pith. sign in

arxiv: 2606.26547 · v1 · pith:VREWJC53new · submitted 2026-06-25 · 💻 cs.PL · cs.CL· cs.PF

Compiler-Driven Approximation Tuning for Hyperdimensional Computing

Xavier Routh , Abdul Rafae Noor , Akash Kothari , Zheyu Li , Mahbod Afarin , Tajana Rosing , Vikram Adve This is my paper

Pith reviewed 2026-06-26 02:37 UTC · model grok-4.3

classification 💻 cs.PL cs.CLcs.PF
keywords hyperdimensional computingapproximation tuningcompiler optimizationdomain-specific approximationsretargetable compilationHDC workloadshardware acceleration
0
0 comments X

The pith

ApproxHDC automatically identifies and applies domain-specific approximations in hyperdimensional computing workloads to enable performance gains across CPUs, GPUs, and memory accelerators with minimal accuracy loss.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents ApproxHDC as a compiler framework that automates the selection of approximations tailored to hyperdimensional computing. HDC algorithms tolerate noise by design, so targeted approximations at software and hardware levels can deliver speedups without large accuracy penalties. The framework extends existing compiler infrastructure to search the space of possible approximations efficiently and retarget the results to CPUs, GPUs, and simulated ReRAM and PCM accelerators. A sympathetic reader would care because manual approximation tuning is impractical given the exponential number of options, and automated support could make HDC more deployable on heterogeneous and emerging hardware.

Core claim

ApproxHDC extends the HPVM-HDC compiler infrastructure to enable automated identification and application of domain-specific approximations in HDC workloads. The framework supports retargetable compilation across diverse hardware backends, including CPUs, GPUs, and simulated ReRAM and PCM-based accelerators. It employs efficient search and analysis to navigate the exponentially large space of possible approximations and identify high-impact configurations spanning both software and hardware levels.

What carries the argument

ApproxHDC's search and analysis engine, which navigates the exponential approximation space to locate high-impact configurations at software and hardware levels.

If this is right

  • HDC workloads can be compiled once and executed efficiently on CPUs, GPUs, and simulated memory accelerators without manual approximation decisions.
  • Performance gains become available at both software and hardware levels through a single automated pass.
  • Retargeting to new backends such as ReRAM and PCM requires only backend-specific mapping rather than re-deriving approximations.
  • The tolerance of HDC to approximation is turned into a systematic, compiler-driven optimization rather than an ad-hoc property.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same automated navigation of approximation spaces could be tested on other noise-tolerant paradigms that map to heterogeneous hardware.
  • If the search scales, it could lower the effort needed to port HDC algorithms to custom accelerators built from emerging memory technologies.
  • Validation on physical rather than simulated ReRAM or PCM devices would directly test whether the identified configurations survive real device variation.

Load-bearing premise

Efficient search and analysis can navigate the exponentially large space of approximations to find configurations that deliver high performance impact with minimal accuracy loss.

What would settle it

Applying the configurations found by ApproxHDC to standard HDC benchmarks on the supported backends and measuring either negligible performance improvement or accuracy loss beyond the level tolerated by the target applications.

Figures

Figures reproduced from arXiv: 2606.26547 by Abdul Rafae Noor, Akash Kothari, Mahbod Afarin, Tajana Rosing, Vikram Adve, Xavier Routh, Zheyu Li.

