Fixed Function Shared Accelerator Generation via ML Fingerprinting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Modern systems-on-a-chip face inefficiencies due to the large number of specialized accelerators that do not cover the full spectrum of workloads, leading to the 'accelerator wall' challenge, where exponential use of accelerators may hit a hard limit, and existing methods for finding similarities between workloads are complex and computationally expensive, such as graph-isomorphism, which requires significant time to discover isomorphism between subgraphs.
Innovation Solution
The development of a lightweight fingerprinting methodology using machine learning methods to identify and generate Fixed-Function Shared Accelerators (FFSAs) by encapsulating a kernel's static and dynamic behavior, generating abstract syntax trees, and creating fingerprinting vectors that encode relevant features, which are then used to predict similarities between workloads and output candidate FFSAs, reducing computational burden and increasing coverage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If graph-isomorphism methods are used to find similarities between workloads, then accuracy in identifying isomorphic cases is improved, but computational time increases significantly (from minutes to months)
Solution Approach 1:
The patent segments the workload analysis into multiple stages: first generating ASTs from source code, then creating fingerprinting vectors that capture essential characteristics, and finally using ML models for similarity prediction. This segmentation allows the system to avoid direct graph-isomorphism computation while maintaining accuracy through progressive refinement of workload representations.
Solution Approach 2:
The patent introduces fingerprinting vectors as an intermediary representation between the original workload code and the similarity comparison process. These vectors encode key characteristics (operation counts, data flow patterns, control structures) and serve as a computationally efficient proxy for full graph-isomorphism analysis, enabling fast preliminary screening before detailed verification.
2Adaptability or versatility
If more specialized accelerators are added to cover more workloads, then workload coverage is improved, but device complexity increases exponentially (accelerator wall)
Solution Approach 1:
The patent enables a single accelerator to perform multiple workloads by identifying and exploiting similarities between different computational kernels. Through the fingerprinting and ML-based similarity detection, the system determines which workloads can share the same hardware accelerator, allowing one accelerator to be universally applicable to multiple workloads rather than requiring dedicated accelerators for each workload.
Solution Approach 2:
The patent merges multiple workload requirements into shared accelerator designs by identifying common computational patterns. When the ML model predicts high similarity between workloads, their hardware implementations are merged into a single accelerator that can handle both workloads, reducing the total number of accelerators needed while maintaining comprehensive workload coverage.
3Productivity
If lightweight fingerprinting methodology is used instead of graph-isomorphism, then computational time is reduced significantly, but measurement precision may be compromised
Solution Approach 1:
The patent performs preliminary analysis by generating ASTs and fingerprinting vectors that capture essential workload characteristics before the actual similarity comparison. This preliminary action creates a compressed representation of workloads that retains the most discriminative features, enabling fast ML-based comparison without sacrificing the accuracy needed to distinguish truly similar workloads from dissimilar ones.
Solution Approach 2:
The patent transforms the workload representation from complex graph structures to simplified fingerprinting vectors with specific parameters (operation counts, data flow characteristics, control structure metrics). This parameter change enables the use of efficient ML similarity measures while preserving the essential characteristics needed for accurate isomorphism detection, achieving both speed and precision.
Data Source
AI summary
Systems and methods for generating fixed function shared accelerators (FFSAs) implement and/or comprise operations of receiving source code, the source code indicating a plurality of workloads to be performed by an electronic circuit; generating a plurality of abstract syntax trees (ASTs) based on the source code, wherein respective ones of the plurality of ASTs include a plurality of nodes corresponding to function instructions; generating a plurality of fingerprinting vectors corresponding to the plurality of ASTs, wherein respective ones of the plurality of fingerprinting vectors encode at least one of a number of nodes, a number of edges, a density, a computation intensity, an operands percentage, a control, or a data dependency; and providing the plurality of fingerprinting vectors to a machine learning (ML) model, wherein the ML model is configured to predict similarities between different ones of the plurality of workloads and to output at least one candidate FFSA.