Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

XR-NPE: High-Throughput Mixed-precision SIMD Neural Processing Engine for Extended Reality Perception Workloads

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A runtime-reconfigurable SIMD engine for XR perception puts FP4 and Posit formats in one datapath, cutting MAC area 42% and power 38% at 28 nm.

desk verdict The XR-NPE design is a real engineering contribution, but its headline savings claims are overstated—Table II shows the group's own Flex-PE at 5.37 pJ/op, so the '2.85x' figure is not best-of-SoTA. read the letter →

arxiv 2508.13049 v1 pith:24YXT4CY submitted 2025-08-18 cs.AR cs.AIcs.CVeess.IV

Tejas Chaudhari , Akarsh J. , Tanushree Dewangan , Mukul Lokhande , Santosh Kumar Vishvakarma This is my paper · ORCID
classification cs.ARcs.AIcs.CVeess.IV
keywords mixed-precisionSIMDneuralprocessingengineextendedrealitypositarithmeticFP4visual-inertialodometryquantization-awaretraininghardwareaccelerator
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

The paper proposes XR-NPE, a single-instruction, multiple-data (SIMD) neural processing engine that reconfigures itself at runtime among FP4, Posit(4,1), Posit(8,0), and Posit(16,1) formats (Posit is a tapered floating-point format). Its target is the perception stack of extended-reality devices—visual inertial odometry, object classification, and eye-gaze extraction—which the paper says accounts for roughly 60% of XR application runtime. The load-bearing idea is that a reconfigurable mantissa-multiplier datapath, aided by layer-adaptive quantization-aware training, can deliver near-FP32 accuracy while sharply reducing memory bandwidth and energy. If the reported silicon numbers hold, the engine yields 42% less area, 38% less power, and 2.85x improved arithmetic intensity over the best compared state-of-the-art MAC design at 28 nm.

What carries the argument

The load-bearing block is the Reconfigurable Mantissa Multiplication and Exponent processing Circuitry (RMMEC), a Karnaugh-map-based 2-bit multiplier that can be composed into larger mantissa products for Posit(8,0) and Posit(16,1) instead of using separate wide multipliers. Because adders and comparators scale linearly with precision while multipliers and shifters grow exponentially, RMMEC attacks the main source of dark silicon in a mixed-precision datapath: idle multiplier hardware. A precision-select control signal reconfigures the lanes as 4x FP4/Posit(4,1), 2x Posit(8,0), or 1x Posit(16,1), and zero-operand multipliers are power-gated, with exceptions (NaN, infinity, subnormal) handled before the accumulate stage.

What would settle it

Re-synthesize XR-NPE and the cited 28-nm MAC designs in the same cell library at the same voltage, precision mix, and throughput, then measure area, power, and pJ per operation; if the 42% area and 38% power reductions do not survive that head-to-head run, the main efficiency claim fails.

Watch

Extended reading notes

Core claim

The paper claims that one SIMD neural-processing engine can cover 4-bit and 8/16-bit formats—FP4, Posit(4,1), Posit(8,0), Posit(16,1)—by reconfiguring its mantissa multipliers instead of instantiating full-width multipliers for every precision. At 28 nm CMOS it reports a 1.72 GHz operating frequency, 0.016 mm² area, and 14 pJ per operation, with 42% less area and 38% less power than the best compared state-of-the-art MAC design. The hardware is paired with layer-adaptive quantization-aware training that assigns each network layer its own precision; on the UL-VIO model this reduces model size from 13.5 MB (FP32) to 2.42 MB while moving translation and rotation error only 0.72 and 0.13 percentage points, respectively, from the FP32 baseline.

Load-bearing premise

The reported area, power, and energy savings are computed against numbers taken from other papers' tables, so they hold only if those baselines were measured under the same process corner, voltage, datapath width, and definition of an operation as XR-NPE.

Editorial extensions

If this is right

  • The engine is the first reported design to support FP4 and Posit(4,1) together with Posit(8,0) and Posit(16,1) in one runtime-reconfigurable SIMD datapath, so an XR device can choose precision per layer without changing hardware.
  • At 28 nm the reported 1.72 GHz clock, 0.016 mm² area, and 14 pJ per operation imply the design can serve both throughput-heavy DNN layers and lower-precision VIO layers from the same MAC array.
  • The AXI-enabled matrix-multiplication co-processor consumes 1.4x fewer LUTs and 1.77x fewer flip-flops and delivers 1.2x better energy efficiency than the compared FPGA accelerator, pointing to lower integration cost in RISC-V-based XR SoCs.
  • On the UL-VIO workload, mixed FP4/Posit precision keeps the model at 2.42 MB with translation and rotation error within about 0.7 and 0.1 percentage points of FP32, making on-device VIO practical in memory-constrained headsets.
  • The co-processor also reports 23% better energy efficiency and 4% better compute density than prior DNN-based VIO accelerators, suggesting one engine can cover multiple XR perception tasks.