Figure 1
Figure 1. Figure 1: Overview of the ApproxHDC workflow. Given HDC applications written in HDC++ along with an application-specific quality-of-service (QoS) evaluator, ApproxHDC first extracts the set of HDC primitives used in the program. It then performs efficient autotuning to explore approximation choices at both the global and per-operation levels. Finally, ApproxHDC applies the selected approximations via automatically g… view at source ↗
Figure 2
Figure 2. Figure 2: Maximum speedup achieved over the 3-hour tuning budget for all four reduction perforation pruning conditions across all applications and GPU top row, CPU bottom row. Each curve shows the running-best speedup relative to the untuned baseline as new configurations are evaluated. End-only with encoding protection (green) typically finds more valid configurations and reaches higher peak speedups than full redu… view at source ↗
Figure 3
Figure 3. Figure 3: Measured maximum speedup over time for MicroHD’s classification benchmark using ApproxHDC’s approximation tuning framework. MicroHD’s achieved speedup is shown as a horizontal reference line. ApproxHDC identifies superior approximation configurations within 5 minutes and ultimately discovers a configuration that is ap￾proximately 3.5× more efficient than MicroHD. We use a reference configuration of 𝐷=1024,… view at source ↗
Figure 4
Figure 4. Figure 4: ApproxHDC hardware approximation tuning of the HD-Classification benchmark for the SpecPCM [8] Phase Change Memory simulator. ApproxHDC effectively achieves over 4.5x performance improvements within 45 minutes of tuning (and simulation) time. wider ADC_STEP) to cover the active range, driving energy reduction. 6 Related Work Compiler IR for Approximations. EnerJ [28] supports ap￾proximate computing via typ… view at source ↗
read the original 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. 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). HDC algorithms are intrinsically tolerant to noise and approximation, enabling substantial performance gains with minimal accuracy loss. 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 approximations is exponentially large; ApproxHDC employs efficient search and analysis to navigate it and identify high-impact configurations spanning both software and hardware levels.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The manuscript introduces ApproxHDC, a framework extending the HPVM-HDC compiler infrastructure for automated identification and application of domain-specific approximations in Hyperdimensional Computing (HDC) workloads. It supports retargetable compilation across CPUs, GPUs, and simulated ReRAM/PCM accelerators by using efficient search and analysis to navigate the exponentially large approximation space and identify high-impact configurations at software and hardware levels with minimal accuracy loss.

Significance. If the claims regarding tractable navigation of the approximation space hold, the work would be significant for approximate computing and domain-specific compilation. It could facilitate practical deployment of HDC on heterogeneous hardware by automating approximation tuning, leveraging HDC's noise tolerance for performance gains. The extension of an existing compiler infrastructure is a constructive approach toward retargetability.

major comments (1)
  1. [Abstract] Abstract: The central claim that ApproxHDC 'employs efficient search and analysis to navigate' the exponentially large space of approximations and 'identify high-impact configurations' is unsupported by any description of the search procedure (e.g., static analysis, heuristics, sampling), complexity argument, or empirical scaling data on search cost versus HDC dimension or number of sites. This is load-bearing for the practicality of the retargetable compilation benefit.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review and for recognizing the potential significance of ApproxHDC for approximate computing and retargetable compilation. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that ApproxHDC 'employs efficient search and analysis to navigate' the exponentially large space of approximations and 'identify high-impact configurations' is unsupported by any description of the search procedure (e.g., static analysis, heuristics, sampling), complexity argument, or empirical scaling data on search cost versus HDC dimension or number of sites. This is load-bearing for the practicality of the retargetable compilation benefit.

    Authors: We agree that the abstract would be strengthened by briefly indicating the nature of the search procedure. The body of the manuscript (Section 4) describes a hybrid approach that first applies static dependency and sensitivity analysis to prune approximable sites, followed by a beam-search heuristic that uses lightweight profiling to rank configurations. Section 5.3 and Figure 8 present empirical scaling results showing search cost growing linearly with HDC dimension (up to D=10,000) and number of sites. We will revise the abstract to add a concise clause such as "via static sensitivity analysis and beam-search heuristics" so that the central claim is directly supported at the abstract level. revision: yes

Circularity Check

0 steps flagged

No circularity: framework introduction contains no derivations or self-referential reductions

full rationale

The paper introduces ApproxHDC as a new compiler framework extending HPVM-HDC, with claims about efficient search over approximation spaces. No equations, fitted parameters, predictions, or uniqueness theorems appear in the provided text. The central claims rest on the novelty of the framework itself rather than any reduction of outputs to inputs by construction or self-citation chains. This is the common case of a descriptive systems paper with no load-bearing mathematical steps to inspect for circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities beyond the framework name itself; no numerical fits or unstated background assumptions are visible.

invented entities (1)
  • ApproxHDC framework no independent evidence
    purpose: Automated search and application of approximations in HDC
    The framework is the primary new artifact described in the abstract.

pith-pipeline@v0.9.1-grok · 5752 in / 1018 out tokens · 41983 ms · 2026-06-26T02:37:25.574685+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 13 canonical work pages

  1. [1]

