Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2026
|
| Online Access: | https://doi.org/10.5281/zenodo.18913865 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866902228754759680 |
|---|---|
| author | Topuria, Nika |
| author_facet | Topuria, Nika |
| contents | <p>This submission presents the Gelron Architecture, a theoretical design study and preliminary engineering specification for a volumetric, solid-state optical Application-Specific Integrated Circuit (ASIC). Designed specifically for Ternary Large Language Model (LLM) inference under the BitNet b1.58 paradigm, the Gelron ASIC functions as a heterogeneous optical-electronic co-processor. It executes <span>$O(N^2)$</span> ternary matrix multiplications optically through a passive 3D-printed polymer mesh, while leveraging a bonded Silicon CMOS interposer for <span>$O(N)$</span> operations, KV-caching, and signal phase-resets. The specification resolves historic optomechanical limitations of 3D optical computing by introducing suspended isotropic waveguides for polarization integrity, in-situ holographic calibration, volumetric EDFA gain-loss reconciliation, and a Wavelength-Division Multiplexed (WDM) throughput model. The resulting architecture offers a physically manufacturable pathway to ultra-low-energy, sub-microsecond AI acceleration.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18913865 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Volumetric Optical ASICs for AI Inference: The Gelron Architecture Topuria, Nika <p>This submission presents the Gelron Architecture, a theoretical design study and preliminary engineering specification for a volumetric, solid-state optical Application-Specific Integrated Circuit (ASIC). Designed specifically for Ternary Large Language Model (LLM) inference under the BitNet b1.58 paradigm, the Gelron ASIC functions as a heterogeneous optical-electronic co-processor. It executes <span>$O(N^2)$</span> ternary matrix multiplications optically through a passive 3D-printed polymer mesh, while leveraging a bonded Silicon CMOS interposer for <span>$O(N)$</span> operations, KV-caching, and signal phase-resets. The specification resolves historic optomechanical limitations of 3D optical computing by introducing suspended isotropic waveguides for polarization integrity, in-situ holographic calibration, volumetric EDFA gain-loss reconciliation, and a Wavelength-Division Multiplexed (WDM) throughput model. The resulting architecture offers a physically manufacturable pathway to ultra-low-energy, sub-microsecond AI acceleration.</p> |
| title | Volumetric Optical ASICs for AI Inference: The Gelron Architecture |
| url | https://doi.org/10.5281/zenodo.18913865 |