Reading between the lines

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

  • An implicit consequence is that the reconfigurable-mantissa approach may transfer to other tapered or low-precision formats, since the same small-multiplier composition underlies every precision the engine supports.
  • The paper leaves precision assignment trained and layer-fixed; a natural extension would be runtime precision selection driven by input difficulty or remaining battery, which the precision-select mechanism already makes possible.
  • The area and power savings are comparisons against specific prior designs, so the headline percentages should be read as evidence about this design point rather than as a general bound on other MAC architectures.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes XR-NPE, a mixed-precision SIMD neural processing engine supporting FP4, Posit(4,1), Posit(8,0), and Posit(16,1) formats via a reconfigurable mantissa-multiplication/exponent-processing block (RMMEC) and selective power gating. The authors report a 1.72 GHz operating frequency, 0.016 mm^2 area, and 14 pJ/op arithmetic intensity at 28 nm CMOS, claiming 2.85x improved arithmetic intensity, 42% area reduction, and 38% power reduction relative to the best state-of-the-art MAC approaches. They also present an AXI-enabled FPGA co-processor with lower LUT/FF counts and improved energy efficiency for VIO. The software side uses quantization-aware training with a sensitivity-based per-layer precision assignment, reporting accuracy numbers for object classification, eye-gaze, and UL-VIO. The code is publicly released.

Significance. If the quantitative claims were substantiated, XR-NPE would be a useful contribution to the XR-perception accelerator literature: the proposed per-format support (FP4, Posit (4,1), Posit (8,0), Posit (16,1)) is broader than most previous MAC engines, and the RMMEC approach to reduce dark silicon is an interesting architectural idea. The public code release and the demonstration on a real VIO workload are also strong points. However, the headline comparisons are undermined by the paper's own Table II, and the lack of a detailed synthesis methodology prevents independent verification. The central claims therefore need substantial revision before the paper can be accepted.

major comments (4)
  1. [Abstract and Section III, Table II] The claim that XR-NPE provides "2.85x improved arithmetic intensity ... compared to the best of state-of-the-art MAC approaches" is directly contradicted by Table II. The proposed 14 pJ/op is larger (worse) than the 5.37 pJ/op of TVLSI'25 [11] (Flex-PE), which is 2.6x more energy-efficient. The 2.85x figure is computed against [24]'s 40 pJ/op, not against the best row in the table. Similarly, the 42% area and 38% power reductions are against [24] only; [11] has lower power (7.3 mW vs 24.1 mW). The paper must either compare against [11] and explain any exclusion, or revise the headline to specify the actual comparison baseline.
  2. [Section III, Table II] The cross-design comparison in Table II is not apples-to-apples. The rows differ in technology (65 nm vs 28 nm), supply voltage (1.2 V, 1.05 V, 1.0 V, 0.9 V), frequency, and presumably datapath width and per-operation definition. Since pJ/op scales strongly with voltage and technology, the claimed 2.85x, 42%, and 38% improvements are not demonstrated as architectural advantages. The authors should provide normalized metrics (e.g., scaled to iso-voltage/technology or report energy-delay product) or clearly state that the comparison is raw and explain why the raw numbers are still meaningful.
  3. [Section III (synthesis methodology)] The paper does not describe the synthesis and power-estimation flow: library version, process corner (SS/TT/FF), temperature, clock constraints, switching activity or toggle rate, and whether power includes memory and clock-tree power. Without these details, the reported 1.72 GHz, 0.016 mm^2, and 24.1 mW cannot be verified or reproduced even with the released code. A clear methodology subsection is required before the hardware metrics can be assessed.
  4. [Section II and Section III (precision selection)] The reported accuracy and model-size results depend on per-layer precision assignments chosen using the sensitivity metric of Eq. (1) and the same validation data used to report the results. There is no validation split, hold-out set, or ablation against uniform-precision alternatives or random precision maps. This risks overfitting the precision map to the reported benchmarks and undermines the claim that the mixed-precision approach is generally effective. Please provide a robustness evaluation (e.g., compare against a fixed-precision baseline with equal bit-width, or report variance across validation folds).
minor comments (6)
  1. [Figure 1] Figure 1 is low-resolution and nearly unreadable; please re-render it with larger fonts and clearly labeled workload categories.
  2. [Table II] Table II lists two rows for TCAS-AI'25 [23] without a distinguishing design name or explanation; this looks like a duplicate or a two-mode entry and should be clarified.
  3. [Section III (terminology)] The paper uses "arithmetic intensity" to mean energy per operation (pJ/op), which conflicts with the standard definition of arithmetic intensity as operations per byte. Define the term explicitly at first use to avoid confusion.
  4. [Section II, Eqs. (1)-(5)] Some symbols in the quantization equations are not fully defined (e.g., 'sc', 'QMxP'', 'n', k in Eq. (3), and the saturation thresholds Wl/Wh in the text). Please provide complete definitions and describe how they are computed per layer.
  5. [Section III, off-chip data movement] The sentence "off-chip data-movement accounts to almost 60% of energy-consumption" lacks a citation; please add a reference or qualify it as the authors' workload analysis.
  6. [References] The reference list contains formatting inconsistencies (e.g., garbled text in several entries like [1], [19], [35]) and should be cleaned up with a consistent style.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: hardware metrics are direct synthesis measurements and accuracy results are empirical post-QAT evaluations, with no load-bearing self-citation chain.