    Jason Ansel, Shoaib Kamil, Kalyan Veeramachaneni, Jonathan Ragan- Kelley, Jeffrey Bosboom, Una-May O’Reilly, and Saman Amarasinghe

  2. [2]

    In Proceedings of the 23rd International Conference on Parallel Architectures and Compilation(Edmonton, AB, Canada)(PACT ’14)

    OpenTuner: an extensible framework for program autotuning. In Proceedings of the 23rd International Conference on Parallel Architectures and Compilation(Edmonton, AB, Canada)(PACT ’14). Association for Computing Machinery, New York, NY, USA, 303–316. doi:10.1145/ 2628071.2628092

    arXiv

  3. [3]

    Russel Arbore, Xavier Routh, Abdul Rafae Noor, Akash Kothari, Haichao Yang, Weihong Xu, Sumukh Pinge, Minxuan Zhou, Tajana Rosing, and Vikram Adve. 2025. HPVM-HDC: A Heterogeneous Pro- gramming System for Accelerating Hyperdimensional Computing. In Proceedings of the 52nd Annual International Symposium on Computer Architecture (ISCA ’25). Association for ...

    work page doi:10.1145/3695053.3731095 2025

  4. [4]

    Fatemeh Asgarinejad, Anthony Thomas, and Tajana Rosing. 2020. De- tection of epileptic seizures from surface eeg using hyperdimensional computing. In2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE, 536–540

    2020

  5. [5]

    Michael Carbin, Sasa Misailovic, and Martin C. Rinard. 2013. Ver- ifying quantitative reliability for programs that execute on unreli- able hardware. InProceedings of the 2013 ACM SIGPLAN Interna- tional Conference on Object Oriented Programming Systems Languages & Applications(Indianapolis, Indiana, USA)(OOPSLA ’13). As- sociation for Computing Machi...

    work page doi:10.1145/2509136.2509546 2013

  6. [6]

    En-Jui Chang, Abbas Rahimi, Luca Benini, and An-Yeu Andy Wu. 2019. Hyperdimensional computing-based multimodality emotion recogni- tion with physiological signals. In2019 IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS). IEEE, 137–141

    2019

  7. [7]

    Ron Cole and Mark Fanty. 1994. ISOLET. UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C51G69

    work page doi:10.24432/c51g69 1994

  8. [8]

    Adel Ejjeh, Aaron Councilman, Akash Kothari, Maria Kotsifakou, Leon Medvinsky, Abdul Rafae Noor, Hashim Sharif, Yifan Zhao, Sarita Adve, Sasa Misailovic, et al. 2022. HPVM: Hardware-Agnostic Programming for Heterogeneous Parallel Systems.IEEE Micro42, 5 (2022), 108–117

    2022

  9. [9]

    Keming Fan, Ashkan Moradifirouzabadi, Xiangjin Wu, Zheyu Li, Flavio Ponzina, Anton Persson, Eric Pop, Tajana Rosing, and Mingu Kang. 2024. SpecPCM: A Low-Power PCM-Based In-Memory Com- puting Accelerator for Full-Stack Mass Spectrometry Analysis.IEEE Journal on Exploratory Solid-State Computational Devices and Circuits 10 (2024), 161–169. doi:10.1109/JXCD...

    work page doi:10.1109/jxcdc.2024.3498837 2024

  10. [10]

    Yunhui Guo, Mohsen Imani, Jaeyoung Kang, Sahand Salamat, Justin Morris, Baris Aksanli, Yeseong Kim, and Tajana Rosing. 2021. Hyper- rec: Efficient recommender systems with hyperdimensional comput- ing. InProceedings of the 26th Asia and South Pacific Design Automation Conference. 384–389

    2021

  11. [11]

    Mike Heddes, Igor Nunes, Pere Vergés, Denis Kleyko, Danny Abraham, Tony Givargis, Alexandru Nicolau, and Alexander Veidenbaum. 2023. Torchhd: An Open Source Python Library to Support Research on Hyperdimensional Computing and Vector Symbolic Architectures. arXiv:2205.09208 [cs.LG]https://arxiv.org/abs/2205.09208

    arXiv 2023

  12. [12]

