Low-Energy Reduced RISC-V Instruction Subset Processor for Tsetlin Machine Inference at the Edge

Chanda Gupta; Himani Panwar; Rishad Shafik; Sanidhya Bhatia; Shaurya Priyadarshi; Sudip Roy

arxiv: 2606.19964 · v1 · pith:YUPZJH3Inew · submitted 2026-06-18 · 💻 cs.LG · cs.AR

Low-Energy Reduced RISC-V Instruction Subset Processor for Tsetlin Machine Inference at the Edge

Chanda Gupta , Sanidhya Bhatia , Shaurya Priyadarshi , Himani Panwar , Rishad Shafik , Sudip Roy This is my paper

Pith reviewed 2026-06-26 18:15 UTC · model grok-4.3

classification 💻 cs.LG cs.AR

keywords Tsetlin MachineRISC-V processoredge inferenceinstruction subset reductionenergy efficiencybinarized neural networksprogrammable corebitwise operations

0 comments

The pith

A reduced-instruction RISC-V processor for Tsetlin Machine inference achieves 29.7 times lower energy use while matching or beating binarized neural network accuracy.

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

The paper develops a domain-specific RISC-V microprocessor by profiling TM workloads, removing unused instructions, and simplifying the datapath and control logic to match TM's bitwise and automata operations. This keeps the core programmable without external hosts or microcode, unlike prior accelerator designs. Evaluation on multiple datasets shows TM reaching up to 88.18 percent accuracy on CIFAR-2 versus 60 percent for BNNs, with execution time cut by as much as 98 percent and average energy reduced by 29.7 times. A sympathetic reader cares because the result offers a middle path between rigid accelerators and general-purpose cores for edge logic-based machine learning.

Core claim

By profiling instructions on TM inference programs and then pruning the instruction set plus datapath simplifications, the authors produce a smaller RV32IM-derived core that retains full programmability for Tsetlin Machine workloads while delivering the reported accuracy, speed, and energy gains over both a baseline RV32IM core and BNN implementations.

What carries the argument

The reduced instruction subset processor obtained by instruction profiling followed by datapath and control-path simplifications tailored to TM bitwise operations and finite-state automata.

If this is right

TM inference becomes feasible on a fully programmable core without tight host coupling or microcode.
Energy for edge TM deployments falls by an average factor of 29.7 while accuracy stays comparable to or higher than BNN baselines.
Execution time for the same TM tasks drops by up to 98 percent across the evaluated datasets.
The same profiling-plus-simplification flow can be repeated for other logic-based or automata-driven ML models.

Where Pith is reading between the lines

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

The reduced core could serve as a template for adding TM-specific custom instructions inside a standard RISC-V toolchain without losing compatibility.
Energy savings observed on simulation may translate to real silicon only after accounting for memory access patterns that were not re-profiled.
Designers of other edge processors might apply the same subset-reduction method to workloads dominated by bitwise logic rather than arithmetic.

Load-bearing premise

Instruction profiling on the chosen datasets fully captures every TM inference program that will ever run, and the resulting simplifications preserve functional correctness and accuracy without needing extra host code or microcode.

What would settle it

Execute a Tsetlin Machine program on the reduced core that exercises an instruction removed during profiling and observe whether accuracy drops or the program fails to complete.

Figures

Figures reproduced from arXiv: 2606.19964 by Chanda Gupta, Himani Panwar, Rishad Shafik, Sanidhya Bhatia, Shaurya Priyadarshi, Sudip Roy.