full rationale

XR-NPE's central claims are hardware measurements (1.72 GHz, 0.016 mm^2, 14 pJ/op from 28-nm synthesis and place-and-route) and empirical accuracy results from quantized models evaluated on KITTI, EfficientNet, and gaze-estimation workloads. None of these results is defined in terms of the conclusion it supports. The sensitivity metric in Eq. (1)-(2) is used to choose per-layer precision, and the reported accuracy is then measured after quantization-aware training; selection shaping the outcome is an optimization procedure, not a definitional reduction. The comparisons to the state of the art are arithmetic ratios from Table II, not derived predictions, though the claim of 'best of state-of-the-art' is internally inconsistent because row [11] reports 5.37 pJ/op versus 14 pJ/op for this work; that is a correctness/support concern, not circularity. Self-citations to [11] and [17] appear as background references and as an API-validation choice, but the core architectural and algorithmic claims do not reduce to those citations; the work provides new synthesis results and public code for reproducibility. No equation is shown to be equivalent to its own input, and no fitted parameter is renamed as a prediction. Therefore no significant circularity is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on self-reported synthesis metrics, a selected set of comparison baselines, and per-layer precision assignments that are treated as part of the method but are not fully disclosed. No new physical or theoretical entities are introduced beyond the hardware blocks themselves, so the invented entities list is empty.

free parameters (3)
  • Per-layer precision assignment = Not disclosed
    The sensitivity metric in Eq. (1) is used to choose which layers run in FP4, Posit(4,1), Posit(8,0), or Posit(16,1), and this selection is made before inference using validation data. The resulting model size and accuracy are then reported as benefits.
  • PACT clipping threshold alpha = Not reported
    Eqs. (6) and (7) use a parameterized clipping activation with a trained threshold alpha per layer, but no values or training schedules are provided.
  • Saturation thresholds Wl and Wh = Data-dependent, not disclosed
    Eqs. (4) and (5) use lower and upper saturation thresholds chosen to align with the learned weight distribution; these thresholds influence the reported accuracy and are not documented.
assumptions (4)
  • domain assumption Cross-paper synthesis and FPGA numbers at 28 nm and 16 nm are directly comparable without normalizing PDK, voltage, frequency, datapath width, or operation definition.
    The main area, power, and energy claims in Tables II-IV are single-row comparisons across independently designed accelerators with different networks and tools.
  • domain assumption Software-emulated FP4/Posit accuracy matches what the hardware would deliver on-chip.
    Accuracy figures in Figs. 5-8 come from quantized algorithmic emulation and QAT, not from measurements on the synthesized XR-NPE.
  • domain assumption The FP4 and Posit arithmetic semantics are correctly implemented in the RMMEC and Quire datapath.
    No testbench results, formal verification, or silicon validation are provided for arithmetic correctness.
  • standard math First-order Taylor expansion of the loss is an adequate proxy for layer sensitivity to quantization.
    Eq. (1) follows standard quantization sensitivity practice, but it is used here to justify layer assignments without a sensitivity analysis or ablation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of XR-NPE: High-Throughput Mixed-precision SIMD Neural Processing Engine for Extended Reality Perception Workloads." pith.science (2026). https://pith.science/paper/24YXT4CY

