Overview
Deploying transformer inference workloads on edge hardware requires evaluating both operational energy consumption and the embodied carbon associated with chip manufacturing. Silicon photonics provides high-bandwidth, low-energy matrix multiplication, yet prior evaluations of microring-resonator-based Dot Product Units (DPUs) have focused primarily on convolutional workloads using analog signaling, without accounting for manufacturing-phase emissions. A comparative evaluation addresses this gap by adapting three photonic DPU organizations to transformer inference under both analog and stochastic signaling schemes, assessing life-cycle carbon across operational and manufacturing metrics.
Research Context
Existing microring-resonator DPU designs were initially configured for convolutional computations. In earlier assessments, these architectures were evaluated using analog signaling paradigms and analyzed exclusively through the lens of operational energy efficiency. However, migrating transformer models to edge environments shifts the environmental profile of the hardware, making the embodied footprint of fabrication an essential factor alongside active inference power.
Approach
The study examines three core microring-resonator DPU architectures: AMW, MAW, and MWA. Each architecture is adapted specifically for transformer inference workloads and implemented under two separate signaling paradigms: conventional analog signaling and stochastic signaling. This produces six discrete accelerator organizations:
- AMW-Analog and AMW-Stochastic
- MAW-Analog and MAW-Stochastic
- MWA-Analog and MWA-Stochastic
Researchers calculated both embodied carbon and operational carbon per inference across five transformer models. To evaluate manufacturing sensitivity, the analysis incorporates four distinct semiconductor foundry electrical grids. Hardware performance is benchmarked against conventional CPUs, as well as enterprise accelerators including the NVIDIA H100 SXM5 GPU and the Google TPU v5p.
Findings
The evaluation revealed significant structural trade-offs between execution speed, operational energy, and physical die requirements:
- Efficiency gains over CPUs: Across the five evaluated transformer models and four foundry grids, the photonic accelerator configurations achieved improvements of four orders of magnitude in Energy-Delay Product, four orders of magnitude in Carbon-Delay Product, and six orders of magnitude in Carbon-Energy Product relative to CPU baselines.
- Impact of stochastic signaling: Transitioning from analog to stochastic signaling reduced operational energy consumption by approximately 91%. However, the conversion increased silicon die area by 4.3× to 4.6× and amplified embodied carbon by 4.4× to 4.8×.
- Shift in carbon dominance: Due to the footprint overhead of stochastic signaling, embodied carbon equaled or exceeded operational carbon per inference in those configurations.
- Primary drivers of carbon footprint: The specific DPU architectural organization and signaling mechanism served as the primary determinants of total carbon impact, exerting greater influence than variations among the four foundry power grids.
- Architecture-level trade-offs: The fastest evaluated design, MWA, did not yield the highest carbon efficiency. Instead, the MAW-Stochastic organization demonstrated the most favorable balance between computational performance and environmental footprint.
Why This Matters
The findings demonstrate that focusing solely on runtime energy misses the majority of emissions when stochastic photonic accelerators are deployed at the edge. Because manufacturing footprint matches or surpasses inference-phase carbon under stochastic schemes, hardware architects must treat physical organization and signaling choices—rather than foundry power sourcing—as the decisive factors in sustainable accelerator design.