PISim: A Unified Photonic Inference Simulator for Hardware-Aware Cross-Layer Design-Space Exploration
Organizations: QPI Laboratories, Inc. Escondido, CA, USA
Abstract
Photonic neural accelerators offer ultra-low latency linear algebra, yet their usable numerical precision, sustainable throughput, and application accuracy remain tightly bounded by surrounding electro-optical interfaces and physical non-idealities. Pre-fabrication design-space exploration (DSE) requires a cross-layer electronic design automation (EDA) framework capable of evaluating disparate optical topologies under identical neural workloads and physical assumptions. We present PISim, a unified Photonic Inference Simulator that couples pretrained deep neural network execution with physics-based models of six accelerator architecture families, six electro-optic modulator technologies, four optical source organizations, and direct or coherent readout schemes. PISim embeds stage-resolved impairment operators---covering control and analog-to-digital quantization, finite extinction ratio, shot noise, correlated relative-intensity noise (RIN), thermal drift, propagation loss, and channel crosstalk---alongside an analytical cycle-level interface timing model to expose electronic-optical bottlenecks. Using simulated MNIST multilayer perceptron (MLP) benchmarks, we demonstrate how PISim identifies optimal precision operating points under multi-impairment stress (peaking at 6 bits for Mach--Zehnder and microring modulators), quantifies a 52.5 percentage-point accuracy gap between VCSEL-array and parallel-fanout laser architectures arising, under the modeled conditions, from power partitioning and correlated source noise, and models interface-limited throughput scaling across emitter-array widths from 32 to 128 channels. This comprehensive framework establishes an extensible, reproducible foundation for accuracy-constrained Pareto optimization prior to photonic tape-out.
Figures & tables
| Architecture Family | Physical Representation & Forward Operator | Primary EDA Design Variables |
|---|---|---|
| Diffractive Optical Networks [ 6 ] | Free-space or slab spatial complex field propagation across phase/amplitude diffractive masks onto detector regions. | Mask resolution, layer axial spacing, diffraction efficiency, transverse misalignment, and detector pixel binning. |
| Mach–Zehnder Interferometer (MZI) Meshes [ 4 , 7 ] | Coherent unitary/arbitrary matrix decomposition ( ) via cascaded beamsplitters and thermo-/electro-optic phase shifters. | Mesh decomposition topology (e.g., Clements [ 7 ] ), phase bit precision, calibration error, and path-dependent insertion loss. |
| Electro-Optic MAC Units [ 5 ] | Point-to-point or broadcast-and-weight intensity/field modulation followed by spatial or temporal photodetector accumulation. | Modulator transfer curve, signed encoding convention, DAC/ADC sampling rates, and matrix tiling factor. |
| Resonator-Based Inference [ 8 ] | WDM spectral routing through cascaded microring resonator (MRR) weight banks with differential balanced readout. | Resonance detuning, WDM channel spacing, thermal drift sensitivity, weight state resolution, and inter-channel crosstalk. |
| VCSEL-Array Processors | Spatially parallel vertical-cavity surface-emitting laser arrays with direct current modulation and optical fan-in/fan-out. | Emitter array width ( ), inter-emitter coupling, nonlinear L-I dynamics, and detector integration window. |
| Free-Space Optical Accelerators [ 9 ] | Spatial light modulation (SLM/DMD) with cylindrical/Fourier lens optical matrix-vector projection onto camera/photodiode arrays. | Spatial pixel encoding, optical point-spread function, SLM frame refresh rate, and sensor readout schedule. |
| Parameter | Documented Value / Setting |
|---|---|
| Dataset | MNIST (test split, indices 0–999) |
| Sample Count | 1,000 samples |
| Network Topology | 3-layer MLP ( ; 50,816 weight bytes) |
| Random Seed | 33 (deterministic PRNG) |
| Reference Accuracy | 97.1% (ideal FP32) |
| Array Width ( ) | channels (E5) |
| Precision | E2 Clean | E3 Stress | E3 Stress | E3 Stress |
|---|---|---|---|---|
| (Bits) | (All Three) | MZM | Ring | QCSE |
| 4 bits | 95.9% | 35.7% | 35.7% | 24.6% |
| 6 bits | 97.0% | 46.4% | 46.4% | 15.9% |
| 8 bits | 97.1% | 44.5% | 44.5% | 19.0% |
| Array Width | VCSEL / | Parallel | Serialized |
|---|---|---|---|
| ( ) | MZM | Laser / MZM | Laser / MZM |
| 32 | 5.563 M/s | 5.563 M/s | 12.399 M/s |
| 64 | 10.710 M/s | 10.710 M/s | 12.496 M/s |
| 128 | 19.584 M/s | 19.584 M/s | 12.545 M/s |
| Min Latency | 51.06 ns | 51.06 ns | 79.71 ns |
Photonic Computing
Photonic computing uses light and optical devices to perform computation, aiming for high throughput and lower energy in machine learning and signal processing.
See all 51 papers