@misc{pith2026250813049,
  author       = {Pith},
  title        = {Pith review of: XR-NPE: High-Throughput Mixed-precision SIMD Neural Processing Engine for Extended Reality Perception Workloads},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/24YXT4CY}},
  note         = {Machine review of arXiv:2508.13049}
}
read the original abstract

This work proposes XR-NPE, a high-throughput Mixed-precision SIMD Neural Processing Engine, designed for extended reality (XR) perception workloads like visual inertial odometry (VIO), object classification, and eye gaze extraction. XR-NPE is first to support FP4, Posit (4,1), Posit (8,0), and Posit (16,1) formats, with layer adaptive hybrid-algorithmic implementation supporting ultra-low bit precision to significantly reduce memory bandwidth requirements, and accompanied by quantization-aware training for minimal accuracy loss. The proposed Reconfigurable Mantissa Multiplication and Exponent processing Circuitry (RMMEC) reduces dark silicon in the SIMD MAC compute engine, assisted by selective power gating to reduce energy consumption, providing 2.85x improved arithmetic intensity. XR-NPE achieves a maximum operating frequency of 1.72 GHz, area 0.016 mm2 , and arithmetic intensity 14 pJ at CMOS 28nm, reducing 42% area, 38% power compared to the best of state-of-the-art MAC approaches. The proposed XR-NPE based AXI-enabled Matrix-multiplication co-processor consumes 1.4x fewer LUTs, 1.77x fewer FFs, and provides 1.2x better energy efficiency compared to SoTA accelerators on VCU129. The proposed co-processor provides 23% better energy efficiency and 4% better compute density for VIO workloads. XR-NPE establishes itself as a scalable, precision-adaptive compute engine for future resource-constrained XR devices. The complete set for codes for results reproducibility are released publicly, enabling designers and researchers to readily adopt and build upon them. https://github.com/mukullokhande99/XR-NPE.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Bhasha-Rupantarika: Algorithm-Hardware Co-design approach for Multilingual Neural Machine Translation

    cs.AR 2025-10 reject novelty 4.0 of 10

    4-bit quantization of NLLB-200 plus a custom FPGA accelerator (NLPE) yields claimed 4.1x smaller and 4.2x faster multilingual translation, but translation quality is never measured.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [11]

    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,” ���� ������ ���� �����, vol. 33, pp. 1610–1623, June 2025

    work page 2025

  2. [24]

    A 3-D Multi-Precision Scalable Systolic FMA Architecture,

    H. Liu, X. Lu, X. Yu, �� ���, “A 3-D Multi-Precision Scalable Systolic FMA Architecture,” ���� ������ �� ���� ��� ����� �, vol. 72, no. 1, pp. 265–276, 2025

    work page 2025

  3. [1]

    Simulating Virtual Environment and Experience for Training, Exergaming, and Edutainment in eXtended Reality (XR): A Survey,

    W. Li, “Simulating Virtual Environment and Experience for Training, Exergaming, and Edutainment in eXtended Reality (XR): A Survey,” in ������������� ���������� �� �������� ������������ ���������� ������ , pp. 309–317, 2023

    work page 2023

  4. [2]

    Augmented Reality and Virtual Reality displays: Emerging technologies and future perspectives,

    J. Xiong, E.-L. Hsiang, �� ���, “Augmented Reality and Virtual Reality displays: Emerging technologies and future perspectives,” ������ ������� � ������������, vol. 10, no. 1, p. 216, 2021

    work page 2021

  5. [3]

    A 25.1-TOPS/W Sparsity-Aware Hybrid CNN-GCN Deep Learning SoC for Mobile Augmented Reality,

    W.-C. Huang, I.-T. Lin, �� ���, “A 25.1-TOPS/W Sparsity-Aware Hybrid CNN-GCN Deep Learning SoC for Mobile Augmented Reality,” ���� ������� �� ����������� ��������, vol. 59, no. 11, pp. 3840–3852, 2024

    work page 2024

  6. [4]

    System-Level Design and Integration of a Prototype AR/VR Hardware Featuring a Custom Low-Power DNN Accelerator Chip in 7nm Technology for Codec Avatars,

