Abstract
Fully Homomorphic Encryption (FHE) is a transformative technology for data privacy, enabling computation on encrypted data without decryption. However, the massive computational overhead of FHE—often four orders of magnitude slower than plaintext processing—remains a significant barrier to adoption. EFFACT introduces a full-stack acceleration platform that prioritizes resource efficiency over raw hardware scaling. By co-designing a custom Instruction Set Architecture (ISA), an LLVM-based compiler, and specialized hardware featuring a streaming memory architecture and reconfigurable functional units, EFFACT achieves state-of-the-art performance density. Experimental results demonstrate that EFFACT delivers up to 11.89× better performance density than previous specialized ASICs while operating with significantly less on-chip memory.
Introduction
Fully Homomorphic Encryption (FHE) is frequently described as the "holy grail" of privacy-preserving computation. By allowing data to remain encrypted throughout its entire processing lifecycle, FHE eliminates the need for users to trust third-party service providers with sensitive information. Despite this promise, the practical application of FHE is hindered by its extreme computational intensity. Existing hardware accelerators, including specialized ASICs and FPGAs, typically attempt to bridge this performance gap through "brute force" scaling—deploying hundreds of megabytes of on-chip SRAM and tens of thousands of multipliers. This approach results in designs that are expensive, power-hungry, and often lack the flexibility to support diverse FHE schemes such as CKKS, BGV, or BFV.
EFFACT challenges this paradigm by demonstrating that a "full-stack" approach focused on resource efficiency can outperform massive, hard-coded designs. By optimizing the interaction between the compiler and the underlying hardware, EFFACT achieves high throughput with a fraction of the hardware resources used by contemporary accelerators.
The EFFACT Architecture: A Full-Stack Approach
The core philosophy of EFFACT is the tight integration of software and hardware. The platform consists of three primary components: a custom ISA, an LLVM-based compiler, and a specialized hardware architecture.
Streaming Memory Access
One of the primary bottlenecks in FHE acceleration is the movement of data between high-bandwidth memory (HBM) and on-chip SRAM. Traditional designs rely on large SRAM buffers to minimize DRAM access. EFFACT, however, utilizes a streaming memory architecture. The compiler identifies data with low reuse and streams it directly from DRAM to the Functional Units (FUs), bypassing the SRAM. This strategy allows EFFACT to operate effectively with only 27MB of on-chip SRAM, a dramatic reduction compared to the 200MB to 512MB required by competing high-performance designs.
Reconfigurable Functional Units
In FHE, the Number Theoretic Transform (NTT) is a critical operation, yet the authors observed that NTT instructions only account for approximately 7% of the total instruction count. The remaining 93% consists of standard arithmetic operations like additions and multiplications. EFFACT introduces reconfigurable functional units that "double-dip" as both NTT and Multiply-Accumulate (MAC) units. When not performing transforms, these units are reconfigured to handle standard arithmetic, maximizing transistor utility and eliminating the need for separate, idle hardware blocks.
Computation Merging and Fixed-Network Automorphism
EFFACT further optimizes performance through mathematical and structural refinements:
- Computation Merging: The platform merges maintenance operations, such as Montgomery transformations and constant scaling, into the Base Conversion (BConv) step. This eliminates redundant processing cycles and reduces overall latency.
- Fixed-Network Automorphism: To handle matrix transpositions without the power and area costs of expensive crossbars, EFFACT employs a "fixed network" that exploits the bit-reversal patterns inherent in NTT-domain data.
Experimental Evaluation
The EFFACT platform was evaluated using four major FHE benchmarks: Fully-packed Bootstrapping (CKKS), Logistic Regression training (HELR), ResNet-20 inference, and Encrypted Database Queries (DBLookup). The results were compared against state-of-the-art (SOTA) accelerators, including CraterLake, BTS, ARK, MAD, and F1.
Key Performance Metrics
- Throughput: ASIC-EFFACT is 13.49× faster than a GPU and 4.93× faster than MAD (the previous leader in efficient design) for bootstrapping operations.
- Resource Efficiency: EFFACT achieves between 1.46× and 11.89× higher performance density (performance per chip area) than prior ASICs. Its power efficiency also surpasses previous works by at least 1.48×.
- FPGA Versatility: The FPGA implementation of EFFACT outperforms SOTA FPGA accelerators by a geometric mean of 1.22× across various benchmarks, proving the design's effectiveness across different hardware substrates.
Discussion and Limitations
While EFFACT represents a significant leap in efficiency, certain trade-offs exist. The streaming model is heavily dependent on High-Bandwidth Memory (HBM); performance may degrade on systems with standard DDR memory where bandwidth is more constrained. Additionally, while EFFACT is the most efficient design per unit of area, "resource-heavy" designs like CraterLake may still offer lower absolute latency for single operations due to their sheer volume of hardware units.
During FPGA prototyping, the authors also encountered routing congestion when scaling to 256 lanes at high clock speeds. This necessitated running the physical prototype at a lower frequency (12.5 MHz) and scaling the results for comparative analysis, suggesting that physical layout remains a challenge for highly parallel FHE designs.
Conclusion
EFFACT demonstrates that the path to practical FHE lies in architectural intelligence rather than just hardware scaling. By leveraging a compiler-driven approach and reconfigurable hardware, EFFACT provides a scalable, efficient, and programmable platform for the next generation of privacy-preserving applications. Future work will focus on optimizing the compiler for "transciphering" tasks, such as switching between symmetric encryption (AES) and FHE, to further broaden the platform's utility.
Summary updated