Pith. sign in

REVIEW 5 major objections 6 minor 37 references

POLARON: Precision-aware On-device Learning and Adaptive Runtime-cONfigurable AI acceleration

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a single runtime-reconfigurable multiply-accumulate engine can span 4/8/16-bit fixed-point, floating-point, and posit formats with accuracy within 1.8% of FP32.

desk verdict A plausible unified MAC design, but the headline accuracy claims for RNN/RL/Transformer workloads are explicitly marked as future work in the text, and the comparison tables have inconsistencies. read the letter →

arxiv 2506.08785 v1 pith:5TMAFHCP submitted 2025-06-10 cs.AR cs.AIcs.CCeess.IV

Mukul Lokhande , Santosh Kumar Vishvakarma This is my paper
classification cs.ARcs.AIcs.CCeess.IV
keywords deeplearningacceleratorstrans-precisioncomputingSIMDprocessingelementspositprecisionfloating-pointmultiply-accumulatevariable-precisionquantizationedgeAIon-device
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is trying to establish that a single multiply-accumulate (MAC) engine, with a datapath that can be reconfigured at runtime, can cover the precision formats edge AI actually needs—4/8/16-bit fixed-point, floating-point, and posit—without the silicon cost of separate datapaths. If true, this matters because edge devices are energy- and area-constrained, and most accelerators either fix one precision or duplicate hardware for several. The proposed PARV-CE engine, embedded in the POLARON accelerator, is reported to achieve up to 2× better power–delay product, 3× less resource usage, and 4× better energy efficiency than state-of-the-art designs while keeping accuracy within 1.8% of a full FP32 baseline. A companion quantization scheme, WILD-QLite, assigns precision layer by layer according to sensitivity, which is what makes the accuracy result plausible.

What carries the argument

The load-bearing object is the PARV-CE pipeline, a five-stage SIMD multiply-accumulate engine whose datapath is morphable across precisions. It unpacks operands into sign, exponent/regime, and mantissa fields; multiplies with a bank of 16 reconfigurable 4-bit modified radix-2 Booth multipliers, a standard digit-recoding multiplication scheme; accumulates through a two-stage adder tree with carry-save adders and a Kulisch-style wide accumulator that avoids intermediate rounding; and normalizes using a leading-zero anticipator with round-toward-positive rounding. Because every precision mode shares this datapath, switching from 4-bit fixed-point to posit16 or FP16 only changes how operand fields are routed and combined. The WILD-QLite quantization algorithm supplies the layer-adaptive choices: a sensitivity metric based on weight quantization error and gradient magnitude, plus learnable PACT clipping, decides which layers run at which width.

What would settle it

Run the RNN, RL, and Transformer workloads on the FPGA prototype (or a fabricated ASIC) and compare their accuracy, power, and resource use against both the FP32 baseline and the Python-emulation numbers; if accuracy degrades by more than 1.8% or the RTL and emulation disagree on identical inputs, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that PARV-CE is the first runtime-adaptable trans-precision vector compute element: one five-stage, SIMD-enabled MAC pipeline that handles variable 4/8/16-bit fixed-point, FP8/16, BF16, and posit8/16 by morphing a fused Posit-16 datapath into narrower lanes. Posit is an alternative floating-point format with a variable-length exponent field. Sixteen 4-bit modified radix-2 Booth multipliers can be combined into 8-bit or 16-bit operand widths, giving up to 16× throughput for the narrowest formats while keeping near-100% hardware utilization. Compared against state-of-the-art MAC units and accelerators on FPGA and 28 nm ASIC, the paper reports up to 2× better power–delay product, 3× lower resource use, and up to 4× better energy efficiency, with accuracy within 1.8% of the FP32 baseline across DNN, RNN, RL, and Transformer workloads.

Load-bearing premise

The load-bearing premise is that the Python software emulation of the custom MAC arithmetic faithfully represents how the hardware will actually behave, so the claimed near-FP32 accuracy for RNN, RL, and Transformer workloads would hold when those workloads run on real silicon.