**Figure 1.** Figure 1: Block diagram of a multiclass Tsetlin machine. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 3.** Figure 3: (a) Instruction count comparison between T1 and T2, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Hardware-software co-design approach for domain [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Figure 5: Execution time (in ms) across multiple inference scenarios for datasets. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 6.** Figure 6: Energy Consumption (in mJ) vs. number of clauses per class across datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗

read the original abstract

Tsetlin Machine (TM) is a logic-based machine learning approach that relies on simple bitwise operations and finite-state automata, which makes it attractive for edge AI deployments. Recent work has focused on co-processor and accelerator designs based on Tsetlin Machines (TMs). Although these designs achieve high performance, they typically depend on tightly coupled interfaces, microcode-style programming, and external host processors, limiting flexibility and ease of programming. In this work, we present a domain-specific RISC-V microprocessor architecture and design flow tailored for TM inference. Leveraging the modular structure of RISC-V, we design a reduced instruction subset processor that retains programmability while targeting improved performance and lower energy consumption for TM workloads. Instruction profiling is employed to guide instruction reduction, followed by datapath and control path simplifications tailored to TM inference. Both the baseline RV32IM core and the proposed reduced core are evaluated across multiple datasets and compared with Binarized Neural Networks (BNNs), which serve as a hardware-efficient baseline due to their reliance on bitwise operations during inference. Results show that TM achieves comparable or higher accuracy (e.g., up to 88.18% on CIFAR-2 compared to 60.0% for BNN) while reducing execution time by up to 98% across multiple datasets. Furthermore, the proposed design achieves an average $29.7\times$ reduction in energy consumption, demonstrating its effectiveness for programmable and efficient edge AI systems.

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.

Desk Editor's Note private letter to a colleague

The paper delivers a profiled reduced-instruction RISC-V core for TM inference with solid reported energy and speed gains over RV32IM, but the pruning step's guarantee of functional equivalence rests on limited dataset coverage.

read the letter

The core contribution is a RISC-V design flow that profiles TM inference workloads on chosen datasets, drops unused instructions from RV32IM, and simplifies the datapath and control logic while keeping the core programmable. This yields the headline numbers: up to 98% shorter execution time and 29.7× lower energy on average, with TM accuracy holding or beating BNN baselines on the tested sets.

What works is the direct application of RISC-V modularity to a narrow but real workload. The comparisons to both the full RV32IM core and BNN hardware are concrete and the energy claim is the kind of number that matters for edge deployment. The design stays general enough that a programmer can still write TM code without a host coprocessor.

The soft spot is exactly the one the stress test flags. Instruction reduction is driven by profiling on specific datasets; nothing in the abstract shows an exhaustive test suite or formal check that every valid TM program produces identical clause and automaton results on the reduced core. If an unprofiled addressing mode or control interaction appears, the accuracy and energy figures could be invalid for that program. The paper would be stronger with either a machine-checked equivalence argument or a broader verification pass.

This is useful reading for anyone building edge logic-ML hardware. It is narrow enough that it will not change general computer architecture, but the implementation details and numbers are worth seeing in full. It deserves a serious referee because the engineering is reproducible in principle and the claims are falsifiable with the right test cases.

Referee Report

1 major / 0 minor

Summary. The manuscript proposes a domain-specific reduced RISC-V processor for Tsetlin Machine (TM) inference at the edge. Instruction profiling on TM workloads guides pruning of the RV32IM subset, followed by datapath and control-path simplifications. The design retains programmability and is evaluated against a baseline RV32IM core and Binarized Neural Network (BNN) baselines across multiple datasets, claiming up to 98% execution-time reduction, average 29.7× energy reduction, and accuracy that is comparable or higher than BNNs (e.g., 88.18% vs. 60.0% on CIFAR-2).