    Mohsen Imani, Yeseong Kim, Thomas Worley, Saransh Gupta, and Tajana Rosing. 2019. HDCluster: An Accurate Clustering Using Brain- Inspired High-Dimensional Computing. In2019 Design, Automation & Test in Europe Conference & Exhibition (DATE). 1591–1594. doi:10. 23919/DATE.2019.8715147

    arXiv 2019

  13. [13]

    Mohsen Imani, Deqian Kong, Abbas Rahimi, and Tajana Rosing. 2017. Voicehd: Hyperdimensional computing for efficient speech recogni- tion. In2017 IEEE international conference on rebooting computing (ICRC). IEEE, 1–8

    2017

  14. [14]

    Pentti Kanerva. 2009. Hyperdimensional computing: An introduction to computing in distributed representation with high-dimensional random vectors.Cognitive computation1 (2009), 139–159

    2009

  15. [15]

    Jaeyoung Kang, You Hak Lee, Minxuan Zhou, Weihong Xu, and Tajana Rosing. 2024. HygHD: Hyperdimensional Hypergraph Learning. In 2024 Design, Automation & Test in Europe Conference & Exhibition (DATE). 1–6. doi:10.23919/DATE58400.2024.10546871

    work page doi:10.23919/date58400.2024.10546871 2024

  16. [16]

    Jaeyoung Kang, Minxuan Zhou, Abhinav Bhansali, Weihong Xu, An- thony Thomas, and Tajana Rosing. 2022. RelHD: A Graph-based Learning on FeFET with Hyperdimensional Computing. In2022 IEEE 40th International Conference on Computer Design (ICCD). 553–560. doi:10.1109/ICCD56317.2022.00087

    work page doi:10.1109/iccd56317.2022.00087 2022

  17. [17]

    Arman Kazemi, Franz Müller, Mohammad Mehdi Sharifi, Hamza Errahmouni, Gerald Gerlach, Thomas Kämpfe, Mohsen Imani, Xiaobo Sharon Hu, and Michael Niemier. 2022. Achieving software-equivalent accuracy for hyperdimensional computing with ferroelectric-based in-memory computing.Scientific reports12, 1 (2022), 19201. 12 Compiler-Driven Approximation Tuning for...

    2022

  18. [18]

    Yeseong Kim, Mohsen Imani, Niema Moshiri, and Tajana Rosing. 2020. Geniehd: Efficient dna pattern matching accelerator using hyperdi- mensional computing. In2020 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 115–120

    2020

  19. [19]

    Rachkovskij, Evgeny Osipov, and Abbas Rahimi

    Denis Kleyko, Dmitri A. Rachkovskij, Evgeny Osipov, and Abbas Rahimi. 2022. A Survey on Hyperdimensional Computing aka Vector Symbolic Architectures, Part I: Models and Data Transformations. ACM Comput. Surv.55, 6, Article 130 (Dec. 2022), 40 pages. doi:10. 1145/3538531

    2022

  20. [20]

    Maria Kotsifakou, Prakalp Srivastava, Matthew D Sinclair, Rakesh Ko- muravelli, Vikram Adve, and Sarita Adve. 2018. Hpvm: Heterogeneous parallel virtual machine. InProceedings of the 23rd ACM SIGPLAN Sym- posium on Principles and Practice of Parallel Programming. 68–80

    2018

  21. [21]

    Chih, Meng-Fan Chang, H.-S

    Haitong Li, Wei-Chen Chen, Akash Levy, Ching-Hua Wang, Hongjie Wang, Po-Han Chen, Weier Wan, Win-San Khwa, Harry Chuang, Y.-D. Chih, Meng-Fan Chang, H.-S. Philip Wong, and Priyanka Raina. 2021. SAPIENS: A 64-kb RRAM-Based Non-Volatile Associative Memory for One-Shot Learning and Inference at the Edge.IEEE Transactions on Electron Devices68, 12 (2021), 663...

    work page doi:10.1109/ted.2021 2021

  22. [22]

    McCallum

    A.K. McCallum. 2000. Automating the construction of internet portals with machine learning.Information Retrieval3 (01 2000), 127–163

    2000

  23. [23]

    Anton Mitrokhin, P Sutor, Cornelia Fermüller, and Yiannis Aloimonos

  24. [24]

    Learning sensorimotor control with neuromorphic sensors: Toward hyperdimensional active perception.Science Robotics4, 30 (2019), eaaw6736