Editorial extensions

If this is right

  • A single accelerator can serve DNN inference, RNN/LSTM, reinforcement learning, and Transformer workloads by reconfiguring precision per layer, eliminating separate MAC arrays for each format.
  • Edge devices could run on-device training, not just inference, because the same datapath supports posit and floating-point accumulation at up to 16× throughput for narrow operands.
  • The reported 3× resource reduction and 2× power–delay improvement mean more MACs can fit on a chip, or the same workload can run with smaller area and lower energy.
  • An accuracy budget under 1.8% means low-precision edge deployment can stay within a couple of percentage points of a full FP32 model for the evaluated networks.
  • Runtime layer-adaptive precision can be paired with early-exit logic, so easy inputs skip layers and compound the energy savings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the software emulation faithfully matches the RTL, the same unified-datapath approach could extend to newer formats such as micro-scaling or bfloat variants, since the pipeline already separates mantissa and exponent handling.
  • Editorial inference: the paper's accuracy claims for RNNs, RL, and Transformers are based on emulation rather than hardware evaluation; running those workloads on the FPGA prototype is a concrete test of whether the 1.8% bound holds end to end.
  • Editorial inference: the layer-sensitivity metric could be converted into an online controller that reassigns precision during training, but the paper only evaluates it as an offline allocation, so a runtime-adaptive training loop is a natural next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes POLARON, an edge-AI accelerator built around a new multi-precision MAC engine (PARV-CE) that supports 4/8/16-bit fixed-point, FP8/BF16/FP16, and Posit8/16 formats in a single reconfigurable datapath. It also introduces WILD-QLite, a layer-adaptive precision/quantization framework, and reports FPGA and ASIC synthesis results claiming up to 2x PDP improvement, 3x resource reduction, and 4x energy-efficiency gains over state-of-the-art designs, together with accuracy within 1.8% of an FP32 baseline across DNNs, RNNs, RL, and Transformers. The hardware contributions are presented through a five-stage pipelined MAC design, a parameterized quantization scheme with a sensitivity metric, and a system-level accelerator architecture with runtime precision control.

Significance. If the performance and accuracy claims were fully supported, the paper would be a useful contribution: a unified multi-format MAC datapath with runtime precision reconfiguration is a relevant direction for edge AI, and the joint design of a quantization algorithm with a reconfigurable SIMD pipeline is a sensible co-design goal. The authors also provide concrete Verilog, FPGA, and ASIC synthesis results, which is a strength. However, the significance is currently undermined by the mismatch between the claimed workload coverage and the evidence: only two CNN models are evaluated, the RNN/RL/Transformer accuracy figures are explicitly deferred to future work, and the comparison baselines are not properly identified. As presented, the central claims of cross-workload accuracy and superiority over SoTA cannot be assessed.