    H. E. Sumbul, T. F. Wu, �� ���, “System-Level Design and Integration of a Prototype AR/VR Hardware Featuring a Custom Low-Power DNN Accelerator Chip in 7nm Technology for Codec Avatars,” in ���� ���� ������ ���������� �������� ���������� ������, pp. 01–08, 2022

    work page 2022

  7. [5]

    An Empirical Study on Meta Virtual Reality Applications: Security and Privacy Perspectives,

    H. Guo, H.-N. Dai, �� ���, “An Empirical Study on Meta Virtual Reality Applications: Security and Privacy Perspectives,” ���� ������������ �� �������� �����������, vol. 51, no. 5, pp. 1437–1454, 2025

    work page 2025

  8. [6]

    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 ���� ������������� ����������� �������� ���������� �������, vol. 68, pp. 1–3, 2025

    work page 2025

Show all 35 references
  1. [7]

    Aspen: A 630 FPS Real-Time Posit-Based Unified Accelerator for Extended Reality Perception Workloads,

    K. Feng, K. Prabhu, K. Bartolone, J. Yu, and P. Raina, “Aspen: A 630 FPS Real-Time Posit-Based Unified Accelerator for Extended Reality Perception Workloads,” in ���� ������ ���������� �������� ���������� ������, pp. 1–3, 2025

    2025

  2. [8]

    Occamy: A 432-Core Dual-Chiplet Dual- HBM2E 768-DP-GFLOP/s RISC-V System for 8-to-64-bit Dense and Sparse Computing in 12-nm FinFET,

    P. Scheffler, T. Benz, �� ���, “Occamy: A 432-Core Dual-Chiplet Dual- HBM2E 768-DP-GFLOP/s RISC-V System for 8-to-64-bit Dense and Sparse Computing in 12-nm FinFET,” ���� ������� �� ����������� ��������, vol. 60, no. 4, 2025

    2025

  3. [9]

    Parallel Accurate Minifloat MACCs for NN Inference on Versal FPGAs,

    H. J. Damsgaard, K. J. HoBfeld, and J. Nurmi, “Parallel Accurate Minifloat MACCs for NN Inference on Versal FPGAs,” ���� ������ ����������� ���� ������ ���� �����, vol. 44, pp. 2181–2194, June 2025

    2025

  4. [10]

    Maestro: A 302 GFLOPS/W and 19.8 GFLOPS RISC-V Vector-Tensor Architecture for Wearable Ultrasound Edge Computing,

    M. Sinigaglia �� ���, “Maestro: A 302 GFLOPS/W and 19.8 GFLOPS RISC-V Vector-Tensor Architecture for Wearable Ultrasound Edge Computing,” ���� ������ �� �������� ��� ������ �, pp. 1–15, 2025

    2025

  5. [12]

    AMD XDNA NPU in Ryzen AI Processors,

    A. Rico, S. Pareek, �� ���, “AMD XDNA NPU in Ryzen AI Processors,” ���� ����� , vol. 44, no. 3, pp. 123–130, 2024

    2024

  6. [13]

    NVIDIA Blackwell Platform: Advancing Generative AI and Accelerated Computing,

    A. Tirumala and R. Wong, “NVIDIA Blackwell Platform: Advancing Generative AI and Accelerated Computing,” in ���� ��� ����� ���� ������ �����, vol. 36, pp. 1–33, 2024

    2024

  7. [14]

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

    H. Tan, J. Zhang, X. He, L. Huang, Y . Wang, and L. Xiao, “A Low-Cost Floating-Point FMA Unit Supporting Package Operations for HPC-AI Applications,” ���� ������ �� �������� ��� ������� ��� ������� ������, vol. 71, pp. 3488–3492, July 2024

    2024

  8. [15]

    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, �� ���, “A Multi-Mode 8k-MAC HW-Utilization- Aware Neural Processing Unit With a Unified Multi-Precision Datapath in 4-nm Flagship Mobile SoC,” ���� ������� �� ����������� ��������, vol. 58, no. 1, pp. 189–202, 2023

    2023

  9. [16]

    RaPiD: AI Accelerator for Ultra-low Precision Training and Inference,

    S. Venkataramani, V . Srinivasan, �� ���, “RaPiD: AI Accelerator for Ultra-low Precision Training and Inference,” �������� ���� ������ ������������� ��������� �� �������� ������������, pp. 153–166, 2021

    2021

  10. [17]

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

    O. Kokane, M. Lokhande, �� ���, “LPRE: Logarithmic Posit-enabled Reconfigurable edge-AI Engine,” in ���� ������������� ��������� �� �������� ��� ������� �������, pp. 1–5, May 2025