Significance. If the reduced core produces functionally equivalent results, the work supplies a programmable, low-energy alternative to tightly coupled TM accelerators or microcode-based co-processors. The direct head-to-head comparisons with both a standard RV32IM core and BNNs on several datasets provide a concrete basis for assessing the claimed efficiency gains.

major comments (1)

[Design flow / instruction reduction and evaluation sections] The headline accuracy and energy claims rest on the assumption that the profiling-driven instruction subset and datapath simplifications produce bit-equivalent clause and automaton outputs for every valid TM inference program. The manuscript does not describe an exhaustive test suite, formal equivalence check, or coverage argument showing that omitted instructions, addressing modes, or control interactions never arise outside the profiled datasets; without this, the reported 88.18% CIFAR-2 accuracy and 29.7× energy figure cannot be treated as verified.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comment point-by-point below, providing clarification on our verification approach and committing to revisions where appropriate.

read point-by-point responses

Referee: [Design flow / instruction reduction and evaluation sections] The headline accuracy and energy claims rest on the assumption that the profiling-driven instruction subset and datapath simplifications produce bit-equivalent clause and automaton outputs for every valid TM inference program. The manuscript does not describe an exhaustive test suite, formal equivalence check, or coverage argument showing that omitted instructions, addressing modes, or control interactions never arise outside the profiled datasets; without this, the reported 88.18% CIFAR-2 accuracy and 29.7× energy figure cannot be treated as verified.

Authors: We acknowledge this is a valid concern for ensuring the claims are robust. Our instruction reduction was derived from dynamic profiling of TM inference workloads across all evaluated datasets (including CIFAR-2), which showed that TM clause evaluation and automaton updates exclusively use a narrow set of RV32IM operations: bitwise AND/XOR, comparisons, and limited arithmetic for state transitions. The TM algorithm's structure (deterministic bitwise clause computation and finite-state updates) inherently excludes many instructions such as multiplication, division, and complex addressing modes. To strengthen the manuscript, we will revise the design flow and evaluation sections to: (1) explicitly list all pruned instructions with profiling coverage metrics, (2) report results from executing the reduced core on additional TM programs generated with varied hyperparameters and random seeds not used in initial profiling, and (3) provide a coverage argument based on the TM inference pseudocode demonstrating that no omitted instructions or control paths can be reached in valid programs. These additions will support the bit-equivalence assumption without requiring formal methods. revision: yes

Circularity Check

0 steps flagged

No circularity; results from direct simulation on profiled workloads with external baselines

full rationale

The paper's core flow (profiling RV32IM instructions on TM datasets, pruning the subset, simplifying datapath/control, then measuring execution time/energy/accuracy on the same datasets vs. BNN baselines) contains no equations, fitted parameters renamed as predictions, or self-citation chains that reduce any headline claim to its inputs by construction. All reported numbers (e.g., 98% time reduction, 29.7× energy reduction, accuracy figures) are presented as outcomes of explicit implementation/simulation, not derived quantities. This is the normal case of a domain-specific hardware design evaluated against external references.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The design rests on standard computer-architecture assumptions about workload-specific ISA reduction and the suitability of TM for bitwise operations; no free parameters, invented entities, or ad-hoc axioms are introduced in the abstract.

axioms (1)

domain assumption RISC-V ISA modularity allows safe removal of unused instructions for a target workload without breaking TM inference correctness.
Invoked to justify creating the reduced core after instruction profiling.

pith-pipeline@v0.9.1-grok · 5821 in / 1332 out tokens · 24579 ms · 2026-06-26T18:15:46.095552+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 2 canonical work pages · 1 internal anchor

[1]

Hardware Accelerator for MobileViT Vision Transformer with Reconfigurable Computation,

S.-F. Hsiao, T.-H. Chao, Y .-C. Yuan, and K.-C. Chen, “Hardware Accelerator for MobileViT Vision Transformer with Reconfigurable Computation,” inProc. of the ISCAS, 2024, pp. 1–4

2024
[2]

Tsetlin Machine-Based Image Classification FPGA Accelerator With On-Device Training,

S. A. Tunheim, L. Jiao, R. Shafik, A. Yakovlev, and O.-C. Granmo, “Tsetlin Machine-Based Image Classification FPGA Accelerator With On-Device Training,”IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 72, no. 2, pp. 830–843, 2025

2025
[3]