major comments (5)
  1. [Section III (final paragraph), abstract, conclusion] The paper explicitly defers the accuracy evidence for most of the claimed workload classes. The abstract and conclusion state that accuracy is retained within 1.8% of an FP32 baseline across DNNs, RNNs, RL, and Transformer models, but the only quantitative statement for non-CNN workloads is: 'The inference accuracy for RNNs, LSTMs stays within 1.2% compared to FP32 and RL between 1%. We mark this as a opportunity for future exploration.' No Transformer accuracy is reported anywhere. This is a direct contradiction, not a mere presentation issue, and it is load-bearing because the paper's value proposition is precision-adaptive execution that preserves accuracy across diverse workloads.
  2. [Table VI, Section III] The accuracy results cover only LeNet-5 (98.3% vs 98.6% baseline) and VGG-16 (68.7%, identical to the baseline). For VGG-16 the proposed quantized design shows zero accuracy change, and for LeNet-5 the difference is 0.3 percentage points. The claimed 'within 1.8%' bound is therefore not established even for the two evaluated models, and it is not clear how the 1.8% figure was derived. The table also does not report any variance, training hyperparameters, or the number of runs, so the reader cannot judge whether the differences are meaningful.
  3. [Tables II, III, VI; Section III] The 'Baseline' rows in the comparison tables are never defined. The manuscript states that 'several prior SoTA designs were re-implemented under identical experimental conditions,' but the baseline used in Tables II, III, and VI appears to be the authors' own earlier Flex-PE design (reference [6]), since the data-type ranges (Ad-FxP-4/8/16, Var-FP8/16, Posit8/16) and the title of [6] match. If so, the '3x reduction in resource usage compared to SoTA' claim is partly a comparison against the authors' own prior work, not against an independent state of the art. The paper must identify which published design each row corresponds to and state the re-implementation methodology; otherwise the efficiency claims are not verifiable.
  4. [Section III (Performance Analysis)] The accuracy evaluation methodology is described only as a 'Python-based software emulation of the custom MAC arithmetic' using FxP-Math and QKeras on an NVIDIA V100. No details are given about how the emulation matches the hardware arithmetic (rounding, accumulator width, special cases such as subnormals or posit regime overflow), how the quantized models were trained (epochs, learning rates, QAT schedule), or how the 'cross-validation against QuestaSim' was performed. The phrase 'cross-validated' is asserted without presenting any quantitative agreement results. This absence of methodological detail makes the accuracy claims, including the 1.8% bound, unverifiable.
  5. [Section II.B, Eqs. (1)-(2)] The sensitivity metric sl in Eq. (1) is introduced without any validation. The expression mixes a weight-space quantization error term (difference of norms) with a gradient-magnitude term and a layer-size normalization, so its units and interpretation are unclear, and no experiment is reported showing that layers ranked by this metric actually benefit from higher precision. WILD-QLite is also never compared against a standard quantization baseline (e.g., PACT alone, LSQ, or simple uniform quantization) in terms of accuracy versus bit-width trade-off. As the algorithm is central to the claimed accuracy preservation, this missing validation is a load-bearing gap.
minor comments (6)
  1. [Abstract, Section IV] The sentence 'The empirical analysis establish PARVCE incorporated POLARON as a scalable and energy-efficient solution' contains a subject-verb agreement error and should read 'The empirical analyses establish...' or 'The empirical analysis establishes...'.
  2. [Section II.A] There are several typos and grammatical errors, for example 'similiar' (should be 'similar'), 'aligened agaisnt' (should be 'aligned against'), and 'fedding' (should be 'feeding'). These should be corrected.
  3. [Section II.B] The text contains an artifact: '-Awithout requiring retraining from scratch.' The leading '-A' appears to be a formatting error and should be removed.
  4. [Section III] The phrase '0.8W W' contains a duplicated 'W' and should be '0.8 W'. Also, the acronym 'ODT' is used without definition; please expand it at first use.
  5. [References [15], [16], [17]] The patent numbers appear malformed: '7,107/305', '8,051/123', and '10,409/604' should likely be '7,107,305', '8,051,123', and '10,409,604' respectively.
  6. [Table IV] The row 'Mantissa Mult.& Exponent Proc.' has missing entries for several designs (TVLSI'22, TCAS-II'24, TCAD'24). If these values are not applicable or were not reported, please state so explicitly in a footnote.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: accuracy claims are under-supported for RNN/RL/Transformer, but no derivation reduces to its inputs.

full rationale

The paper's derivation chain is not circular. The WILD-QLite quantization equations (1)-(7) are stated in full and their accuracy is evaluated directly on LeNet-5 and VGG-16 in Table VI, while the hardware metrics are synthesis results from Vivado and Design Compiler. There is no equation or fitted parameter that is renamed as a prediction, and no step where a central claim is defined in terms of its own output. Self-citations to the authors' prior Flex-PE work [6] are used as motivation and prior art, not as a load-bearing theorem that forces the present results. The paper does assert accuracy bounds for RNNs, LSTMs, and RL in Section III ('The inference accuracy for RNNs, LSTMs stays within 1.2% compared to FP32 and RL between 1%. We mark this as a opportunity for future exploration.'), and no Transformer accuracy is reported anywhere; however, this is an evidence gap and an internal contradiction, not a circular reduction. The 'Baseline' rows in Tables II, III, and VI are internal comparisons, but nothing indicates they were constructed from or defined by the proposed outcome. In summary, the paper's weaknesses are empirical completeness and claim-evidence alignment, not circular derivation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. Its engineering entities (PARV-CE, POLARON, WILD-QLite) are modular designs, not invented postulates. The free parameters listed are inherent to the quantization algorithm. The axioms are standard domain assumptions or unvalidated claims specific to this work's evaluation methodology.

