Scaling Photonic Tensor Cores with Unary and Homodyne Designs

Ishan Thakkar; Oluwaseun Alo

arxiv: 2604.14664 · v1 · submitted 2026-04-16 · ⚛️ physics.optics · cs.AR

Scaling Photonic Tensor Cores with Unary and Homodyne Designs

Oluwaseun Alo , Ishan Thakkar This is my paper

Pith reviewed 2026-05-10 10:52 UTC · model grok-4.3

classification ⚛️ physics.optics cs.AR

keywords photonic tensor coresmicroring resonatorsunary encodinghomodyne accumulationoptical computingscalabilityparallelismoptical neural networks

0 comments

The pith

Unary encoding and homodyne accumulation offer the strongest path to higher parallelism in photonic microring tensor cores.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper compares five photonic microring tensor core designs through one shared optical power model. It identifies circuit ordering, unary encoding, and homodyne accumulation as the main factors that determine how far the cores can scale. Unary encoding and homodyne accumulation emerge as the more effective options for increasing the level of parallel operations. This matters for efforts to build larger optical systems that perform matrix multiplications efficiently, as photonic hardware could otherwise face limits from power and loss constraints. A sympathetic reader sees the work as identifying which design principles to favor when trying to expand these cores.

Core claim

Analysis of five photonic microring tensor core designs shows that circuit ordering, unary encoding, and homodyne accumulation shape scalability, with unary encoding and homodyne accumulation providing the strongest improvements to parallelism.

What carries the argument

Unary encoding and homodyne accumulation mechanisms evaluated within the common optical power model for the five microring designs.

If this is right

Unary encoding supports more parallel multiplications under fixed optical power limits.
Homodyne accumulation reduces losses relative to other methods and permits larger core sizes.
Circuit ordering influences scalability but less strongly than the encoding and accumulation choices.
Together these elements allow more efficient photonic execution of tensor operations at higher parallelism.
The results indicate which architectural features should be prioritized to increase the size of optical tensor processors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Future photonic designs could embed unary encoding and homodyne accumulation to push past present integration limits.
The same design principles may transfer to other optical computing platforms that face similar power and loss trade-offs.
Prototype fabrication and testing under real noise conditions would provide a direct check on the model's accuracy.

Load-bearing premise

The common optical power model accurately captures performance and losses across all five designs without unmodeled noise, crosstalk, or fabrication variations.

What would settle it

Experimental data from a fabricated unary homodyne microring tensor core showing power consumption or parallelism levels that differ markedly from the model's predictions.

Figures

Figures reproduced from arXiv: 2604.14664 by Ishan Thakkar, Oluwaseun Alo.

read the original abstract

We analyze five photonic microring tensor core designs with a common optical power model. The results show that circuit ordering, unary encoding, and homodyne accumulation shape scalability, with the last two offering the strongest path to higher parallelism.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Unary encoding and homodyne accumulation look like the stronger options for parallelism in these microring designs, but only under a simplified power model that may miss real photonic noise and crosstalk.

read the letter

The main thing to know is that the paper compares five microring photonic tensor core variants under one shared optical power model and concludes that unary encoding plus homodyne accumulation best support scaling parallelism, with circuit ordering as a secondary factor. This is a direct result of their side-by-side analysis rather than a restatement of prior work. They do a solid job keeping the model consistent across designs so the differences in encoding and accumulation can be isolated cleanly. That kind of controlled comparison is useful when designers are choosing between photonic accelerator options for AI workloads. The thinking is straightforward and stays grounded in the existing literature on microring-based tensor cores. The soft spot is the model itself. It treats optical power and losses in a common way but does not appear to include crosstalk between rings, thermal noise, or fabrication-induced resonance shifts. Those effects grow with more parallel elements, so the claimed advantages for unary and homodyne could shrink once real hardware constraints are added. No sensitivity checks or noise simulations are mentioned to test how robust the scalability numbers are. This paper is aimed at researchers modeling or building photonic hardware accelerators. It is not a paradigm shift but gives a practical ranking of design choices that could inform early decisions. The analysis is coherent on its own terms and shows honest engagement with the problem. I would send it to peer review so referees can check whether the power model needs extra terms or whether the parallelism limits hold up under more complete simulations.

