tsmc cowos and advanced packaging technologies
on this page
packaging landscape
market dynamics: cowos capacity constrained through 2025, prices +10-20% 1
technology evolution: 2.5d cowos → 3d soic → 3.5d xdsip progression
silicon size: broadcom xdsip enables 6,000mm² packages with 12 hbm stacks 2
osat partnerships: ase handles 40-50% of outsourced cowos-s 1
ai demand driver: nvidia h100/h200, amd mi300, custom hyperscaler chips
2026 capacity: tsmc targeting 2x cowos output, new fabs coming online
packaging constrains ai chip production. cowos and similar technologies enable integration of logic, memory, and interconnects in ai accelerators.
cowos technology overview
architecture
cowos (chip on wafer on substrate) - tsmc’s 2.5d packaging platform:3
tl;dr
| component | function | specifications |
|---|---|---|
| silicon interposer | interconnect platform | up to 2,500mm², 65nm process |
| through silicon vias (tsv) | vertical connections | 10μm diameter, 100μm pitch |
| micro-bumps | die-to-interposer bonds | 40μm pitch, copper pillar |
| c4 bumps | package-to-substrate | 150μm pitch |
| redistribution layers | signal routing | 4-6 layers, sub-2μm l/s |
cowos variants and evolution
tsmc cowos configurations:3
- cowos-s: standard silicon interposer, 2,500mm² maximum
- cowos-l: local silicon interconnect (lsi) bridges, reduced cost
- cowos-r: organic interposer with embedded bridges
- soic: 3d chip stacking, direct cu-cu bonding, sub-1μm pitch
performance characteristics
cowos benefits
bandwidth: 8.5 tb/s between logic and hbm (8-stack configuration)
power efficiency: 50% lower than traditional packaging at same bandwidth
signal integrity: shortened interconnects reduce latency 10x
thermal management: integrated cooling solutions for 700w+ tdp
yield optimization: chiplet approach improves overall yield vs monolithic
broadcom’s 3.5d xdsip innovation
technology details
broadcom’s extreme dimension system in package specifications:2
- hybrid architecture: 2.5d interposer with 3d die stacking
- face-to-face bonding: direct cu-cu connections, sub-micron pitch
- scale: 6,000mm² total silicon area
- hbm integration: 12 hbm stacks per package
- production: early 2026
xdsip vs traditional cowos
tl;dr
| metric | cowos-s | 3.5d xdsip | improvement |
|---|---|---|---|
| max silicon area | 2,500mm² | 6,000mm² | 2.4x |
| hbm stacks | 8 | 12 | 1.5x |
| interconnect density | 40μm pitch | sub-1μm f2f | 40x |
| vertical integration | single layer | multi-layer 3d | new capability |
| power delivery | traditional | backside power | 20% efficiency gain |
applications
xdsip applications:
- hyperscale ai training chips (>1000w tdp)
- multi-die cpu/gpu complexes
- disaggregated architectures
- openai custom inference chip4
manufacturing process flow
cowos production steps
production sequence
step 1: fabricate silicon interposer wafer (65nm process)
step 2: create through silicon vias (tsv) and redistribution layers
step 3: thin interposer wafer to 100μm
step 4: mount logic dies and hbm stacks via micro-bumps
step 5: underfill and molding compound application
step 6: attach to organic substrate with c4 bumps
step 7: final assembly and test
cycle time: 8-12 weeks
quality and yield
manufacturing considerations:
- warpage control: stress management across packages
- alignment precision: sub-micron accuracy
- thermal cycling: 700w+ heat load reliability
- known good die: pre-assembly testing
- defect density: single defect destroys $50,000+ package
yield: 60-80% for mature cowos-s
supply chain and capacity
tsmc capacity
cowos capacity:1
tl;dr
| facility | current capacity | 2026 target | investment |
|---|---|---|---|
| taiwan ap6 | 15k wafers/month | 30k wafers/month | $2.8b |
| taiwan ap2 | 8k wafers/month | 12k wafers/month | $900m |
| japan atp | 0 | 5k wafers/month | $1.6b |
| total | 23k wafers/month | 47k wafers/month | $5.3b |
demand exceeds supply through 2026.
