TSMC Reports 98% CoWoS Yield at 5.5x Reticle
TSMC reported at the OCP APAC Summit on August 11 that its 5.5x-reticle CoWoS advanced-packaging products are in volume production with yields consistently above 98% across multiple AI customer products, according to TrendForce. Vice President Jun He also outlined a roadmap toward CoWoS packages beyond 14x reticle size in 2029, while identifying memory and ABF substrates as supply-chain constraints.
TSMC reported that its 5.5x-reticle CoWoS advanced-packaging products have entered volume production with yields consistently above 98% across multiple AI customer products. According to TrendForce's coverage of the August 11 OCP APAC Summit, some products have reached 99% yield.
The figures were presented by Jun He, TSMC vice president of Advanced Packaging Technology and Services. XenoSpectrum, citing on-site reporting from Economic Daily News, cautions that the common benchmark is stable yield above 98%, while 99% is a peak reported for some products. TSMC has not publicly disclosed the calculation denominator or the specific manufacturing stage represented by those yield figures, so they should not be read as complete-system pass rates.
Larger CoWoS packages on the roadmap
CoWoS, or Chip-on-Wafer-on-Substrate, is TSMC's family of advanced-packaging technologies for integrating large logic dies and high-bandwidth memory in a single package. The 5.5x designation refers to package scale relative to the lithography reticle limit, the maximum area exposed in one lithographic step.
TrendForce reports that TSMC expects a 14x-reticle CoWoS package to enter production in 2028, followed by packages beyond 14x in 2029. The same report says a 14x package could integrate roughly 10 large compute dies and 20 HBM stacks. TSMC's North America Technology Symposium roadmap, reported by Tom's Hardware in April, described packages of 14 reticles and above, with up to 24 HBM5E stacks, by 2029.
The differing publication dates and roadmap descriptions leave the precise 2028-2029 sequencing unclear. The common point in the reporting is that TSMC is extending package-scale integration far beyond earlier CoWoS roadmaps that topped out at 9.5 reticles, according to Tom's Hardware.
For accelerator designers, increasing package area can create room for more compute chiplets and HBM stacks without requiring one monolithic die beyond reticle limits. Industry work on comparable multi-die systems, however, treats package yield as only one production variable: memory availability, substrate supply, thermal behavior, power delivery, test coverage, and system-level reliability can each constrain shipment volumes.
Capacity and material constraints remain relevant
He said TSMC had nearly doubled CoWoS capacity annually over the previous three years and operated 10 advanced-packaging facilities, according to TrendForce. He added, "We haven't fully met demand yet, but we're getting very close."
He also identified memory and Ajinomoto build-up film, or ABF, substrates as AI supply-chain bottlenecks, TrendForce reported. That distinction matters because a mature interposer and assembly flow does not remove constraints in the surrounding bill of materials. HBM stack availability and substrate lead times can limit final accelerator output even when packaging yields improve.
TSMC also discussed its System-on-Integrated-Chips, or SoIC, hybrid-bonding roadmap. TrendForce, citing TechNews, reported that SoIC pitch reached 6 microns in 2025 and is targeted to narrow to 4.5 microns by 2029. The report also cited TSMC's previously stated availability target for A14-to-A14 SoIC production in 2029, with 1.8x higher die-to-die I/O density than N2-on-N2 SoIC.
The reported CoWoS yield is therefore a useful indicator of manufacturing maturity at a large package scale, but it is not a standalone measure of available AI compute. For ML infrastructure teams, the near-term relevance remains tied to accelerator supply, memory configurations, and the packaging choices made by GPU and custom-silicon vendors.
Key Points
- 1TSMC reported stable CoWoS yields above 98% at 5.5x reticle scale, indicating volume-production maturity for very large AI packages.
- 2Roadmap reporting places 14x and larger CoWoS packages in the 2028-2029 period, expanding potential compute-die and HBM integration.
- 3Higher packaging yield does not remove industry-wide constraints, as HBM supply and ABF substrates can still limit accelerator shipment volumes.
Scoring Rationale
The reported yield and scaling roadmap concern a core manufacturing bottleneck for AI accelerators, affecting the hardware supply available to model-training and inference operators. The story is highly relevant to infrastructure planning, although it is a packaging update rather than a new broadly deployable AI product or model.
Sources
Public references used for this report.
Practice with real Ad Tech data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Ad Tech problems