free parameters (4)
  • Learnable clipping parameter alpha (PACT) = not specified, learned during QAT
    Controls the activation clipping range in Eq. (6-7); adjusted per layer during training to minimize quantization error.
  • Saturation thresholds W_l and W_h = computed from weight distribution, not given
    Define the quantization range in Eqs. (4-5); effectively tuned to the learned weight statistics.
  • Scale factor scale(k) = based on mean(|W|), specific values not given
    Normalizes weights in Eq. (3); depends on the empirical weight distribution and bit-width.
  • Layer-wise bit-width allocation = not specified
    The sensitivity metric in Eq. (1-2) is meant to guide bit-width assignment, but the mapping from sensitivity to bit-width is not defined, likely involving manual tuning or heuristics.
assumptions (5)
  • domain assumption MAC operations account for over 90% of DNN workload operations
    Invoked in the Introduction to justify the focus on MAC units. This is a well-known empirical claim, but it is not derived or verified in the paper.
  • domain assumption Posit-8 is comparable to FP16 and Posit-16 to FP32 in accuracy for AI workloads
    Stated in Section II-B without supporting experiments; used to claim that Posit support covers a wide range of AI tasks.
  • ad hoc to paper The sensitivity metric sl in Eq. (1) correctly measures per-layer quantization impact
    Defined by the authors without validation against known sensitivity measures. The metric is used to justify layer-adaptive precision, but its causal link to final accuracy is assumed.
  • domain assumption The Python emulation on V100 and the RTL implementations are functionally equivalent
    The paper claims cross-validation against QuestaSim, but no verification details or coverage figures are provided. The accuracy results rely entirely on this assumption.
  • domain assumption Re-implementations of prior SoTA designs under 'identical experimental conditions' are fair and accurate
    Needed for the efficiency comparisons; the paper does not describe how these re-implementations were done or verified, so the validity of the improvement ratios is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of POLARON: Precision-aware On-device Learning and Adaptive Runtime-cONfigurable AI acceleration." pith.science (2026). https://pith.science/paper/5TMAFHCP

@misc{pith2026250608785,
  author       = {Pith},
  title        = {Pith review of: POLARON: Precision-aware On-device Learning and Adaptive Runtime-cONfigurable AI acceleration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5TMAFHCP}},
  note         = {Machine review of arXiv:2506.08785}
}
read the original abstract

The increasing complexity of AI models requires flexible hardware capable of supporting diverse precision formats, particularly for energy-constrained edge platforms. This work presents PARV-CE, a SIMD-enabled, multi-precision MAC engine that performs efficient multiply-accumulate operations using a unified data-path for 4/8/16-bit fixed-point, floating point, and posit formats. The architecture incorporates a layer adaptive precision strategy to align computational accuracy with workload sensitivity, optimizing both performance and energy usage. PARV-CE integrates quantization-aware execution with a reconfigurable SIMD pipeline, enabling high-throughput processing with minimal overhead through hardware-software co-design. The results demonstrate up to 2x improvement in PDP and 3x reduction in resource usage compared to SoTA designs, while retaining accuracy within 1.8% FP32 baseline. The architecture supports both on-device training and inference across a range of workloads, including DNNs, RNNs, RL, and Transformer models. The empirical analysis establish PARVCE incorporated POLARON as a scalable and energy-efficient solution for precision-adaptive AI acceleration at edge.

Figures

Figures reproduced from arXiv: 2506.08785 by the authors.

Figure 1
Figure 1. Decades-long growth in hardware FLOPS, DRAM, and interconnect 7 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The detailed micro-architecture of 5-stage Precision-Aware Runtime [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Proposed Precision-aware On-device Learning and Adaptive Runtime [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages

  1. [6]

    Flex-PE: Flexible and SIMD Multiprecision Processing Element for AI Workloads,

    M. Lokhande, G. Raut, and S. K. Vishvakarma, “Flex-PE: Flexible and SIMD Multiprecision Processing Element for AI Workloads,” IEEE Trans. VLSI Syst. , pp. 1–14, 2025

    work page 2025

  2. [1]

    Next-generation domain-specific accelerators: From hard- ware to system,

    Y . S. Shao, “Next-generation domain-specific accelerators: From hard- ware to system,” in IEEE CICC, pp. 1–5, 2024

    work page 2024

  3. [2]

    How to keep pushing ml accelerator performance? know your rooflines!,

    M. Verhelst, L. Benini, and N. Verma, “How to keep pushing ml accelerator performance? know your rooflines!,” IEEE Journal of Solid- State Circuits, pp. 1–18, 2025

    work page 2025

  4. [3]

