FHE Accelerator Coefficient Ordering for Lower Bootstrapping Dataflow
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The execution of fully homomorphic encryption (FHE) programs is hindered by complex bootstrapping processes that require significant computational and memory resources, along with high intra-chip data movement, leading to impractical bandwidth and power consumption.
Innovation Solution
Optimize the order of polynomial coefficients within an FHE accelerator by selecting an optimal ordering that minimizes data transfers and permutations, using a management module to configure permute units and perform permutations efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the bootstrapping process is executed frequently to maintain decryption accuracy, then the reliability of FHE operations is improved, but the computational and memory resource consumption increases
Solution Approach 1:
The patent pre-calculates and stores permutation lookup tables during system initialization, so that during bootstrapping operations, the system only needs to perform simple table lookups and combinations rather than complex real-time permutations. This preliminary preparation reduces the computational burden during actual FHE operations while maintaining decryption accuracy.
Solution Approach 2:
The patent divides the bootstrapping process into distinct phases: permutation phase, multiplication phase, and rescaling phase. By segmenting the process, the system can optimize each phase independently and manage computational resources more efficiently, reducing the overall complexity while maintaining reliability.
2Reliability
If polynomial permutations are performed during bootstrapping, then the correctness of FHE operations is maintained, but the intra-chip data movement and bandwidth requirements become impractical
Solution Approach 1:
The patent pre-computes permutation results and stores them in lookup tables during system initialization. During actual bootstrapping operations, the system retrieves pre-computed permutation data from these tables rather than performing real-time permutations, dramatically reducing intra-chip data movement and bandwidth requirements while maintaining operational correctness.
Solution Approach 2:
The patent creates and stores copies of permutation tables in on-chip memory during initialization. These copied lookup tables allow the system to perform permutations by simple memory access rather than complex computational operations, reducing the energy and bandwidth consumption associated with real-time polynomial permutations.
3Reliability
If polynomial permutations are performed during bootstrapping, then the FHE scheme parameters are maintained, but the power consumption reaches impractical levels
Solution Approach 1:
The patent performs power-consuming permutation computations during system initialization and stores the results in lookup tables. During actual FHE operations, the system only needs to perform low-power table lookups and simple combinations, maintaining FHE scheme parameters while reducing power consumption by orders of magnitude during operational phases.
Data Source
AI summary
A method and device for optimizing dataflow load in an accelerator of a fully homomorphic encryption (FHE) program are provided. The accelerator is configured with a FHE network including a plurality of permute units, and the method includes obtaining a set of program parameters; obtaining a set of optional orderings; determining optimal program parameters to match an ordering of the set of optimal orderings to yield a required dataflow load; and modifying a FHE program to place coefficients in the permute units and perform the permutations based on the optimal program parameters and matching ordering.