AI & Computingpreprint2026-09-13

roject JANUS: A Spatial Residue Number System Photonic AI Architecture with Non-Volatile Phase-Change Routing and Monolithic 3D Multi-Physics Simulation Suite (Release 4.0)

Open access0 citations

Abstract

## OverviewProject JANUS is a constraint-aware, bounded-exact optoelectronic computing architecture designed to eliminate the analog optical bottleneck by replacing analog amplitude accumulation with deterministic Spatial One-Hot Residue Number System (RNS) permutations over Z₁₇ Fermat rings. By routing single-wavelength optical pulses through 4-stage asymmetric binary tree networks of non-volatile Sb₂S₃ phase-change switches with zero static hold power, JANUS scales from low-power INT4 neural inference up to Exact INT64 deterministic precision at 100 GHz wave-pipelining with zero static hold power. --- ## What's New in Release 4.0 (Tapeout & 16-Tree Integration Milestone):- Asymmetric 16-Tree Fermat Core Migration: Formally unified all optical routing and formal proofs to the 4-stage Asymmetric 16-Tree topology (16 active waveguides + 1 dark zero channel), reducing insertion loss to 1.61 dB and switch counts to 240 per multiplier.- Synthesized & DRC-Verified GDS II Silicon Tapeout: Complete multi-stratum physical layout for the JANUS Mini-16 accelerator (3.20 mm × 3.20 mm): • Segmented, collision-free CMOS mask layers (Range 100–199) and optical layers (Range 1–29). • 4-level balanced binary H-tree clock distribution network (Layer 161/0, Metal 6) to eliminate intra-tile clock skew across all 16 heterogeneous tiles. • 4-layer concentric moisture barrier chip seal ring (Layer 190/0). • 4-sided wire-bond I/O pad frame (Layer 182/0, 120 µm pitch). • Through-Dielectric Via (Cu TDV) vertical interconnect arrays (8 µm diameter, Layer 30/0).- Expanded Primary Treatise (39-Page IEEEtran Manuscript): Incorporates comprehensive derivations, Galois field permutation proofs, 16 formal verification theorems, and full system benchmarks vs. NVIDIA H100 SXM5.- Updated 9-Page IEEE Co-Simulation Report: Includes high-resolution 300 DPI GDS II layout micrographs and closed-loop multi-physics validation across Maxwell FDTD (Meep), thermo-mechanical FEM (Elmer), transient circuit SPICE (Xyce), digital RTL (Icarus Verilog), and formal Python arithmetic proofs. --- ## Included Documents in Release 4.0:1. JANUS_IEEE_Manuscript.pdf — Primary Foundational Architecture Treatise (39 Pages, IEEEtran).2. JANUS_Mini16_Simulation_Report.pdf — 5-Tier Multi-Physics Co-Simulation & Physical GDS II Report (9 Pages, IEEEtran).3. JANUS_Mini16_CMOS_Architecture.pdf — 65nm LP/GP CMOS Digital Base Die & Mixed-Signal Micro-Architecture Spec. --- ## Intellectual Property & Patent Reference- Filing: Indian Patent Application No. 202611052791 (Patent Pending)- Title: A Spatial Residue Number System Photonic AI Architecture with Non-Volatile Phase-Change Routing and Monolithic 3D Heterogeneous Stacking- Inventor: Deepanshu Bhardwaj --- ## Open-Access Repositories & Interactive Web Platform- Interactive Web Platform & 3D WebGL Strata Model: https://janus-photonic-hardware.vercel.app- Hardware & GDS II Layout Repository: https://github.com/horizonseekerik/janus-photonic-hardware- 5-Tier Co-Simulation Suite Repository: https://github.com/horizonseekerik/janus-simulation

// Source

View paper (DOI)Open access versionOpenAlexZenodo (CERN European Organization for Nuclear Research)Published 2026-09-13

Authors: Daya Bhardwaj