    Co-Optimization of GPU AI Chip: Technology, Design, System and Algorithms,

    J. R. Hu, L. Liu, et al., “Co-Optimization of GPU AI Chip: Technology, Design, System and Algorithms,” in IEEE IEDM, pp. 1–4, 2024

    work page 2024

  5. [4]

    AI and Memory Wall,

    A. Gholami, Z. Yao, et al., “AI and Memory Wall,” IEEE Micro, vol. 44, no. 3, pp. 33–39, 2024

    work page 2024

  6. [5]

    MEGA.mini: A Universal Generative AI Processor with a New Big/Little Core Architecture for NPU,

    D. Han and A. P. Chandrakasan, “MEGA.mini: A Universal Generative AI Processor with a New Big/Little Core Architecture for NPU,” in IEEE ISSCC, vol. 68, pp. 1–3, 2025

    work page 2025

  7. [7]

    Parallel Accurate Minifloat MACCs for Neural Network Inference on Versal FPGAs,

    H. J. Damsgaard, K. J. Hoßfeld, et al. , “Parallel Accurate Minifloat MACCs for Neural Network Inference on Versal FPGAs,” IEEE Trans. Comput.-Aided Design Integr. Circuits Syst. , pp. 1–1, 2025

    work page 2025

  8. [8]

    A Reconfigurable Processing Element for Multiple- Precision Floating/Fixed-Point HPC,

    B. Li, K. Li, et al., “A Reconfigurable Processing Element for Multiple- Precision Floating/Fixed-Point HPC,” IEEE Trans. Circuits Syst. II , vol. 71, no. 3, pp. 1401–1405, 2024

    work page 2024

Show all 37 references
  1. [9]

    A Low-Cost Floating-Point Dot-Product-Dual- Accumulate Architecture for HPC-Enabled AI,

    H. Tan, L. Huang, et al., “A Low-Cost Floating-Point Dot-Product-Dual- Accumulate Architecture for HPC-Enabled AI,” IEEE Trans. Comput.- Aided Design Integr. Circuits Syst. , vol. 43, no. 2, pp. 681–693, 2024

    2024

  2. [10]

    A Low-Cost Floating-Point FMA Unit Supporting Package Operations for HPC-AI Applications,

    H. Tan, J. Zhang, et al. , “A Low-Cost Floating-Point FMA Unit Supporting Package Operations for HPC-AI Applications,” IEEE Trans. Circuits Syst. II , vol. 71, no. 7, pp. 3488–3492, 2024

    2024

  3. [11]

    RAMAN: A Reconfigurable and Sparse tinyML Accelerator for Inference on Edge,

    A. Krishna, S. Rohit Nudurupati, et al. , “RAMAN: A Reconfigurable and Sparse tinyML Accelerator for Inference on Edge,” IEEE Internet of Things Journal , vol. 11, no. 14, pp. 24831–24845, 2024. TABLE VI EVALUATION METRICS COMPARISON BETWEEN DIVERSE SOTA AI ACCELERATOR DESIGNS...

    2024

  4. [12]

    A Multi-Mode 8k-MAC HW-Utilization- Aware Neural Processing Unit With a Unified Multi-Precision Datapath in 4-nm Flagship Mobile SoC,

    J.-S. Park, C. Park, et al. , “A Multi-Mode 8k-MAC HW-Utilization- Aware Neural Processing Unit With a Unified Multi-Precision Datapath in 4-nm Flagship Mobile SoC,” IEEE Journal of Solid-State Circuits , vol. 58, no. 1, pp. 189–202, 2023

    2023

  5. [13]

    Unified Posit/IEEE-754 Vector MAC Unit for Transprecision Computing,

    L. Crespo, P. Tom ´as, N. Roma, and N. Neves, “Unified Posit/IEEE-754 Vector MAC Unit for Transprecision Computing,” IEEE Trans. Circuits Syst. II, vol. 69, no. 5, pp. 2478–2482, 2022

    2022

  6. [14]

    Multiple-Mode-Supporting Floating- Point FMA Unit for Deep Learning Processors,

    H. Tan, G. Tong, L. Huang, et al., “Multiple-Mode-Supporting Floating- Point FMA Unit for Deep Learning Processors,” IEEE Trans. VLSI Syst., vol. 31, no. 2, pp. 253–266, 2023

    2023

  7. [15]