Referee Report

1 major / 0 minor

Summary. The paper analyzes five photonic microring tensor core designs under a shared optical power model. It concludes that circuit ordering, unary encoding, and homodyne accumulation determine scalability limits, with unary encoding and homodyne accumulation providing the strongest route to higher parallelism.

Significance. If the common optical power model holds under realistic conditions, the comparative analysis offers practical guidance for scaling photonic tensor cores in optical neural network hardware. The identification of unary and homodyne approaches as particularly promising is a concrete design insight that could inform future device fabrication and system-level integration.

major comments (1)

The central results rest on the assumption that a single optical power model accurately captures losses, power scaling, and parallelism limits across all five designs without significant unmodeled effects. The skeptic note correctly flags that inter-ring crosstalk, thermal noise, and fabrication-induced resonance shifts are known to be important in microring systems and could erode the reported advantages of unary and homodyne designs at higher parallelism; the manuscript should include at least a sensitivity analysis or bounding argument showing that these effects do not reverse the ordering of the designs.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive comments and the recommendation for major revision. We address the single major comment below and will incorporate the requested analysis into the revised manuscript.

read point-by-point responses

Referee: The central results rest on the assumption that a single optical power model accurately captures losses, power scaling, and parallelism limits across all five designs without significant unmodeled effects. The skeptic note correctly flags that inter-ring crosstalk, thermal noise, and fabrication-induced resonance shifts are known to be important in microring systems and could erode the reported advantages of unary and homodyne designs at higher parallelism; the manuscript should include at least a sensitivity analysis or bounding argument showing that these effects do not reverse the ordering of the designs.

Authors: We agree that inter-ring crosstalk, thermal noise, and fabrication-induced resonance shifts are important practical effects in microring systems that are not explicitly modeled in our common optical power model. The model is deliberately scoped to enable a consistent, first-order comparison of power scaling and loss across the five designs, isolating the roles of circuit ordering, encoding, and accumulation. In the revised manuscript we will add a new subsection containing a sensitivity analysis and bounding argument. This will estimate the degradation in effective SNR due to these effects and demonstrate that the relative ordering of the designs (with unary encoding and homodyne accumulation remaining most scalable) holds for realistic parameter ranges drawn from the literature. We will also expand the existing skeptic note to cross-reference this new analysis. revision: yes

Circularity Check

0 steps flagged

No circularity: results follow from comparative analysis under shared model

full rationale

The paper performs a comparative analysis of five microring tensor core designs under a common optical power model, deriving scalability conclusions about circuit ordering, unary encoding, and homodyne accumulation directly from that model's calculations of power, losses, and parallelism limits. No step reduces a claimed prediction or first-principles result to its own inputs by construction, no fitted parameter is relabeled as a derivation, and no load-bearing uniqueness or ansatz is imported via self-citation. The central claims remain independent of the paper's own outputs and rest on external photonic design principles.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract mentions a common optical power model but provides no details on its parameters, assumptions, or any new entities; no free parameters, axioms, or invented entities can be identified from available text.

pith-pipeline@v0.9.0 · 5318 in / 924 out tokens · 38255 ms · 2026-05-10T10:52:17.371902+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

7 extracted references · 7 canonical work pages

[1]

HolyLight: A nanophotonic accelerator for deep learning in data centers,

W. Liuet al., “HolyLight: A nanophotonic accelerator for deep learning in data centers,”IEEE/ACM DATE, 2019

work page 2019
[2]

DEAP-CNN: Digital electronics and analog photonics for convolutional neural networks,

V . Bangariet al., “DEAP-CNN: Digital electronics and analog photonics for convolutional neural networks,”IEEE JSTQE, 2020

work page 2020
[3]

