Approximate Multiplier Units for Energy-Efficient LSTM Acceleration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing neural network models, particularly LSTM, face challenges in achieving high energy efficiency and throughput on energy-constrained mobile devices due to their compute-intensive and memory-bandwidth requirements, making them unsuitable for scalable deployment in IoT and mobile applications.
Innovation Solution
The design of an energy-efficient LSTM accelerator (ELSA) that employs approximate multiplications and elastic pipelining, utilizing compact and low-power computation units with hierarchical control to synchronize operations and maximize throughput, while maintaining accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard RNN/LSTM models are used to capture temporal dependencies and achieve high accuracy, then measurement precision is improved, but use of energy and device complexity increase significantly
Solution Approach 1:
The multiplication operation is segmented into two phases: an approximate multiplication phase that produces a preliminary result quickly, followed by a correction phase that refines the result. This segmentation allows the system to achieve high accuracy through the correction step while maintaining low energy consumption through the efficient approximate phase, resolving the contradiction between accuracy and energy use.
Solution Approach 2:
The system dynamically changes the precision parameter based on operational requirements. During normal operation, approximate multiplication with lower precision is used to save energy. When higher accuracy is needed, the correction mechanism adjusts the precision parameters to refine the result, allowing the system to adapt between energy efficiency and accuracy as needed.
2Measurement precision
If standard RNN/LSTM models are used to capture temporal dependencies and achieve high accuracy, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The computational process is divided into an approximate multiplication stage and a correction stage. The approximate stage uses simpler logic structures that reduce device complexity, while the correction stage selectively applies additional computational steps only when needed, maintaining accuracy without requiring the entire system to be complex.
Solution Approach 2:
Instead of always performing full-precision multiplication, the system applies approximate multiplication (partial action) for most operations where full precision is not critical. The correction mechanism then applies additional precision only partially, when and where needed, reducing overall device complexity while maintaining required accuracy.
3Manufacturing precision
If exact multiplication is used in neural network computation units, then manufacturing precision is improved, but productivity decreases due to increased computation time
Solution Approach 1:
The multiplication operation is segmented into an approximate phase that processes data quickly with lower precision, followed by a correction phase that refines selected results to full precision. This segmentation allows the majority of operations to proceed rapidly while maintaining overall computation precision through the correction step, thereby increasing throughput without sacrificing accuracy.
Solution Approach 2:
The system applies full-precision computation (excessive action) only selectively to critical operations or when error thresholds are exceeded, while using approximate computation for the majority of operations. This partial application of high-precision processing maintains manufacturing precision where needed while significantly improving overall productivity.
4Measurement precision
If compute-intensive operations are performed on large dimensional data in LSTM models, then measurement precision is improved, but use of energy increases
Solution Approach 1:
The compute-intensive multiplication operations are segmented into approximate and correction phases. The approximate phase handles the bulk of computational work with reduced energy requirements, while the correction phase selectively refines results to maintain accuracy. This segmentation resolves the contradiction by distributing the energy burden across two stages with different precision requirements.
Solution Approach 2:
The system dynamically changes computation parameters based on data characteristics and operational context. For large dimensional data, the system switches to approximate multiplication with lower energy consumption, then applies correction parameters selectively to maintain accuracy, allowing the system to adapt energy usage to the specific computational demands.
Data Source
AI summary
Disclosed is neural network circuitry having a first plurality of logic cells that is interconnected to form neural network computation units that are configured to perform approximate computations. The neural network circuitry further includes a second plurality of logic cells that is interconnected to form a controller hierarchy that is interfaced with the neural network computation units to control pipelining of the approximate computations performed by the neural network computational units. In some embodiments the neural network computation units include approximate multipliers that are configured to perform approximate multiplications that comprise the approximate computations. The approximate multipliers include preprocessing units that reduce latency while maintaining accuracy.