Compiler-Driven Approximation Tuning for Hyperdimensional Computing

Hyperdimensional Computing (HDC) is a brain-inspired paradigm that represents information as high-dimensional vectors. Because HDC is inherently resilient to noise, it allows for aggressive approximation—such as reduced precision or skipped computations—that can significantly improve performance and energy efficiency. However, the design space for these approximations is combinatorially large, especially when considering both software-level transformations and hardware-specific knobs (e.g., ADC resolution in ReRAM). The authors introduce ApproxHDC, a compiler-driven framework that automates the identification and application of these approximations.

ApproxHDC extends the existing HPVM-HDC compiler infrastructure. It performs static analysis to identify HDC primitives, constructs a multi-dimensional approximation search space, and uses efficient search techniques to find configurations that meet user-specified Quality-of-Service (QoS) constraints. By integrating both software-level techniques (like binarization and loop perforation) and hardware-level knobs (like multi-level cell configuration), the framework enables retargetable optimization across CPUs, GPUs, and simulated in-memory accelerators like ReRAM and PCM.

ApproxHDC significantly outperforms manual or non-compiler-integrated approaches. By leveraging HDC-specific domain knowledge, the framework prunes the approximation search space by up to 86 orders of magnitude, allowing for the discovery of high-quality configurations in minutes rather than hours or days. Experimental results demonstrate substantial performance gains: up to 17.25x speedup on CPUs, 15.02x on GPUs, and 4.69x on SpecPCM accelerators. Furthermore, the framework outperforms the state-of-the-art MicroHD library by 3.49x, demonstrating the value of compiler-driven, fine-grained approximation control.

As traditional scaling (Moore's Law) slows, domain-specific accelerators and approximate computing are becoming essential for maintaining performance gains in machine learning. ApproxHDC provides a systematic way to bridge the gap between high-level HDC algorithms and heterogeneous hardware backends. By automating the complex trade-off between accuracy and performance, this work makes HDC more accessible and efficient for resource-constrained environments like edge AI and biomedical signal processing.