    Multiply-accumulate unit for single- instruction/multiple-data instructions,

    D. Deng, A. Jebson, et al. , “Multiply-accumulate unit for single- instruction/multiple-data instructions,” Sept. 2006. US Patent 7,107/305

    2006

  8. [16]

    Multipurpose functional unit with combined integer and floating-point multiply-add pipeline,

    M. Siu and S. Oberman, “Multipurpose functional unit with combined integer and floating-point multiply-add pipeline,” Nov. 2011. US Patent 8,051/123

    2011

  9. [17]

    Apparatus and method for performing multiply-and-accumulate-products operations,

    M. A. Kennedy and N. Burgess, “Apparatus and method for performing multiply-and-accumulate-products operations,” Sept. 2019. US Patent 10,409/604

    2019

  10. [18]

    A Configurable Floating-Point Multiple-Precision Processing Element for HPC and AI Converged Computing,

    W. Mao, K. Li, et al., “A Configurable Floating-Point Multiple-Precision Processing Element for HPC and AI Converged Computing,” IEEE Trans. VLSI Syst. , vol. 30, no. 2, pp. 213–226, 2022

    2022

  11. [19]

    TSUNAMI: Triple Sparsity-Aware Ultra Energy- Efficient NN Training Accelerator With Multi-Modal Iterative Pruning,

    S. Kim, J. Lee, et al., “TSUNAMI: Triple Sparsity-Aware Ultra Energy- Efficient NN Training Accelerator With Multi-Modal Iterative Pruning,” IEEE Trans. Circuits Syst. I , vol. 69, no. 4, pp. 1494–1506, 2022

    2022

  12. [20]

    QuaRL: Quantization for fast and envi- ronmentally sustainable reinforcement learning,

    S. Krishnan, M. Lam, et al. , “QuaRL: Quantization for fast and envi- ronmentally sustainable reinforcement learning,” IEEE Trans. Machine Learning Research, 2022

    2022

  13. [21]

    MSDF-Based MAC for Energy-Efficient Neural Networks ,

    S. M. Cherati, M. Barzegar, and L. Sousa, “ MSDF-Based MAC for Energy-Efficient Neural Networks ,” IEEE Trans. VLSI Syst. , pp. 1–12, Feb. 2025

    2025

  14. [22]

    LPRE: Logarithmic Posit-enabled Reconfigurable edge-AI Engine,

    O. Kokane et al. , “LPRE: Logarithmic Posit-enabled Reconfigurable edge-AI Engine,” IEEE ISCAS, 2025

    2025

  15. [23]

    Low-Precision Mixed- Computation Models for Inference on Edge,

    S. Azizi, M. Nazemi, M. Kamal, and M. Pedram, “Low-Precision Mixed- Computation Models for Inference on Edge,” IEEE Trans. Very Large Scale Integr. Syst., vol. 32, p. 1414–1422, Aug. 2024

    2024

  16. [24]

    An Empirical Approach to Enhance Performance for Scalable CORDIC-Based DNNs,

    G. Raut, S. Karkun, and S. K. Vishvakarma, “An Empirical Approach to Enhance Performance for Scalable CORDIC-Based DNNs,” ACM Trans. Reconfigurable Technol. Syst., vol. 16, June 2023

    2023

  17. [25]

    QuantMAC: Enhancing Hardware Performance in DNNs With Quantize Enabled Multiply-Accumulate Unit,

    N. Ashar, G. Raut, V . Treevedi,et al., “QuantMAC: Enhancing Hardware Performance in DNNs With Quantize Enabled Multiply-Accumulate Unit,” IEEE Access, vol. 12, pp. 43600–43614, 2024

    2024

  18. [26]

    A two-stage operand trimming ap- proximate logarithmic multiplier,

    R. Pilipovi ´c, P. Buli´c, and U. Lotri ˇc, “A two-stage operand trimming ap- proximate logarithmic multiplier,” IEEE Trans. Circuits Syst. I , vol. 68, no. 6, pp. 2535–2545, 2021

    2021

  19. [27]

    An Efficient and Flexible Accelerator Design for Sparse CNNs,

    X. Xie, J. Lin, Z. Wang, and J. Wei, “An Efficient and Flexible Accelerator Design for Sparse CNNs,” IEEE Trans. Circuits Syst. I , vol. 68, no. 7, pp. 2936–2949, 2022

    2022

  20. [28]