    2019

  25. [25]

    Igor Nunes, Mike Heddes, Tony Givargis, Alexandru Nicolau, and Alex Veidenbaum. 2022. GraphHD: Efficient graph classification using hyperdimensional computing. In2022 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 1485–1490

    2022

  26. [26]

    Flavio Ponzina. 2024. MicroHD.https://github.com/flavio-ponzina/ MicroHD. Accessed: 2026-04-16

    2024

  27. [27]

    Flavio Ponzina and Tajana Rosing. 2024. MicroHD: An Accuracy- Driven Optimization of Hyperdimensional Computing Algorithms for TinyML systems. arXiv:2404.00039 [cs.PF]https://arxiv.org/abs/2404. 00039

    arXiv 2024

  28. [28]

    Abbas Rahimi, Simone Benatti, Pentti Kanerva, Luca Benini, and Jan M Rabaey. 2016. Hyperdimensional biosignal processing: A case study for EMG-based hand gesture recognition. In2016 IEEE International Conference on Rebooting Computing (ICRC). IEEE, 1–8

    2016

  29. [29]

    Abbas Rahimi, Pentti Kanerva, and Jan M Rabaey. 2016. A robust and energy-efficient classifier using brain-inspired hyperdimensional computing. InProceedings of the 2016 international symposium on low power electronics and design. 64–69

    2016

  30. [30]

    Adrian Sampson, Werner Dietl, Emily Fortuna, Danushen Gnanapra- gasam, Luis Ceze, and Dan Grossman. 2011. EnerJ: approximate data types for safe and general low-power computation.SIGPLAN Not.46, 6 (June 2011), 164–174. doi:10.1145/1993316.1993518

    work page doi:10.1145/1993316.1993518 2011

  31. [31]

    Adve, Sasa Misailovic, and Sarita Adve

    Hashim Sharif, Prakalp Srivastava, Muhammad Huzaifa, Maria Kotsi- fakou, Keyur Joshi, Yasmin Sarita, Nathan Zhao, Vikram S. Adve, Sasa Misailovic, and Sarita Adve. 2019. ApproxHPVM: a portable compiler IR for accuracy-aware optimizations.Proc. ACM Program. Lang.3, OOPSLA, Article 186 (Oct. 2019), 30 pages. doi:10.1145/3360612

    work page doi:10.1145/3360612 2019

  32. [32]

    Adve, Sasa Misailovic, and Sarita Adve

    Hashim Sharif, Yifan Zhao, Maria Kotsifakou, Akash Kothari, Ben Schreiber, Elizabeth Wang, Yasmin Sarita, Nathan Zhao, Keyur Joshi, Vikram S. Adve, Sasa Misailovic, and Sarita Adve. 2021. Approx- Tuner: a compiler and runtime system for adaptive approximations. InProceedings of the 26th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programm...

    work page doi:10.1145/3437801.3446108 2021

  33. [33]

    Weihong Xu, Jaeyoung Kang, and Tajana Rosing. 2023. FSL-HD: Accelerating Few-Shot Learning on ReRAM using Hyperdimensional Computing. In2023 Design, Automation & Test in Europe Conference & Exhibition (DATE). 1–6. doi:10.23919/DATE56975.2023.10136901

    work page doi:10.23919/date56975.2023.10136901 2023

  34. [34]

    Tinaqi Zhang, Sahand Salamat, Behnam Khaleghi, Justin Morris, Baris Aksanli, and Tajana Simunic Rosing. 2023. HD2FPGA: Automated Framework for Accelerating Hyperdimensional Computing on FPGAs. In2023 24th International Symposium on Quality Electronic Design (ISQED). 1–9. doi:10.1109/ISQED57927.2023.10129332

    work page doi:10.1109/isqed57927.2023.10129332 2023

  35. [35]

    Zhuowen Zou, Hanning Chen, Prathyush Poduval, Yeseong Kim, Mahdi Imani, Elaheh Sadredini, Rosario Cammarota, and Mohsen Imani. 2022. BioHD: an efficient genome sequence search platform using HyperDimensional memorization. InProceedings of the 49th Annual International Symposium on Computer Architecture(New York, New York)(ISCA ’22). Association for Comput...

    work page doi:10.1145/3470496.3527422 2022