Explicit Sign-Magnitude Encoders Enable Power-Efficient Multipliers

Felix Arnold, Maxence Bouvier, Ryan Amaudruz, Renzo Andri, Lukas CavigelliarXiv 2025 · DOI · OpenReview

TL;DR

This paper introduces a novel approach to designing power-efficient multipliers for AI hardware by decomposing them into an encoder and a multiplier. By converting inputs to a sign-magnitude format before multiplication, they achieve significant power savings, especially for typical AI workloads. This method offers a way to boost efficiency without sacrificing functional equivalence.

What Problem Does This Paper Address?

Modern AI accelerators, like TPUs and NVIDIA Tensor Cores, rely on massive numbers of parallel processing elements, with multipliers being a critical component. Optimizing these multipliers for both power and area efficiency is crucial for achieving high performance under stringent constraints. While the two's complement (TC) representation is standard for signed integers due to its ease of use in addition, the sign-magnitude (SM) representation is known to be more power-efficient for multiplication, particularly when input values are centered around zero. However, directly replacing TC multipliers with SM ones can break system compatibility. This paper addresses the challenge of leveraging the power benefits of SM multiplication within a TC-dominated system without compromising overall functionality.

Methods

The core of the proposed method is a decomposition strategy:

  1. Encoder Block: A dedicated encoder converts the input operands from the standard two's complement (TC) format to a sign-magnitude (SM) format.
  2. Multiplier Block: An SM multiplier then performs the multiplication.
  3. Output Formatting: The result is then converted back to the TC format to maintain system compatibility.

Crucially, the encoder and multiplier blocks are synthesized and optimized individually. This allows for specialized optimization tailored to each component's specific task. The authors also employ a sophisticated design space exploration (DSE) framework based on Majority-Inverter Graph (MIG) optimization. This framework iteratively applies various synthesis recipes to find optimal circuits, guided by metrics like transistor count and switching activity (SwAct). The SwAct is estimated using a post-synthesis simulation model that accounts for transistor counts in fan-out cells to better approximate dynamic power consumption.

Datasets

Not discussed in the paper. The evaluation uses simulated input streams based on normal distributions with varying standard deviations (σ=2, 3, 4) to mimic typical AI workload distributions.

Key Insights and Ideas

  • Decomposition is Key: Breaking down the multiplier into an encoder and a dedicated SM multiplier is the primary enabler for power savings. This allows for specialized optimization of each stage.
  • SM for Multiplication Wins: Sign-magnitude representation is inherently more power-efficient for multiplication, especially for inputs near zero, which are common in AI.
  • Compatibility Achieved: By converting the output back to two's complement, the proposed method ensures seamless integration into existing systems that expect TC format.
  • Synthesis Tools Miss Opportunities: Standard EDA tools often optimize multipliers as monolithic blocks, failing to exploit the potential gains from decomposition and specialized representations.
  • Slight Range Deviation is Acceptable: Relaxing the full TC range (e.g., by clipping -8 to -7) can yield substantial power improvements, and this deviation is often acceptable in AI applications.
  • DSE is Powerful: Advanced design space exploration techniques are essential for finding optimal configurations and pushing the boundaries of power efficiency.

Key Results

  • Power Savings: For a realistic input stream (normal distribution, σ=3), the decomposed approach shows up to 12.9% lower switching activity compared to a standard TC multiplier.
  • Aggressive Savings: When the input range is slightly reduced (e.g., -7 to +7), switching activity reductions can reach up to 33%.
  • End-to-End SM: A fully SM-based multiplier (input and output) demonstrated a drastic 75.5% improvement in switching activity for σ=2.
  • Area Trade-offs: SM-based multipliers generally have a lower transistor count (down to 204 transistors for SM-SM vs. 420 for TC-TC baseline). However, the addition of the encoder can slightly increase the total transistor count for some configurations (e.g., SME-TC configuration B shows a 10% increase in total transistors compared to the baseline).
  • Further Optimization: The guided DSE framework can yield an additional 5-10% improvement in power efficiency.

Limitations and Future Work

  • Bit-width Restriction: The study is limited to 4-bit multipliers. While 8-bit evaluations suggest reproducibility, larger bit-widths might present different challenges.
  • Fixed-Point Only: The evaluation focuses solely on fixed-point representations. Extending to floating-point formats could reveal further insights.
  • Simplified SwAct Model: The SwAct model uses transistor count as a proxy for cell complexity, which is a simplification of dynamic behavior.
  • Synthesis-Only: The analysis is limited to synthesis. Place-and-route and timing analysis are needed for a complete system-level evaluation.
  • Design Space Bias: The approach starts from specific points (e.g., Verilog star operator). Understanding how these starting points affect generality is needed.
  • Limited Data Representations: Only a few TC and SM variants were explored. Other encodings or enhancements might be beneficial.
  • Depth Not Primary Focus: While power is optimized, depth (related to speed) was not the main objective and could be further improved.
  • Optimization Effort: Not all configurations were subjected to the highest level of optimization effort.
  • Application Regime: The SM encoding was applied only to the multiplier. Exploring its use across the entire data path (weights, activations) is a promising direction.

Future work will focus on addressing these limitations, including exploring higher bit-widths, extending to floating-point, incorporating physical design aspects, and automating the search for optimal encodings.

Summary updated

Abstract

This work presents a method to maximize power-efficiency of fixed point multiplier units by decomposing them into sub-components. First, an encoder block converts the operands from a two's complement to a sign magnitude representation, followed by a multiplier module which performs the compute operation and outputs the resulting value in the original format. This allows to leverage the power-efficiency of the Sign Magnitude encoding for the multiplication. To ensure the computing format is not altered, those two components are synthesized and optimized separately. Our method leads to significant power savings for input values centered around zero, as commonly encountered in AI workloads. Under a realistic input stream with values normally distributed with a standard deviation of 3.0, post-synthesis simulations of the 4-bit multiplier design show up to 12.9% lower switching activity compared to synthesis without decomposition. Those gains are achieved while ensuring compliance into any production-ready system as the overall circuit stays logic-equivalent. With the compliance lifted and a slightly smaller input range of -7 to +7, switching activity reductions can reach up to 33%. Additionally, we demonstrate that synthesis optimization methods based on switching-activity-driven design space exploration can yield a further 5-10% improvement in power-efficiency compared to a power agnostic approach.
BibTeX
@misc{arnold2025explicit,
  author = {Felix Arnold and Maxence Bouvier and Ryan Amaudruz and Renzo Andri and Lukas Cavigelli},
  title = {{Explicit Sign-Magnitude Encoders Enable Power-Efficient Multipliers}},
  year = {2025},
  eprint = {2507.18179},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.18179},
}
Figure