ENLighten Photonic Transformer Accelerator

• ENLighten is a hardware–software co-design framework that fuses Lighten’s PTC-aware compression with a reconfigurable photonic accelerator for efficient Transformer inference.
• The methodology addresses bottlenecks like electro–optic conversion and limited on-chip photonic resources using structured sparsity and dynamic power gating.
• Empirical results on ImageNet show up to a 2.5× energy–delay product improvement with minimal accuracy loss on models such as ViT-Base.

Searching arXiv for ENLighten and related photonic Transformer accelerator work. Searching arXiv for "ENLighten Lighten Transformer photonic accelerator". Searching arXiv for "Lightning-Transformer photonic Transformer accelerator arXiv". ENLighten denotes a hardware–software co-design framework for efficient photonic acceleration of large Transformer inference. In this usage, the term combines Lighten, a post-hoc, photonic-tensor-core-aware compression flow, with ENLighten, a reconfigurable photonic accelerator whose dynamically adaptive tensor cores support fine-grained sparsity and full power gating of inactive optical paths. The framework is motivated by two scaling bottlenecks in photonic Transformer acceleration—electro–optic conversion and data-movement overheads, and the mismatch between limited on-chip photonic tensor-core resources and Transformer scale—and is evaluated on ImageNet with DeiT-Small, DeiT-Base, and ViT-Base models. On ViT-Base, Lighten prunes 50% of parameters with about 1% top-1 accuracy drop after only 3 epochs of fine-tuning within an hour, and deployment on ENLighten yields a $2.5\times$ improvement in energy–delay product over the state-of-the-art photonic Transformer accelerator (Zhu et al., 2 Oct 2025).

1. Scope, terminology, and disambiguation

In this context, ENLighten refers specifically to the photonic-computing framework introduced in “ENLighten: Lighten the Transformer, Enable Efficient Optical Acceleration” (Zhu et al., 2 Oct 2025). The name is structurally tied to its two components: Lighten on the software side and ENLighten on the hardware side. The former compresses Transformer weight matrices into a low-rank component plus a structured-sparse component aligned to photonic tensor-core granularity; the latter executes those compressed models on reconfigurable photonic hardware (Zhu et al., 2 Oct 2025).

The term should be distinguished from several unrelated low-light image enhancement systems that use similar naming. These include “Enlighten Anything: When Segment Anything Model Meets Low-Light Image Enhancement” (Zhao et al., 2023), “Enlighten-Your-Voice: When Multimodal Meets Zero-shot Low-light Image Enhancement” (Zhang et al., 2023), the unpaired GAN-based “EnlightenGAN” (Jiang et al., 2019), local mask-conditioned enhancement via Region-Aware Normalization for Local Enhancement (Yao et al., 2022), and the anime-scenery enhancement framework that trains versions of EnlightenGAN under a Data Relativistic Uncertainty framework (Gao et al., 26 Dec 2025). The shared naming reflects thematic association with “enlightening,” but the photonic ENLighten is a systems-and-architecture proposal rather than an image restoration model.

2. Bottlenecks in photonic Transformer acceleration

ENLighten is organized around two bottlenecks identified for scaling photonic acceleration of Transformers. The first is the cost of electro–optic conversion and data movement. As model size grows, conversion between electrical and optical domains—via DAC/MZM modulation on the input side and photodetection/TIA/ADC on the output side—together with movement of large weight tensors, can dominate energy and latency. In prior photonic Transformer accelerators, these overheads can erode more than half of the energy cost as model size scales (Zhu et al., 2 Oct 2025).

The second bottleneck is the mismatch between limited photonic tensor-core resources and Transformer scale. Photonic devices are bulky, and current fabrication limits constrain on-chip PTC density. As a result, larger models such as ViT-Base require frequent time-multiplexed reuse of a small pool of cores, which dilutes throughput gains (Zhu et al., 2 Oct 2025).

The reported scaling trend is explicit. Recent photonic Transformer accelerators show rapidly diminishing returns: throughput improvements drop from hundreds of $\times$ to only tens of $\times$, and energy-efficiency gains fall from about $30\times$ to under $10\times$ when scaling from ViT-Tiny to ViT-Base at 8-bit precision (Zhu et al., 2 Oct 2025). This makes compression and hardware reconfiguration co-dependent rather than separable optimizations.

3. Lighten: PTC-aware compression flow

Lighten is a three-level compression pipeline designed to preserve accuracy under tight retraining budgets while producing sparsity that maps efficiently to photonics. Its basic decomposition for a weight matrix $W \in \mathbb{R}^{m \times n}$ is

$W \approx U_r \Sigma_r V_r^{\top} + S,$

where the first term is low-rank and $S$ is a structured-sparse residual aligned to photonic tensor-core granularity (Zhu et al., 2 Oct 2025).

