Three Months of a Wafer
TSMC to UNAM to packaging — the end-to-end production timeline.
Prerequisites
What you'll learn here
- Walk through a SIDRA Y1 wafer's full production timeline day-by-day
- Explain the logistics chain (TSMC-UNAM-Taiwan)
- List the wafer-by-wafer process steps
- Identify time-critical steps (lithography, CMP, test)
- State production-acceleration strategies
Hook: 3 Months from Order to Package
A SIDRA Y1 chip:
- Weeks 1-4: TSMC 28 nm CMOS substrate.
- Weeks 5-7: UNAM workshop BEOL HfO₂ + memristors.
- Weeks 8-9: Wafer test + dicing.
- Weeks 10-11: Packaging (Taiwan).
- Week 12: Final test + shipment.
Total: 12 weeks = 3 months. Each wafer 38 dies = 27 final chips (after yield).
This chapter details every stage.
Intuition: Multi-Site, Single Flow
TSMC Taiwan: 28 nm CMOS substrate
↓ flight (4-week process + 1-week ship)
UNAM Bilkent: BEOL memristor
↓ in-house (3 weeks)
ASELSAN test: wafer test
↓ flight (1 week)
Amkor Taiwan: packaging
↓ flight (2 weeks)
SIDRA test lab: final test + shipment
↓ Türkiye + worldMulti-site logistics is complex. Y10+ goal: all-domestic (mini-fab in Türkiye).
Formalism: Wafer Flow
Stage 1: TSMC 28 nm CMOS substrate (4 weeks).
Customer (SIDRA): tape-out file (GDS-II) → TSMC → wafer fab.
- 200 mm wafer, 38-die layout.
- 28 nm CMOS substrate: transistors, M1-M2 metal.
- Multi-Project Wafer (MPW): SIDRA shares wafer with other customers → cheap.
Output: “bare-CMOS” wafer, particle-clean.
Cost: $1000-5000/wafer (MPW shared).
Stage 2: UNAM BEOL (3 weeks).
Wafer arrives at UNAM (air cargo, special crate). At the workshop:
Week 5:
- Wafer intake test + clean.
- M3 metal layer.
- Memristor BEOL: ALD HfO₂ (2 µm active area).
Week 6:
- DUV lithography (memristor pattern).
- ICP-RIE etch.
- CMP planarization.
Week 7:
- M4 metal (BL).
- M5-M20 upper metal layers.
- Pad opening.
Output: full SIDRA Y1 wafer (substrate + BEOL).
Stage 3: Wafer test (1 week).
Probe-card test on every die (chapter 6.9):
- DC test.
- Crossbar reference.
- Memory test.
Failed dies marked. Yield: 75% → 28 good dies/wafer.
ASELSAN’s test lab can be used (in Türkiye).
Stage 4: Dicing (2 days).
Wafer cut → individual dies. Diamond saw or laser. UNAM can do it.
Stage 5: Packaging (2 weeks).
Dies sent to Amkor/ASE in Taiwan. FC-BGA (chapter 5.13):
- Wafer bumping.
- Flip-chip die-to-substrate.
- Underfill.
- Heat spreader.
Yield: 95% → 26-27 packaged chips/wafer.
Stage 6: Final test (1 week).
Each packaged chip system test:
- PCIe enumeration.
- Full crossbar test.
- Thermal stress.
- AI benchmark (MNIST).
Yield: 95% → 25 ship-ready chips/wafer.
Shipment: crate + documentation + certification.
Time-critical steps:
| Step | Time | Bottleneck |
|---|---|---|
| TSMC | 4 weeks | Supplier schedule (3-6 months queue) |
| UNAM ALD | 2 hours × 3 layers | Reactor count |
| Lithography | 1 hour × 4 masks | DUV slow, single wafer |
| Etch | 1 hour × 4 layers | RIE chamber |
| CMP | 1 hour × 5 layers | CMP station |
| Test | 30 s × 38 dies = 20 min | Probe card speed |
| Packaging | 2 weeks | Taiwan logistics |
Single-wafer total: 3 months. Parallel work-in-progress (WIP) means daily output is 1-2 wafers.
Acceleration:
- 2× ALD reactors → ALD time halved.
- Domestic packaging (mini-fab in Türkiye) → save 2-week logistics.
- Domestic TSMC alternative (none today) → 4 weeks → in-house.
Y10 mini-fab cuts the whole cycle to 6 weeks (2× faster).
WIP management:
Workshop runs continuously. Each week 5 wafers in (from TSMC), 5 wafers out (shipping). 12-week look-ahead.
Pipeline: 60 wafers in-progress at any time. Each at a different stage.
Production planning:
class WaferPlanner:
def schedule(self, target_chips):
wafers_needed = target_chips / 25 # final yield
# TSMC order
order_tsmc(wafers_needed, lead_time_weeks=4+queue_weeks)
# UNAM capacity reserve
reserve_workshop(wafers_needed, weeks=3)
# Packaging order
order_amkor(wafers_needed, lead_time_weeks=2)Cost summary:
Per chip:
- TSMC wafer share: $40 (1000/25).
- UNAM BEOL: $20.
- Test: $10.
- Packaging: $30.
- Logistics: $5.
- Margin: $50.
- Customer price: ~$155-250.
Experiment: 100K-Chip Production Plan
Target: 100K chips in 2027.
Math:
- 100K / 25 = 4000 wafers.
- 4000 wafers / 50 weeks = 80 wafers/week.
Capacity:
- UNAM 5 wafers/week → 16× growth needed.
- Or 16 parallel workshops (impractical).
- Mini-fab needed (chapter 7.5).
Logistics:
- TSMC order: 4000 wafers/year = 80 wafers/week within TSMC schedule.
- Cost: 4000 × 4M/year wafer purchase.
- Packaging: 4000 × 120K/year. Taiwan contract.
Personnel:
- UNAM 14× growth: 14 × 15 = 210 engineers. Too large.
- Mini-fab: 50 engineers + 100 technicians. More efficient.
Conclusion: UNAM doesn’t suffice for 100K/year Y1 production. Mini-fab transition mandatory 2027-2028.
Quick Quiz
Lab Exercise
Y1 production-speed analysis.
(a) Single wafer 12 weeks. Daily output via pipeline?
- 1-2 wafers/day (parallel WIP).
(b) Annual output 7000 chips → cost per chip?
- 250 sale → $150 margin.
(c) Investment for 100K chips?
- Mini-fab 20M/year.
(d) Payback time?
- 100K × 15M/year. Mini-fab 15M = 4.7 years.
(e) Feasible for Türkiye?
- 5-year ROI is fine. Strategic (semiconductor sovereignty) bonus.
Cheat Sheet
- 3 months/wafer: TSMC + UNAM + test + packaging.
- TSMC: 4 weeks CMOS substrate.
- UNAM: 3 weeks BEOL memristor.
- Test + packaging: 5 weeks.
- Output/wafer: 25 ship-ready chips (yield 67%).
- Y1 workshop: 7K chips/year.
- Y10 mini-fab: 1M chips/year.
Vision: Single-Site Production
- Y1: Multi-site (TSMC + UNAM + Taiwan).
- Y3: Domestic packaging (Türkiye).
- Y10: Mini-fab Türkiye = single site.
- Y100: Full fab Türkiye.
- Y1000: Wafer-scale, single wafer = single chip.
Further Reading
- Next chapter: 7.3 — TSMC 28nm MPW Process
- Previous: 7.1 — Cleanrooms and ISO Classes
- Wafer fab flow: Plummer, Deal, Griffin, Silicon VLSI Technology.
- MPW: TSMC tape-out services.