13]. I use α = 2. A higher α makes splitting look more attractive, so my example may understate the saving.
- Numbers are rounded off.
- Several sources are old (Moore 1965 and 1975, LaPedus 2012).
- The focus is on the cost case for chiplets; it does not compare chiplets with continued transistor scaling or claim that chiplets will replace it.
- The model assumes testing catches every defective die; in practice some slip through [15].
About The Author:
Hey! I’m Saket and I’m a 17-year-old based in Chennai, India. I’m currently focusing on the semiconductor field, while simultaneously exploring other STEM fields.
If you liked this article, follow my profile to see more of my content! If you have any feedback, opportunities, or would simply like to connect with me, here’s my LinkedIn: https://www.linkedin.com/in/gagan-saisaket-saranam-a2986a3a9/?isSelfProfile=true
References/Sources:
- [1] Bengio, Y., et al. (2025). International AI Safety Report 2025, section 1.3.3, pp. 54–55. https://internationalaisafetyreport.org/publication/international-ai-safety-report-2025
- [2] Moore, G. E. (1965). Cramming more components onto integrated circuits. Electronics, 38(8), 114–117. https://www.cs.utexas.edu/~fussell/courses/cs352h/papers/moore.pdf
- [3] Moore, G. E. (1975). Progress in digital integrated electronics. IEEE International Electron Devices Meeting Technical Digest, 11–13. https://www.lithoguru.com/scientist/CHE323/Moore1975.pdf
- [4] Hu, C. (2010). Device fabrication technology. In Modern Semiconductor Devices for Integrated Circuits (Ch. 3). Pearson. https://www.chu.berkeley.edu/wp-content/uploads/2020/01/Chenming-Hu_ch3.pdf
- [5] Desai, S. B., Madhvapathy, S. R., Sachid, A. B., Llinas, J. P., Wang, Q., Ahn, G. H., Pitner, G., Kim, M. J., Bokor, J., Hu, C., Wong, H.-S. P., & Javey, A. (2016). MoS₂ transistors with 1-nanometer gate lengths. Science, 354(6308), 99–102. https://doi.org/10.1126/science.aah4698 (open access: https://escholarship.org/uc/item/5157q7f4). Hair comparison from the press release: Lawrence Berkeley National Laboratory. (2016, October 6). Smallest. Transistor. Ever. https://foundry.lbl.gov/2016/10/06/smallest-transistor-ever-2/
- [6] Choi, C. Q. (2022, March 11). These transistor gates are just one carbon atom thick: Researchers may have hit a hard limit. IEEE Spectrum. https://spectrum.ieee.org/smallest-transistor-one-carbon-atom
- [7] Intel Corporation. (2012). Moore’s Law fun facts: Exactly how small is 22 nanometers? [Fact sheet]. https://www.intel.com/content/dam/www/public/us/en/documents/corporate-information/history-moores-law-fun-facts-factsheet.pdf
- [8] Van den hove, L. (2022, August 2). 20-year semiconductor roadmap: Tearing down the walls. imec. https://www.imec-int.com/en/articles/20-year-roadmap-tearing-down-walls
- [9] ASML. Light and lasers. https://asml.com/technology/lithography-principles/light-and-lasers
- [10] Xu, D., Gillijns, W., Wu, S., Adams, A., Wiaux, V., Philipsen, V., Zhang, X., Tejnil, E., & Fenger, G. (2024). OPC and modeling solution towards 0.55NA EUV stitching. Proceedings of SPIE, 12953 (Optical and EUV Nanolithography XXXVII). https://doi.org/10.1117/12.3010519
- [11] MIT OpenCourseWare. (2008, March 11). Yield modeling [Lecture 10, 2.830J/6.780J/ESD.63J; slides crediting May & Spanos (2006), Ch. 5]. https://ocw.mit.edu/courses/2-830j-control-of-manufacturing-processes-sma-6303-spring-2008/4aff1e21de13870355ef44dbe71f45c6_lecture10.pdf
- [12] IEEE Technology Navigator. Integrated circuit yield. https://technav.ieee.org/topic/integrated-circuit-yield/
- [13] Feng, Y., & Ma, K. (2022). Chiplet Actuary: A quantitative cost model and multi-chiplet architecture exploration. Proceedings of the 59th ACM/IEEE Design Automation Conference (DAC), 121–126. https://doi.org/10.1145/3489517.3530428 (open access: https://arxiv.org/pdf/2203.12268)
- [14] Palesko, A., & Palesko, C. (2014). Cost and yield comparison of wafer-to-wafer, die-to-wafer, and die-to-die bonding. IMAPS International Symposium on Microelectronics. https://imapsource.org/api/v1/articles/57531-cost-and-yield-comparison-of-wafer-to-wafer-die-to-wafer-and-die-to-die-bonding.pdf
- [15] LaPedus, M. (2012, December 13). Welcome to the “probably good die” era. Semiconductor Engineering. https://semiengineering.com/welcome-to-the-probably-good-die-era/
- [16] Teradyne. (2026, July 27). AI chiplet architectures redefining test insertions. https://www.teradyne.com/2026/07/27/ai-chiplet-architectures-redefining-test-insertions
- [17] Intel Foundry. Advanced packaging innovations | Chip packages. https://www.intel.com/content/www/us/en/foundry/packaging.html
- [18] Laube, P. 3D stacking. Halbleiter.org (last updated October 15, 2026). https://www.halbleiter.org/en/advanced-packaging/3d-stacking/
- [19] TSMC. CoWoS®. TSMC 3DFabric. https://3dfabric.tsmc.com/english/dedicatedFoundry/technology/cowos.htm
- [20] Mathur, R., Chao, C.-J., Liu, R., Tadepalli, N., Chandupatla, P., Hung, S., Xu, X., Sinha, S., & Kulkarni, J. P. (2020). Thermal analysis of a 3D stacked high-performance commercial microprocessor using face-to-face wafer bonding technology. Proceedings of the IEEE 70th Electronic Components and Technology Conference (ECTC), 541–547. https://doi.org/10.1109/ECTC32862.2020.00091 (open access: https://arxiv.org/pdf/2007.16179)