1.1 The Deep Learning Revolution and its Implications for Computer Architecture and Chip Design,

J. Dean, “1.1 The Deep Learning Revolution and its Implications for Computer Architecture and Chip Design,” inProc. of the ISSCC, 2024, pp. 8–14

2024
[4]

Fast and Compact Tsetlin Machine Inference on CPUs Using Instruction-Level Optimization,

Y . Zeng, S. Duan, R. Shafik, and A. Yakovlev, “Fast and Compact Tsetlin Machine Inference on CPUs Using Instruction-Level Optimization,” in Proc. of the ISTM, 2025, pp. 44–47

2025
[5]

RISC-V Instruction Set Architecture Extensions: A Survey,

E. Cui, T. Li, and Q. Wei, “RISC-V Instruction Set Architecture Extensions: A Survey,”IEEE Access, vol. 11, pp. 24 696–24 711, 2023

2023
[6]

Granmo, The Tsetlin machine – a game theoretic bandit driven approach to optimal pattern recognition with proposi- tional logic (2021), arXiv:1804.01508 [cs.AI]

O. C. Granmo, “The Tsetlin Machine – A Game Theoretic Bandit Driven Approach to Optimal Pattern Recognition with Propositional Logic,” 2021. [Online]. Available: https://arxiv.org/abs/1804.01508

work page arXiv 2021
[7]

ETHEREAL: Energy-efficient and High-throughput Inference using Compressed Tsetlin Machine,

S. Duan, R. Shafik, and A. Yakovlev, “ETHEREAL: Energy-efficient and High-throughput Inference using Compressed Tsetlin Machine,” in Proc. of the IWASI, 2025, pp. 1–6

2025
[8]

Learning Dynamics, Pattern Recognition Capability and Interpretability of the Tsetlin Machine,

O. Tarasyuk, A. Gorbenko, T. Rahman, L. Jiao, O.-C. Granmo, R. Shafik, and A. Yakovlev, “Learning Dynamics, Pattern Recognition Capability and Interpretability of the Tsetlin Machine,”Pattern Recognition, vol. 174, p. 113028, 2026

2026
[9]

Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1

M. Courbariaux, I. Hubara, D. Soudry, R. El-Yaniv, and Y . Bengio, “Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to + 1 or-1,” 2016. [Online]. Available: https://arxiv.org/abs/1602.02830

work page internal anchor Pith review Pith/arXiv arXiv 2016
[10]

RISC-V GNU Toolchain,

RISC-V Collab, “RISC-V GNU Toolchain,” [Online]. Available: https://github.com/riscv-collab/riscv-gnu-toolchain, 2025, [Accessed:Feb,2026]

2025
[11]

REDRESS: Generating Compressed Models for Edge Inference Using Tsetlin Machines,

S. Maheshwari, T. Rahman, R. Shafik, A. Yakovlev, A. Rafiev, L. Jiao, and O.-C. Granmo, “REDRESS: Generating Compressed Models for Edge Inference Using Tsetlin Machines,”IEEE TPAMI, vol. 45, no. 9, pp. 11 152–11 168, 2023

2023
[12]

Available: https://github.com/cair/tmu/tree/main, 2025, [Accessed:Feb,2026]

Cair, “TMU,” [Online]. Available: https://github.com/cair/tmu/tree/main, 2025, [Accessed:Feb,2026]

2025
[13]

Available: https://github.com/ultraembedded/riscv, 2025, [Accessed:Feb,2026]

Ultraembedded, “RISC-V,” [Online]. Available: https://github.com/ultraembedded/riscv, 2025, [Accessed:Feb,2026]

2025
[14]

CIFAR-2,

A. Krizhevsky, V . Nair, and G. Hinton, “CIFAR-2,” [Online]. Available: https://keras.io/api/datasets/cifar10/, 2017, [Accessed:Feb,2026]

2017
[15]