osat partnership model
tsmc osat partners:1
- ase/spil: 40-50% of cowos-s backend, nt$4.1b expansion
- amkor: 70,000-80,000 units/year, arizona facility
- jiangsu changjiang electronics: china capacity
- powertech technology: memory integration
tsmc focuses on silicon steps, osats handle assembly.
cost structure
cowos packaging economics:
- interposer cost: $300-500 per 2,500mm² wafer
- assembly cost: $200-400 per package
- test cost: $100-200 per unit
- total cowos cost: $600-1,100 per package
- percentage of chip cost: 10-15% for high-end ai chips
pricing: +10-20% increases through 20251
competitive landscape
alternative packaging
tl;dr
| company | technology | status | key customers |
|---|---|---|---|
| intel | emib, foveros | production | intel, limited external |
| samsung | i-cube, x-cube | ramping | samsung, select customers |
| amd/xilinx | 3d chiplet | production | amd internal |
| ase | focos, vipkg | production | multiple idms |
tsmc: 80%+ high-end packaging market share.
roadmap
2025-2026 developments
- cowos-xl: 3,000mm² interposers
- soic-x: enhanced 3d stacking
- info_b: bridge technology for mobile
2027-2028 horizon
- optical interposers: photonic interconnects
- 4,000mm² interposers: further scaling
- system-in-wafer: full wafer-scale integration
beyond 2028
- quantum packaging: cryogenic compatibility
- bio-compatible packages: implantable chips
- self-assembling systems: automated chiplet integration
applications and use cases
production deployments
cowos applications
nvidia h100/h200: 8 hbm3/3e stacks, 700w tdp
amd mi300x: 8 hbm3 stacks, chiplet architecture
nvidia gb200: dual-chip design, 12 hbm3e stacks
apple m3 ultra: dual m3 max dies, unified memory
google tpu v5: custom design with broadcom
new applications
applications:
- openai inference chip: 3.5d xdsip, 12 hbm stacks4
- meta training clusters: custom interconnects
- quantum-classical hybrids: mixed temperature packaging
- edge ai servers: cowos-l/r variants
- autonomous vehicle compute: automotive-grade
challenges and limitations
technical limits
technical limitations:
- reticle size: 858mm² maximum per exposure
- yield degradation: exponential with package size
- thermal density: 3d stack heat concentration
- signal integrity: cross-package high-speed signals
- mechanical stress: cte mismatches
economic limits
economic constraints:
- nre costs: $10-50m custom design
- development cycles: 12-18 months
- limited second sources
- customer concentration
- expansion capital requirements
supply chain risks
supply chain dependencies:
- abf substrate: shortage beyond silicon
- underfill materials: specialized chemistry, few suppliers
- test equipment: limited for large packages
- skilled workforce: specialized expertise required
- geographic concentration: taiwan risk
future outlook
market outlook
packaging market:5
- 2024: $44 billion
- 2025: $52 billion (+18%)
- 2026: $61 billion (+17%)
- 2030: $120 billion projection
ai accelerators: 60%+ of high-end demand.
technology trends
technology trends:
- chiplet standardization: ucie consortium
- heterogeneous integration: mixed process nodes
- in-package optics: silicon photonics
- advanced cooling: embedded microfluidics
- ai-optimized design: ml-driven optimization
implications
semiconductor implications:
- packaging equals process node importance
- design-foundry-osat integration
- chiplet ecosystem business models
- monolithic to disaggregated shift
- packaging ip differentiation
references
[3] tsmc. (2025). cowos - taiwan semiconductor manufacturing company limited.
[5] idtechex. (2025). advanced semiconductor packaging 2025-2035: forecasts, technologies, applications.