Architecture-Level Modeling of Photonic Deep Neural Network Accelerators

Tanner Andrulis, Gohar Irfan Chaudhry, Vinith M. Suriyakumar, Joel S. Emer, Vivienne SzeIspass 2024 · DOI

TL;DR

This paper introduces a new modeling tool that bridges the gap between photonic hardware and deep learning workloads. By adapting existing compute-in-memory (CiM) modeling techniques, they enable accurate, full-system evaluation of photonic accelerators, revealing that data conversion and DRAM access can dominate energy consumption. Their work demonstrates a 3x energy improvement through optimizations that reduce these overheads.

What Problem Does This Paper Address?

Deep Neural Networks (DNNs) are computationally intensive, driving the need for specialized hardware accelerators. Photonics offers a promising avenue for accelerating DNNs due to its potential for low-energy, high-speed optical interconnects and analog computation. However, realizing these benefits requires careful consideration of the entire system, including the energy costs associated with converting data between different domains (digital-electrical, analog-electrical, analog-optical) and fetching data from memory (DRAM). Without a comprehensive modeling tool, it's difficult for researchers to accurately evaluate novel photonic designs, explore the vast design space, and understand the true system-level energy implications. This paper addresses this gap by providing a methodology and tool to model these full-system considerations for photonic DNN accelerators.

Methods

The core contribution is the adaptation of the open-source CiMLoop tool, originally designed for modeling compute-in-memory (CiM) systems, to accurately model photonic DNN accelerators. The authors highlight the similarities between CiM and photonic systems, particularly their reliance on multiple domains (digital-electrical, analog-electrical) and the challenge of high cross-domain conversion energy. They augment CiMLoop's component library with models for photonic components like microring resonators, lasers, and photodiodes. This allows the tool to capture data movement, reuse opportunities, and energy costs across digital-electrical (DE), analog-electrical (AE), and analog-optical (AO) domains. The tool then maps DNN workloads onto the specified photonic architecture to estimate full-system area, energy, and throughput.

Datasets

The paper evaluates their modeling approach using two standard DNN workloads: VGG16 and AlexNet. They also reference the Albireo [6] photonic system as the architecture under study.

Key Insights and Ideas

  • CiM Modeling Adapts to Photonics: The surprising insight is that existing CiM modeling frameworks can be effectively extended to model photonic systems, leveraging the shared challenges of multi-domain computation and data conversion overheads.
  • Data Conversion and DRAM are Bottlenecks: The modeling reveals that for photonic accelerators, the energy cost of converting data between domains (e.g., electrical to optical) and fetching data from DRAM can significantly outweigh the energy saved by optical interconnects or analog computation, especially in aggressively scaled systems.
  • Reuse is Key to Energy Efficiency: Increasing data reuse within the analog-optical (AO) and analog-electrical (AE) domains can drastically reduce the need for costly data conversions, leading to substantial energy savings.
  • Accurate Throughput Modeling is Crucial: Standard performance metrics can be misleading. The authors demonstrate that their model accurately captures underutilization issues arising from varying DNN tensor shapes, which significantly impacts actual throughput compared to ideal scenarios.

Key Results

  • Model Validation: The developed model shows an average energy error of only 0.4% compared to reported results for the Albireo system.
  • Throughput Accuracy: The model accurately predicts lower throughput for Albireo on VGG16 and AlexNet than ideal scenarios, attributing this to underutilization caused by different DNN weight tensor shapes.
  • DRAM Dominance in Aggressive Scaling: For aggressively scaled photonic systems, DRAM can consume up to 75% of the total system energy.
  • Significant Energy Savings: By implementing optimizations like input/output batching and keeping data on-chip in global buffers, the authors achieved a 3x (67%) reduction in system energy for aggressively scaled Albireo.
  • Data Converter Energy Reduction: Exploring architectural variations that increase data reuse in AO and AE domains reduced data converter energy by 42% and overall accelerator energy by 31%.

Limitations and Future Work

The authors acknowledge that their model, while powerful, relies on the accuracy of the underlying component models. Future work could involve expanding the component library to include more advanced photonic devices and exploring a wider range of architectural optimizations. They also suggest that this modeling framework can facilitate co-design between photonic hardware and DNN algorithms, enabling more holistic system-level optimization.

Summary updated

BibTeX
@inproceedings{andrulis2024architecture,
  author = {Tanner Andrulis and Gohar Irfan Chaudhry and Vinith M. Suriyakumar and Joel S. Emer and Vivienne Sze},
  title = {{Architecture-Level Modeling of Photonic Deep Neural Network Accelerators}},
  booktitle = {Ispass},
  year = {2024},
  doi = {10.1109/ispass61541.2024.00040},
}
Figure