Glass Substrate Partnership

Design Tools for Glass Substrate Adoption

Absolics is shipping glass core substrates to customers today. But your customers lack the design tools to use them. Genesis Glass PDK fills that gap.

The Glass Substrate Design Gap

The Problem

Your customers (AMD, Intel, NVIDIA) have decades of silicon/organic PDKs. They have zero glass substrate design tools. Every customer evaluation stalls at 'how do I design on glass?'

The Solution

Genesis Glass PDK: 29 solvers, 15.92M+ long-format ML rows, 2.95M live BEM rows, 901K separately versioned strict buyer-regime additions, and 4 substrate classes + 21 materials. From GDS layout to full impedance + yield + SI report in one pipeline. Your customers get a design kit bundled with every glass substrate sample.

The Impact

Faster customer evaluations → faster design wins → faster volume ramp. Every $1 spent on design enablement accelerates $100 of substrate revenue.

What Genesis Glass PDK Provides

TGV Impedance Extraction

Multiconductor BEM solver calibrated to 3.57% MAE vs IEEE. Covers all Absolics glass types. 1,000x speedup via ML surrogate.

GDS-to-Yield Pipeline

Upload customer GDS → get impedance, S-parameters, eye diagram, yield prediction, and cost optimization in one report.

Signal Integrity Analysis

Eye diagram at 28-224 Gbps, differential pair extraction, PDN impedance profiling. Cadence/HFSS/SPICE export.

Independent FEM Validation

BEM predictions cross-checked against IEEE literature, FastHenry2 inductance, and AWS Palace FEM. Meep FDTD output is preserved as a retracted audit artifact, not a current convergence claim.

Partnership Model

Option A: Bundle with substrate samples. Every customer who receives Absolics glass substrate samples also gets access to Genesis Glass PDK. This accelerates their evaluation from months to weeks.

Option B: Co-develop glass-specific PDK. We customize the PDK for Absolics-specific glass compositions, adding your proprietary material data to the impedance database. Your customers get the most accurate design tools in the industry.

Option C: License for internal R&D. Your process engineers use Genesis to optimize TGV geometry, predict yield, and design golden kit reference structures for customer evaluation kits.

Productization fleet (5 net-new endpoints)
From "predicts impedance" to "ships fabricable design"
POST /v1/glass-pdk/geometry-pareto — multi-objective Pareto front (Z₀ + IL + crosstalk + yield)
POST /v1/glass-pdk/drc-validate — glass-interposer DRC with IPC/SEMI defect codes
POST /v1/glass-pdk/validate-against-measurement — surrogate-vs-literature with honest sim-vs-VNA flag
POST /v1/coupons/export-fab — fab-ready bundle (GDS + 12-layer stack-up + per-foundry SOW + $55K-$165K Amkor cost band)
GET /v1/glass-pdk/cross-solver-matrix — 100-geom × 5-solver disagreement witness
All 5 endpoints have 32/32 pytest contract test coverage + buyer adversarial harness coverage. Verify yourself in 5 min: bash scripts/audit/buyer_verify.sh.
Per-regime cross-physics cal heads · v2-clean retrain (2026-05-05)
6/6 regimes deployed · ULTRA_HIGH_FREQ 90.5% · HBM4 90.3%
First successful cross-physics calibration on Genesis. Per-regime BEM-vs-Palace μ-correction across all 6 production regimes. Pooled 17.10% → 4.71% (72.5% pooled gap closure, avg 74.0% across regimes, 4/6 elite >80%): ULTRA_HIGH_FREQ 90.5% · HBM4 90.3% · WIDE_PITCH 87.6% · UCIE 81.6% · EXTREME_TIGHT 50.6% · MMWAVE 43.4% gap closure. Pattern reproduces across 6/6 regimes — not a single-experiment artifact. σ-correction head + μ-correction heads on the BEM quasi-static envelope. Production OFF by default; activate via CHIPLETOS_PALACE_RESIDUAL_HEAD=1.
Track A · DD-grade coverage
95% conformal interval + per-axis OOD on every prediction
Every /v1/glass-pdk/predict-impedance response carries a distribution-free 95% Palace-truth coverage interval — HBM4 q=8.38 Ω, cov 94.46%; UCIE q=6.03 Ω, cov 95.41%; EXTREME_TIGHT q=35.10 Ω, cov 94.42%; WIDE_PITCH q=9.84 Ω, cov 94.44%; pooled fallback q=9.11 Ω — AND a per-axis OOD severity diagnostic across all 7 input axes (in/borderline/out per d/p/t/Dk/Df/freq/wall). For an Absolics glass-substrate handoff, this means every Z₀ promise on a customer datasheet has a board-defensible mathematical bound, not a marketing approximation.
Track B · OSS distribution + agent ecosystem
30 MCP tools · 10 agents · open-source diff-pair scaling law
Apache-2.0 OSS slice pip install glass-tgv-diffpair publishes the universal scaling law log(Z₀_diff)=0.338·log(sep/d), R²=0.918, n=2.1M across 5 commercial glasses (Eagle XG, AF32, Borofloat33, EN-A1, Fused Silica). Companion DAC 2026 paper drafted. MCP surface 11 → 30 wires the full endpoint set into Claude Desktop / Cursor / Codex agents (HBM4 Signoff / Inverse Design / Coupon RFQ / DRC Fixer / Cross-Solver Verifier / Pareto Explorer / Yield Risk / Interface Signoff / Provenance Auditor / Glass PDK Assistant). Controlled-transparency open/closed split (closed: trained weights, cal heads, conformal quantiles, MNDA glass corpora, fab-coupon export).
Composite lab_readiness_score
One 0-100 score on every panel-level Z₀ response
Single 0-100 buyer-facing score with verdict bands (send_to_lab ≥95 / send_with_extra_qc 80-94 / hold 60-79 / reject <60) on /predict-impedance + /geometry-from-target + /coupons/export-fab. Pro-rata weights across 5 active confidence gates (cross_solver 24 + conformal 24 + per_axis_ood 18 + public_data 18 + ensemble 16 = 100). For Absolics panel-level workflows, the score gates whether a CMP-recess + overlay σ uncertainty band on a panel Z₀ claim is fab-ready or whether more cross-physics evidence is needed before the customer ships a coupon. No fabricated proxy values: any gate without measured signals contributes None and triggers partial_score=true; fab-coupon export refuses bundles when partial.
3/5 cross-solver matrix wired
BEM ✓ Palace ✓ FastHenry2 ✓ · OpenEMS / gprMax honest skip
100-geometry × N-solver disagreement matrix at 3 of 5 wired (+ 4/5 spot-checks at 1-geom and 10-geom; witness benchmarks/sprint42/cross_solver_matrix_{1geom,10geom}_4of5_2026_05_01.json). BEM always-available + Palace via spack-installed binary + FastHenry2 via local fasthenry binary (Z₀ = √(L_FH / C_analytical_coax) honestly disclosed as hybrid; dispatch key = FastHenry2_L_plus_analytical_C). OpenEMS + gprMax return None via try-import skip when binary not on PATH — never fabricates a number. Witness: docs/HONEST_FALLBACKS.md Fallback 5.
Request Technical EvaluationView Benchmarks