Statlog (Vehicle Silhouettes),

P. Mowforth and B. Shepherd, “Statlog (Vehicle Silhouettes),” UCI Machine Learning Repository, [Online]. Available: https://archive.ics.uci.edu/ml/datasets/Statlog+(Vehicle+Silhouettes), 2023, [Accessed:Feb,2026]

2023
[16]

Gesture Phase Segmentation,

R. Madeo, P. Wagner, and S. Peres, “Gesture Phase Segmentation,” UCI Machine Learning Repository, [Online]. Available: https://archive.ics.uci.edu/ml/datasets/Gesture+Phase+Segmentation, 2014, [Accessed: Feb, 2026]

2014
[17]

Gas Sensor Array Drift Dataset,

A. Vergara, “Gas Sensor Array Drift Dataset,” UCI Machine Learning Repository, [Online]. Available: https://archive.ics.uci.edu/ml/datasets/Gas+Sensor+Array+Drift+Dataset, 2012, [Accessed: Feb, 2026]

2012
[18]

EMG Data for Gestures,

N. Krilova, I. Kastalskiy, V . Kazantsev, V . A. Makarov, and S. Lobov, “EMG Data for Gestures,” UCI Machine Learning Repository, [Online]. Available: https://archive.ics.uci.edu/ml/datasets/EMG+data+for+gestures, 2019, [Accessed: Feb, 2026]

2019
[19]

Fashion-MNIST,

H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-MNIST,” Keras, [Online]. Available: https://keras.io/api/datasets/fashion mnist/, 2017, [Accessed: Feb, 2026]

2017

[1] [1]

Hardware Accelerator for MobileViT Vision Transformer with Reconfigurable Computation,

S.-F. Hsiao, T.-H. Chao, Y .-C. Yuan, and K.-C. Chen, “Hardware Accelerator for MobileViT Vision Transformer with Reconfigurable Computation,” inProc. of the ISCAS, 2024, pp. 1–4

2024

[2] [2]

Tsetlin Machine-Based Image Classification FPGA Accelerator With On-Device Training,

S. A. Tunheim, L. Jiao, R. Shafik, A. Yakovlev, and O.-C. Granmo, “Tsetlin Machine-Based Image Classification FPGA Accelerator With On-Device Training,”IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 72, no. 2, pp. 830–843, 2025

2025

[3] [3]

1.1 The Deep Learning Revolution and its Implications for Computer Architecture and Chip Design,

J. Dean, “1.1 The Deep Learning Revolution and its Implications for Computer Architecture and Chip Design,” inProc. of the ISSCC, 2024, pp. 8–14

2024

[4] [4]

Fast and Compact Tsetlin Machine Inference on CPUs Using Instruction-Level Optimization,

Y . Zeng, S. Duan, R. Shafik, and A. Yakovlev, “Fast and Compact Tsetlin Machine Inference on CPUs Using Instruction-Level Optimization,” in Proc. of the ISTM, 2025, pp. 44–47

2025

[5] [5]

RISC-V Instruction Set Architecture Extensions: A Survey,

E. Cui, T. Li, and Q. Wei, “RISC-V Instruction Set Architecture Extensions: A Survey,”IEEE Access, vol. 11, pp. 24 696–24 711, 2023

2023

[6] [6]

Granmo, The Tsetlin machine – a game theoretic bandit driven approach to optimal pattern recognition with proposi- tional logic (2021), arXiv:1804.01508 [cs.AI]

O. C. Granmo, “The Tsetlin Machine – A Game Theoretic Bandit Driven Approach to Optimal Pattern Recognition with Propositional Logic,” 2021. [Online]. Available: https://arxiv.org/abs/1804.01508

work page arXiv 2021

[7] [7]

ETHEREAL: Energy-efficient and High-throughput Inference using Compressed Tsetlin Machine,

S. Duan, R. Shafik, and A. Yakovlev, “ETHEREAL: Energy-efficient and High-throughput Inference using Compressed Tsetlin Machine,” in Proc. of the IWASI, 2025, pp. 1–6