At the first level, Lighten performs layer-wise low-rank plus structured-sparse decomposition. The objective is activation-aware: for layer $l$, with calibration activation matrix $X_{\text{calib},l}$, it forms

$\times$0

and minimizes a weighted reconstruction error so that compression respects real activation statistics rather than only weight-space error (Zhu et al., 2 Oct 2025). The low-rank update is obtained by truncated SVD on the residual, while structured sparsity is imposed through PTC-aware column selection.

The sparse pattern is matched to PTC tiling. PTCs tile $\times$1 into blocks of size $\times$2, and Lighten uses column-wise structured sparsity at a granularity equal to the PTC’s column dimension $\times$3 across the full horizontal dimension of $\times$4. It computes the $\times$5 norm of column-vectors of length $\times$6, selects the top $\times$7 columns per block, and condenses them leftward into dense submatrices of size $\times$8. This reduces the footprint from $\times$9 to $\times$0, and those condensed submatrices map directly to the sparse engine (Zhu et al., 2 Oct 2025). Column pruning is preferred over row pruning because it aligns with output-stationary dataflow and reduces ADC and buffer overhead.

After the initial decomposition, Lighten applies local low-rank adaptation. Small-rank adapters $\times$1, with rank at most $\times$2, are optimized on a small calibration set and then merged back into the factors. The stated purpose is to boost fidelity without increasing parameter count, using only a few gradient steps (Zhu et al., 2 Oct 2025).

The second level is a fast batch-wise rank allocator. Since redundancy differs substantially across layers, Lighten avoids uniform rank budgets. It defines the layer error indicator

$\times$3

normalizes those errors across layers, and performs batch-wise greedy rank allocation with coarse-to-fine increments tied to the PTC base granularity $\times$4. ViT-Base search completes in approximately 10 minutes on a single GPU, and the procedure reuses precomputed full-rank factors rather than recomputing SVDs (Zhu et al., 2 Oct 2025).

The third level is decoupled two-stage distillation. Stage 1 performs one epoch of block-wise feature alignment,

$\times$5

over attention and MLP blocks. Stage 2 then applies logit-level distillation for the remaining epochs,

$\times$6

In the reported setup, DeiT-Small trains for 6 epochs and ViT-Base for 3 epochs, achieving approximately 1% accuracy drop at 50% pruning (Zhu et al., 2 Oct 2025).

4. ENLighten hardware architecture

ENLighten builds on the Lightning-Transformer optical accelerator architecture, but augments it with a reconfigurable sparse engine and adaptive tensor cores so that Lighten’s structured sparsity becomes physically exploitable rather than merely representable (Zhu et al., 2 Oct 2025).

Its compute primitive is implemented by a dynamically operated photonic tensor core. One cycle computes

$\times$7

with $\times$8 and $\times$9, producing $30\times$0 outputs using $30\times$1 vertical inputs, $30\times$2 horizontal outputs, and $30\times$3 wavelengths. The tensor core uses arrays of optical dot-product engines with wavelength-division multiplexing. Within each input waveguide, a demultiplexer splits light into wavelength channels; each channel is modulated by a Mach–Zehnder modulator with the corresponding input or weight value, and the recombined signals interfere through directional couplers to realize inner products. DACs drive MZMs, photodetectors with TIAs and shared ADCs capture outputs, and both weights and activations are 8-bit (Zhu et al., 2 Oct 2025).

The defining hardware modification is broadband, calibration-free light redistribution. ENLighten replaces the first two splitter stages upstream of each PTC with tunable, broadband MZIs that use multimode interferometers rather than wavelength-sensitive directional couplers at outputs. Each arm has a 200 $30\times$4m thermal phase shifter. The design exposes only discrete modes: equal split $30\times$5 when phase shifters are disabled, and full switching $30\times$6 or $30\times$7 at phase extrema $30\times$8. According to the reported simulations, these extrema are wavelength-insensitive across WDM channels and avoid arbitrary split ratios that would require heavy calibration in WDM systems (Zhu et al., 2 Oct 2025).

This mechanism enables fine-grained row gating. Switching one second-stage splitter to $30\times$9 suppresses light to one quarter of PTC rows, giving $10\times$0 granularity; multiple such units allow flexible combinations that trade accuracy against efficiency. The paper describes this as enabling full power gating of inactive photonic paths and post-fabrication flexibility in sparsity granularity (Zhu et al., 2 Oct 2025).

Architecturally, ENLighten contains a dense engine and a reconfigurable sparse engine. The dense engine inherits dense DPTCs at full resolution and supports photonic input broadcast, so the same input can feed multiple tiles while minimizing per-tile DAC and modulation power. The sparse engine executes the condensed $10\times$1 submatrices produced by Lighten, uses adaptive granularity such as cores of height 8 with $10\times$2 row gating, and relies on a lightweight input indexer because it cannot use input broadcast. Both engines use output-stationary accumulation with shared ADC/TIA banks (Zhu et al., 2 Oct 2025).