SCONNA: A stochastic computing based optical accelerator for ultra-fast, energy-efficient inference of integer-quantized CNNs,

S. S. Vatsavaiet al., “SCONNA: A stochastic computing based optical accelerator for ultra-fast, energy-efficient inference of integer-quantized CNNs,”IEEE IPDPS, 2023

work page 2023
[4]

ASTRA: A stochastic transformer neural network accelerator with silicon photonics,

S. Afifi, O. Alo, I. Thakkar, and S. Pasricha, “ASTRA: A stochastic transformer neural network accelerator with silicon photonics,”ACM TECS, vol. 25, no. 1, Article 12, Jan. 2026

work page 2026
[5]

HEANA: A hybrid time-amplitude analog optical accelerator with flexible dataflows for energy-efficient CNN inference,

S. S. Vatsavai, V . S. P. Karempudi, and I. Thakkar, “HEANA: A hybrid time-amplitude analog optical accelerator with flexible dataflows for energy-efficient CNN inference,”ACM TODAES, vol. 30, no. 2, Feb. 2025

work page 2025
[6]

SPOGA: Scaling analog photonic accelerators for byte-size, integer GEMM kernels,

O. A. Alo, S. S. Vatsavai, and I. Thakkar, “SPOGA: Scaling analog photonic accelerators for byte-size, integer GEMM kernels,” arXiv:2407.06134, Jul. 2024

work page arXiv 2024
[7]

Photonic reconfigurable accelerators for efficient inference of CNNs with mixed-sized tensors,

S. S. Vatsavai and I. G. Thakkar, “Photonic reconfigurable accelerators for efficient inference of CNNs with mixed-sized tensors,”IEEE TCAD, vol. 41, no. 11, 2022

work page 2022

[1] [1]

HolyLight: A nanophotonic accelerator for deep learning in data centers,

W. Liuet al., “HolyLight: A nanophotonic accelerator for deep learning in data centers,”IEEE/ACM DATE, 2019

work page 2019

[2] [2]

DEAP-CNN: Digital electronics and analog photonics for convolutional neural networks,

V . Bangariet al., “DEAP-CNN: Digital electronics and analog photonics for convolutional neural networks,”IEEE JSTQE, 2020

work page 2020

[3] [3]

SCONNA: A stochastic computing based optical accelerator for ultra-fast, energy-efficient inference of integer-quantized CNNs,

S. S. Vatsavaiet al., “SCONNA: A stochastic computing based optical accelerator for ultra-fast, energy-efficient inference of integer-quantized CNNs,”IEEE IPDPS, 2023

work page 2023

[4] [4]

ASTRA: A stochastic transformer neural network accelerator with silicon photonics,

S. Afifi, O. Alo, I. Thakkar, and S. Pasricha, “ASTRA: A stochastic transformer neural network accelerator with silicon photonics,”ACM TECS, vol. 25, no. 1, Article 12, Jan. 2026

work page 2026

[5] [5]

HEANA: A hybrid time-amplitude analog optical accelerator with flexible dataflows for energy-efficient CNN inference,

S. S. Vatsavai, V . S. P. Karempudi, and I. Thakkar, “HEANA: A hybrid time-amplitude analog optical accelerator with flexible dataflows for energy-efficient CNN inference,”ACM TODAES, vol. 30, no. 2, Feb. 2025

work page 2025

[6] [6]

SPOGA: Scaling analog photonic accelerators for byte-size, integer GEMM kernels,

O. A. Alo, S. S. Vatsavai, and I. Thakkar, “SPOGA: Scaling analog photonic accelerators for byte-size, integer GEMM kernels,” arXiv:2407.06134, Jul. 2024

work page arXiv 2024

[7] [7]

Photonic reconfigurable accelerators for efficient inference of CNNs with mixed-sized tensors,

S. S. Vatsavai and I. G. Thakkar, “Photonic reconfigurable accelerators for efficient inference of CNNs with mixed-sized tensors,”IEEE TCAD, vol. 41, no. 11, 2022

work page 2022