2025

[8] [8]

Learning Dynamics, Pattern Recognition Capability and Interpretability of the Tsetlin Machine,

O. Tarasyuk, A. Gorbenko, T. Rahman, L. Jiao, O.-C. Granmo, R. Shafik, and A. Yakovlev, “Learning Dynamics, Pattern Recognition Capability and Interpretability of the Tsetlin Machine,”Pattern Recognition, vol. 174, p. 113028, 2026

2026

[9] [9]

Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1

M. Courbariaux, I. Hubara, D. Soudry, R. El-Yaniv, and Y . Bengio, “Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to + 1 or-1,” 2016. [Online]. Available: https://arxiv.org/abs/1602.02830

work page internal anchor Pith review Pith/arXiv arXiv 2016

[10] [10]

RISC-V GNU Toolchain,

RISC-V Collab, “RISC-V GNU Toolchain,” [Online]. Available: https://github.com/riscv-collab/riscv-gnu-toolchain, 2025, [Accessed:Feb,2026]

2025

[11] [11]

REDRESS: Generating Compressed Models for Edge Inference Using Tsetlin Machines,

S. Maheshwari, T. Rahman, R. Shafik, A. Yakovlev, A. Rafiev, L. Jiao, and O.-C. Granmo, “REDRESS: Generating Compressed Models for Edge Inference Using Tsetlin Machines,”IEEE TPAMI, vol. 45, no. 9, pp. 11 152–11 168, 2023

2023

[12] [12]

Available: https://github.com/cair/tmu/tree/main, 2025, [Accessed:Feb,2026]

Cair, “TMU,” [Online]. Available: https://github.com/cair/tmu/tree/main, 2025, [Accessed:Feb,2026]

2025

[13] [13]

Available: https://github.com/ultraembedded/riscv, 2025, [Accessed:Feb,2026]

Ultraembedded, “RISC-V,” [Online]. Available: https://github.com/ultraembedded/riscv, 2025, [Accessed:Feb,2026]

2025

[14] [14]

CIFAR-2,

A. Krizhevsky, V . Nair, and G. Hinton, “CIFAR-2,” [Online]. Available: https://keras.io/api/datasets/cifar10/, 2017, [Accessed:Feb,2026]

2017

[15] [15]

Statlog (Vehicle Silhouettes),

P. Mowforth and B. Shepherd, “Statlog (Vehicle Silhouettes),” UCI Machine Learning Repository, [Online]. Available: https://archive.ics.uci.edu/ml/datasets/Statlog+(Vehicle+Silhouettes), 2023, [Accessed:Feb,2026]

2023

[16] [16]

Gesture Phase Segmentation,

R. Madeo, P. Wagner, and S. Peres, “Gesture Phase Segmentation,” UCI Machine Learning Repository, [Online]. Available: https://archive.ics.uci.edu/ml/datasets/Gesture+Phase+Segmentation, 2014, [Accessed: Feb, 2026]

2014

[17] [17]

Gas Sensor Array Drift Dataset,

A. Vergara, “Gas Sensor Array Drift Dataset,” UCI Machine Learning Repository, [Online]. Available: https://archive.ics.uci.edu/ml/datasets/Gas+Sensor+Array+Drift+Dataset, 2012, [Accessed: Feb, 2026]

2012

[18] [18]

EMG Data for Gestures,

N. Krilova, I. Kastalskiy, V . Kazantsev, V . A. Makarov, and S. Lobov, “EMG Data for Gestures,” UCI Machine Learning Repository, [Online]. Available: https://archive.ics.uci.edu/ml/datasets/EMG+data+for+gestures, 2019, [Accessed: Feb, 2026]

2019

[19] [19]

Fashion-MNIST,

H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-MNIST,” Keras, [Online]. Available: https://keras.io/api/datasets/fashion mnist/, 2017, [Accessed: Feb, 2026]

2017