5. Mapping flow, scheduling, and empirical performance

The system-level flow is explicitly co-designed. For each Transformer layer, a small calibration set is used to compute $10\times$3, choose a global target compression $10\times$4, and fix a per-block sparsity ratio $10\times$5 aligned to PTC column granularity. Layer-wise decomposition alternates between truncated SVD for the low-rank term and $10\times$6-norm-based column selection for the structured-sparse term, followed by local low-rank adaptation. Model-wise search then adjusts layer ranks until the global budget is met. Minimal fine-tuning follows in two stages, after which low-rank factors and uncompressed layers are mapped to the dense engine, while condensed sparse submatrices are mapped to the sparse engine with the reconfigurable splitters set to the desired granularity (Zhu et al., 2 Oct 2025).

Scheduling is static: low-rank and dense workloads run on the dense engine, while structured-sparse workloads run on the sparse engine. Sparse cores can also execute dense workloads in full operating mode, though dynamic partitioning is identified as future work (Zhu et al., 2 Oct 2025).

The evaluation uses ImageNet classification with DeiT-Small (22.1M parameters), DeiT-Base (86.6M), and ViT-Base (86.5M). Compression is reported at 30%, 40%, and 50% parameter reduction. In zero-shot settings, Lighten consistently outperforms sparse-only WANDA, pure low-rank compression, and OATS, which uses uniform rank and sparsity assignments. After limited fine-tuning, DeiT-Base and ViT-Base at 50% pruning incur approximately 1% top-1 accuracy drop after only 3 epochs, while DeiT-Small at 30% pruning shows approximately 1.5 points drop (Zhu et al., 2 Oct 2025).

Ablation results clarify the contribution of each component. The rank allocator substantially improves accuracy, with the example of DeiT-Small showing a $10\times$7 gain over uniform OATS at 50% reduction. Local low-rank adaptation adds a further $10\times$8 zero-shot accuracy gain on DeiT-Small at 50% reduction. Two-stage distillation gives the best recovered accuracy, with the cited examples of DeiT-Small reaching 74.75% at 50% reduction and DeiT-Base reaching 80.70% (Zhu et al., 2 Oct 2025).

Quantization and noise sensitivity are also evaluated. With 8-bit weights and activations via PTQ4ViT, the additional loss is minimal due to Hessian-aware PTQ, and with 3% random noise injected into weights and activations, the accuracy drop remains under approximately 1% (Zhu et al., 2 Oct 2025).

At the hardware level, the reported improvements are framed in terms of energy, delay, and energy–delay product. The conceptual model is

$10\times$9

$W \in \mathbb{R}^{m \times n}$0

and

$W \in \mathbb{R}^{m \times n}$1

On ViT-Base with 50% compression, the framework reports approximately 40% total energy savings and $W \in \mathbb{R}^{m \times n}$2 inference speedup compared to the uncompressed dense baseline running on LT-Base-scaled, and the combined Lighten-plus-ENLighten system achieves a $W \in \mathbb{R}^{m \times n}$3 EDP improvement over the state-of-the-art photonic Transformer accelerator (Zhu et al., 2 Oct 2025).

6. Assumptions, limitations, and research significance

The framework assumes WDM-capable photonic tensor cores and broadband-tolerant light redistribution through MZI–MMI splitters with thermal phase shifters. Fabrication variability and thermal drift are acknowledged; the symmetric splitter design is intended to minimize wavelength sensitivity and calibration needs, but large-system integration still requires baseline calibration and thermal control. Reconfiguration is intentionally discrete—$W \in \mathbb{R}^{m \times n}$4 and $W \in \mathbb{R}^{m \times n}$5 only—because arbitrary split ratios would compromise WDM tolerance (Zhu et al., 2 Oct 2025).

Several operating constraints are explicit. Memory hierarchy and electro–optic interfaces remain significant energy contributors even after compression. Models with low redundancy, or tasks that require very fine sparsity granularity, may need smaller compression ratios to preserve accuracy. Sparse engines cannot exploit input broadcast, so at low compression ratios the added encoding load reduces latency gains. The framework also identifies autoregressive LLMs as less suitable for current PTC acceleration because they are memory-bound, whereas the present focus is on compute-bound ViTs (Zhu et al., 2 Oct 2025).

The stated future directions include enhanced structured sparsity, mixed-precision optics, dynamic scheduling and core partitioning between dense and sparse phases, and tighter electrical–optical integration to reduce off-chip memory traffic and interface overheads (Zhu et al., 2 Oct 2025). A plausible implication is that ENLighten’s main contribution is not only a faster accelerator or a better compression method in isolation, but a concrete demonstration that photonic efficiency at Transformer scale depends on matching algorithmic sparsity structure to the physical granularity and gating properties of the optical substrate.

Within that framing, ENLighten occupies a specific position in the photonic AI literature: it treats post-hoc compression, rank allocation, distillation, tensor-core granularity, sparse mapping, and optical power gating as elements of a single optimization problem rather than as independent layers of the stack (Zhu et al., 2 Oct 2025).