    ShortcutFusion: From Tensorflow to FPGA- Based Accelerator With a Reuse-Aware Memory Allocation for Shortcut Data,

    D. T. Nguyen, H. Je, et al., “ShortcutFusion: From Tensorflow to FPGA- Based Accelerator With a Reuse-Aware Memory Allocation for Shortcut Data,” IEEE Trans. Circuits Syst. I, vol. 69, no. 6, pp. 2477–2489, 2022

    2022

  21. [29]

    A Real-Time Object Detection Pro- cessor With XNOR-Based Variable-Precision Computing Unit,

    W. Lee, K. Kim, W. Ahn, et al. , “A Real-Time Object Detection Pro- cessor With XNOR-Based Variable-Precision Computing Unit,” IEEE Trans. VLSI Syst. , vol. 31, no. 6, pp. 749–761, 2023

    2023

  22. [30]

    Edge-Side Fine-Grained Sparse CNN Accelerator With Efficient Dynamic Pruning Scheme,

    B. Wu, T. Yu, K. Chen, and W. Liu, “Edge-Side Fine-Grained Sparse CNN Accelerator With Efficient Dynamic Pruning Scheme,” IEEE Trans. Circuits Syst. I , vol. 71, no. 3, pp. 1285–1298, 2024

    2024

  23. [31]

    A high-throughput full-dataflow mo- bilenetv2 accelerator on edge fpga,

    W. Jiang, H. Yu, and Y . Ha, “A high-throughput full-dataflow mo- bilenetv2 accelerator on edge fpga,” IEEE Trans. Comput.-Aided Design Integr. Circuits Syst., vol. 42, no. 5, pp. 1532–1545, 2023

    2023

  24. [32]

    Dedicated FPGA Implementation of the Gaussian TinyYOLOv3 Accelerator,

    S. Ki, J. Park, and H. Kim, “Dedicated FPGA Implementation of the Gaussian TinyYOLOv3 Accelerator,” IEEE Trans. Circuits Syst. II , vol. 70, no. 10, pp. 3882–3886, 2023

    2023

  25. [33]

    A Low-Latency FPGA Accelerator for YOLOv3- Tiny With Flexible Layerwise Mapping and Dataflow,

    M. Kim, K. Oh, et al., “A Low-Latency FPGA Accelerator for YOLOv3- Tiny With Flexible Layerwise Mapping and Dataflow,” IEEE Trans. Circuits Syst. I , vol. 71, no. 3, pp. 1158–1171, 2024

    2024

  26. [34]

    A 29.12- TOPS/W Vector Systolic Accelerator With NAS-Optimized DNNs in 28-nm CMOS,

    K. Li, M. Huang, A. Li, S. Yang, Q. Cheng, and H. Yu, “A 29.12- TOPS/W Vector Systolic Accelerator With NAS-Optimized DNNs in 28-nm CMOS,” IEEE Journal of Solid-State Circuits , pp. 1–12, 2025

    2025

  27. [35]

    A 12.4TOPS/W @ 136GOPS AI-IoT System- on-Chip with 16 RISC-V , 2-to-8b Precision-Scalable DNN Acceleration and 30%-Boost Adaptive Body Biasing,

    F. Conti, D. Rossi, et al., “A 12.4TOPS/W @ 136GOPS AI-IoT System- on-Chip with 16 RISC-V , 2-to-8b Precision-Scalable DNN Acceleration and 30%-Boost Adaptive Body Biasing,” in ISSCC, pp. 21–23, 2023

    2023

  28. [36]

    A 28nm Energy-Area-Efficient Row-based pipelined Training Accelerator with Mixed FXP4/FP16 for On-Device Transfer Learning,

    W. Lu, H.-H. Pei, et al. , “A 28nm Energy-Area-Efficient Row-based pipelined Training Accelerator with Mixed FXP4/FP16 for On-Device Transfer Learning,” in IEEE ISCAS, pp. 1–5, 2024

    2024

  29. [37]

    PL-NPU: An Energy-Efficient Edge-Device DNN Training Processor With Posit-Based Logarithm-Domain Comput- ing,

    Y . Wang, D. Deng, et al., “PL-NPU: An Energy-Efficient Edge-Device DNN Training Processor With Posit-Based Logarithm-Domain Comput- ing,” IEEE Trans. Circuits Syst. I, vol. 69, no. 10, pp. 4042–4055, 2022

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.