ApproxHDC: Approximation Tuning for Hyperdimensional Computing
Abstract
As Moore's law reaches its physical and economic limits, domain-specific approaches are increasingly employed to accelerate machine learning workloads. Hyperdimensional Computing (HDC) represents one such emerging paradigm, offering an alternative to conventional deep learning techniques. Rooted in cognitive models of computation, HDC is designed bottom-up with hardware efficiency as a first-class objective. Consequently, HDC workloads map naturally to heterogeneous hardware platforms, including CPUs, GPUs, and FPGAs, as well as emerging in-memory computing technologies such as Resistive RAM (ReRAM) and Phase-Change Memory (PCM). The intrinsic tolerance of HDC algorithms to noise and approximation enables substantial performance gains with minimal loss in inference quality. In this work, we introduce ApproxHDC, a framework for automated identification and application of domain-specific approximations in HDC workloads. ApproxHDC extends the HPVM-HDC compiler infrastructure to enable retargetable compilation across diverse hardware backends, including CPUs, GPUs, and simulated ReRAM and PCM-based accelerators. The space of possible software and hardware approximations in HDC is exponentially large, making exhaustive exploration intractable; ApproxHDC employs efficient search and analysis techniques to navigate this space and identify high-impact approximations. We further integrate hardware-level approximation mechanisms within the HDC accelerator to demonstrate ApproxHDC's ability to exploit both software and hardware approximations. Experimental evaluation shows that ApproxHDC achieves up to 17.25x speedup on CPUs, 14.97x on GPUs, and 4.69x on SpecPCM while staying within user-specified quality-of-service constraints, outperforming the state-of-the-art MicroHD framework by 3.49x. HDC-specific search space pruning reduces the design space by 71 to 491 orders of magnitude automatically and up to 89 more with developer guidance, enabling discovery of high-quality configurations within minutes.