The Glass Ceiling of Interconnects: Modeling Signal Skew in TGV-Based 2nm Heterogeneous Systems

By Rizowan Ahmed (@riz1raj)
Senior Technology Analyst | Covering Enterprise IT, Hardware & Emerging Trends

If you are still banking on organic build-up films (ABF) to carry the I/O load for 2026-era chiplet designs, the industry's collective pivot toward glass substrates suggests a significant shift in packaging strategy. This transition is a response to the thermal and mechanical limitations of polymers at advanced nodes. The adoption of Through-Glass Vias (TGV) introduces complex signal integrity (SI) challenges that modern models must address, specifically deterministic skew induced by via-to-via pitch scaling.

The Transition from Organic Substrates to Glass

The limitations of organic substrates have become a primary bottleneck for Heterogeneously Integrated 2nm APUs. The Coefficient of Thermal Expansion (CTE) mismatch between silicon (approximately 2.6 ppm/°C) and organic materials (roughly 17 ppm/°C) can lead to warping during the reflow process of high-density chiplet bonding. Glass, including aluminosilicate and fused silica variants, offers a CTE closer to 3.0–8.0 ppm/°C, providing the structural rigidity required for next-generation UCIe interconnects.

While glass eliminates the "glass weave effect" found in FR4, it introduces new variables. In systems targeting low end-to-end latency for real-time AI inference, jitter management is critical. When scaling TGV pitches to accommodate the high I/O count of 2nm compute dies, the electromagnetic coupling between vias becomes a significant source of signal degradation.

Modeling the TGV Skew: Beyond the Quasi-Static Assumption

Traditional modeling techniques often rely on quasi-static approximations that are insufficient at the high Nyquist frequencies required for 112G and 224G SerDes. To accurately model signal skew in glass-based Through-Glass Vias, engineers utilize full-wave 3D electromagnetic solvers such as Ansys HFSS or Cadence Sigrity Aurora.

Key Parameters for Skew Modeling:

• Dielectric Constant (Dk) Stability: Glass offers a stable Dk (typically 3.8 to 5.2) across a wide frequency range, providing more consistency than many organic resins.
• TGV Sidewall Roughness: The laser-induced etching process for TGVs creates micro-scale roughness. At high frequencies, the skin effect forces current to the periphery, where this roughness increases effective resistance.
• Pitch-Induced Crosstalk: As TGV pitch scales down, the fringe fields from adjacent signal vias penetrate the glass substrate, causing phase shifts that manifest as skew in differential pairs.
• Via Tapering: TGVs often exhibit a slight hourglass or conical shape due to the laser-drilling and wet-etching process. This impedance discontinuity is a driver of reflection-induced skew.

The 2nm APU Bottleneck: Latency Sensitivity

System-level latency targets are highly sensitive to skew in the substrate within heterogeneous integration architectures. In advanced APUs, data movement between HBM stacks, NPUs, and I/O dies must be precisely orchestrated. If the UCIe or BoW (Bunch of Wires) interface experiences skew-induced bit errors, the resulting Retransmission (ARQ) cycles at the Link Layer can impact the latency budget.

Benchmarks in glass substrate signal integrity indicate that when TGV pitch is reduced without compensatory shielding, differential skew can impact the eye diagram at 112Gbps per lane. Proper modeling and shielding are required to maintain signal margins.

Advanced Mitigation: Differential Pair Balancing and Shielding

To combat skew in high-density glass environments, the industry is moving toward Ground-Signal-Signal-Ground (GSSG) via patterns. While this consumes real estate, it is a standard method for protecting signal integrity. Furthermore, hardware designers must balance the use of clock recovery circuits (CDRs) against the power penalties associated with masking substrate-level design deficiencies.

Quantifying the Impact of TGV Geometry

Mathematical rigor in modeling must account for the Dissipation Factor (Df) of the glass. While glass generally has a lower Df than organic materials (typically <0.002), high-density TGV arrays create parasitic capacitance that shifts the resonant frequency of the interconnect. Modeling tools must account for the Maxwell-Wagner effect at the interface of the copper via and the glass dielectric to ensure accurate skew predictions.

Technical Specification Comparison:

• Substrate Material: High-Purity Fused Silica (HPFS) vs. Advanced ABF.
• TGV Aspect Ratio: 10:1 (Current) to 20:1 (Projected).
• Target Reliability: High-performance computing standards for post-FEC Bit Error Rate (BER).

The Reality of 2nm Integration

The move to 2nm involves a significant packaging evolution. The interconnect density required for 2nm compute engines demands TGV pitches that push the limits of laser precision. Alignment is critical, as even minor deviations can lead to common-mode noise. Consequently, SI-aware routing has become a standard requirement in EDA tools.

Thermal management of glass-based APUs is also a factor. While glass is structurally stable, its thermal conductivity is lower than silicon, which can create localized thermal gradients. Because Dk can be temperature-dependent, these gradients may translate into signal skew, necessitating electro-thermal aware modeling.

Industry Outlook

The industry is moving toward the commercialization of 2nm APUs utilizing glass substrates for the high-performance computing (HPC) market. Success in this transition depends on mastering TGV pitch scaling while maintaining a controlled skew profile across the interconnect fabric. We anticipate a shift toward Hybrid Bonding on Glass, where the TGV becomes part of a complex, multi-level 3D interconnect structure. Integrating full-wave EM solvers into the design workflow is essential for managing the physics of glass-based substrates.

---