Amkor's R-LTC: Why Bigger AI Packages Need New Bonding
AI accelerators are no longer limited by how small a transistor can get. They are increasingly limited by how large a package can be assembled without bending. Amkor's R-LTC talk at IMAPS 2026 is a useful window into that problem: it is a bonding method designed around warpage, not speed.
Below is what the public record actually supports, why big packages warp, and how the main bonding methods differ in where they put heat.
KEY TAKEAWAYS
1. TSMC says it is producing 5.5-reticle CoWoS today and plans a 14-reticle version for 2028, roughly 2.5x the area. Warpage scales with that footprint.
2. Copper expands about 6.4x more than silicon per degree (16.7 vs 2.6 ppm/°C). The wider the heated zone during bonding, the more that mismatch turns into bending.
3. Amkor's R-LTC heats the joint from below with a laser while a bond head above controls height and tilt. Amkor cites better thermal uniformity, lower chiplet stress and warpage control, but no measured figures are public yet.
What Amkor actually said
Amkor's event page for IMAPS 2026 (Boston, September 28 to October 1) lists a talk by Woongpyo Lee, Senior Director and Advanced Interconnection Project Leader, titled Mitigating Thermal Stress and Warpage in Heterogeneous Chiplet Integration Using Reverse Laser Thermo-Compression Bonding (R-LTC). The company says the talk shows how R-LTC improves thermal uniformity, reduces chiplet-side stress and enhances reliability through warpage control, for next-generation 2.5D and 3D integration.
A year earlier, Seokha Na of Amkor Technology Korea described the approach to Semiconductor Engineering as "essentially [a] combination of R-LAB and TCB that targets high and large warpage modules," adding that bump wetting "is mainly done by the reverse laser heating."
What I could not verify
Social posts circulating the story attached a "3x warpage reduction," a 54 x 46 mm package, a $12 billion U.S. investment and customer prepayment deals. None of these appear in Amkor's official event page, and I could not locate the slides. They are left out of this post.
Packages are getting very large
A reticle is the largest area a lithography scanner can print in one shot, about 26 x 33 mm. No single die can exceed it, so more compute means more chiplets and HBM stacks placed side by side on an interposer.
In its April 2026 North America Technology Symposium release, TSMC said it is "now producing 5.5-reticle size CoWoS" and that a 14-reticle version, holding roughly 10 large compute dies and 20 HBM stacks, "is slated for production in 2028." A 40-reticle SoW-X system-on-wafer is expected in 2029. In area terms that is roughly 4,700 mm² today versus 12,000 mm² in 2028 (my arithmetic, assuming 858 mm² per reticle).
Amkor's own program at the same conference flags the same issue: its chiplets/FCBGA talk lists thermal management, signal integrity, power delivery, warpage and design complexity as the challenges of large-format packages.
Why big packages bend
A package is a stack of materials that grow by different amounts when heated. Silicon expands at 2.6 ppm/°C, copper at 16.7, FR4-class organic substrate at 11 to 17, and tin-lead solder at 27, according to figures cited by Semiconductor Engineering.
Bonding means melting solder, so the stack goes hot and comes back down. Thermo-compression bonding (TCB) runs at 250°C to 400°C depending on the metal. As layers cool at different rates, they pull on each other and the package curves. The same mismatch over a larger area produces a larger absolute deflection.
Warpage shows up as opens at the package edge (bumps that never touch the pad), bridges (bumps squeezed into neighbors) and residual stress that can crack chiplets or joints later in life. At these sizes, bonding becomes a thermal-management problem as much as a placement problem.
Three ways to make the joint
Mass reflow puts the whole assembly through an oven. It is the cheapest and fastest option but struggles with large, thin substrates and fine bumps.
TCB uses a heated head that presses each die down. Warpage control improves, but throughput falls and the heat travels through the die to reach the bumps.
Laser-assisted bonding (LAB) heats locally and briefly. Independent work shows why that matters: a 2024 IMAPS paper from Université de Sherbrooke and IBM Canada bonded a 25 x 26 mm silicon die with LAB and cut post-assembly die warpage by about 50% versus mass reflow. The same paper found the laser beam had to extend beyond the die edge to get repeatable, defect-free joints, a reminder that large dies are hard to heat evenly.
R-LTC moves the laser underneath. Per Amkor's description, the laser passes through a laser-transparent stage block to heat the joint from below, while the bond head above combines temperature, pressure and force to hold standoff height, alignment and die tilt. Amkor says it transfers less heat to the module than TCB.
Where Amkor's numbers sit
Amkor reported Q2 2026 net sales of $1.90bn, up 26% year on year. Advanced products were $1.557bn, about 82% of sales by my calculation, and computing was 22% of end-market mix. Q3 guidance is $1.95bn to $2.05bn, with 2026 capex of roughly $2.5bn to $3.0bn.
None of that is attributable to R-LTC. Amkor has not disclosed which products, customers or timelines use it. I read the IMAPS talk as a signal of where assembly engineering effort is going, not as a revenue event.
What I actually watch
| Checkpoint | Why it matters |
|---|---|
| IMAPS paper publication | First measured warpage and stress data |
| TSMC 14-reticle CoWoS timing | Sets the size target for assembly |
| Amkor Q3 results (late Oct) | Computing mix and capex range |
| Bonder order disclosures | Whether laser tools gain share vs TCB |
Value chain read-through
| Segment | What warpage changes |
|---|---|
| OSATs | Choice of bonding process |
| Bonding equipment | TCB vs laser tool demand |
| Package substrates | Large-area flatness |
| Underfill and materials | Residual stress relief |
Risks to this view
• R-LTC's benefits are company-described. There is no independent measurement in the public record yet.
• Production use, customers and timing are undisclosed. A conference paper is not a product ramp.
• A laser-transparent stage plus head control adds process complexity. Throughput has not been published.
• Hybrid bonding, panel-level substrates and new materials may solve the same problem by other routes.
Package size is now a first-order variable for AI hardware, and warpage is the tax on size. The bonding methods that keep that tax low deserve as much attention as the interposers they serve.
Sources: Amkor IMAPS Symposium 2026 event page (accessed Oct 2, 2026); Semiconductor Engineering, "Mitigating Warpage In Multi-Chiplet Systems" (Sep 22, 2025); Sarr et al., "Laser Assisted Bonding for Flip Chip Interconnection of Very Large Chips," IMAPS Symposium 2024; TSMC 2026 North America Technology Symposium release (Apr 2026); Amkor Q2 2026 results, SEC 8-K Exhibit 99.1 (Jul 27, 2026). Everything here is from public sources.
Comments
Post a Comment