    2025

  11. [18]

    Cheshire: A lightweight, linux-capable risc-v host platform for domain-specific accel- erator plug-in,

    A. Ottaviano, T. Benz, P. Scheffler, and L. Benini, “Cheshire: A lightweight, linux-capable risc-v host platform for domain-specific accel- erator plug-in,” ���� ������������ �� �������� ��� ������� ��� ������� ������, vol. 70, no. 10, pp. 3777–3781, 2023

    2023

  12. [19]

    8-bit transformer inference and fine-tuning for edge accelerators,

    J. Yu, K. Prabhu, Y . Urman, R. M. Radway, E. Han, and P. Raina, “8-bit transformer inference and fine-tuning for edge accelerators,” in 6 ����������� �� ��� ���� ��� ������������� ���������� �� ������������� ������� ��� ����������� ��������� ��� ��������� �������� ������ �, A...

    2024

  13. [20]

    Low-Precision Mixed-Computation Models for Infer- ence on Edge,

    S. Azizi �� ���, “Low-Precision Mixed-Computation Models for Infer- ence on Edge,” ���� ������ ���� �����, vol. 32, no. 8, pp. 1414–1422, 2024

    2024

  14. [21]

    Hnpu: An adaptive dnn training processor utilizing stochastic dynamic fixed- point and active bit-precision searching,

    D. Han, D. Im, G. Park, Y . Kim, S. Song, J. Lee, and H.-J. Yoo, “Hnpu: An adaptive dnn training processor utilizing stochastic dynamic fixed- point and active bit-precision searching,” ���� ������� �� ����������� ��������, vol. 56, no. 9, pp. 2858–2869, 2021

    2021

  15. [22]

    UL-VIO: Ultra-Lightweight Visual- Inertial Odometry with Noise Robust Test-Time Adaptation,

    J. Park, S. Y . Chun, and M. Seok, “UL-VIO: Ultra-Lightweight Visual- Inertial Odometry with Noise Robust Test-Time Adaptation,” in ���� ����� ������ � ���� ����� ���� �������� ����������� ������ ������ ��������� ���������� �� ����� ������������ ���� �����, (Berlin, Hei- delbe...

    2024

  16. [23]

    A Configurable Floating-Point Fused Multiply-Add Design with Mixed Precision for AI Accelerators,

    F. Niknia, Z. Wang, �� ���, “A Configurable Floating-Point Fused Multiply-Add Design with Mixed Precision for AI Accelerators,” ���� ������ �� ���� ��� ����� ��, pp. 1–15, 2025

    2025

  17. [25]

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

    H. Tan, L. Huang, �� ���, “A Low-Cost FP Dot-Product-Dual-Accumulate Architecture for HPC-Enabled AI,” ���� ������ ����������� ���� ������ ���� �����, vol. 43, pp. 681–693, Feb. 2024

    2024

  18. [26]

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

    L. Crespo, P. Tom ´as, and N. Roma, “Unified Posit/IEEE-754 Vector MAC Unit for Transprecision Computing,” ���� ������ �� �������� ��� ����� ��, vol. 69, pp. 2478–2482, May 2022

    2022

  19. [27]

    Dedicated FPGA Implementation of the Gaussian TinyYOLOv3 Accelerator,

    S. Ki, J. Park, and H. Kim, “Dedicated FPGA Implementation of the Gaussian TinyYOLOv3 Accelerator,” ���� ������ �������� ����� ��� ���� ������, vol. 70, pp. 3882–3886, Oct. 2023

    2023

  20. [28]

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

    M. Kim, K. Oh, �� ���, “A Low-Latency FPGA Accelerator for YOLOv3- Tiny With Flexible Layerwise Mapping and Dataflow,” ���� ������ �������� ����� �, vol. 71, pp. 1158–1171, Mar. 2024

    2024

  21. [29]

    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,” ���� ������ �������� ����� �, vol. 71, pp. 1285–1298, Mar. 2024

    2024

  22. [30]

    Vision meets robotics: The kitti dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The kitti dataset,” ������������� ������� �� �������� �������� ������, 2013

    2013

  23. [31]

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

    K. Li, M. Huang, �� ���, “A 29.12-TOPS/W Vector Systolic Accelerator With NAS-Optimized DNNs in 28-nm CMOS,” ���� ������� �� ������ ����� ��������, pp. 1–12, 2025

    2025

  24. [32]

    MSDF-Based MAC for Energy-Efficient Neural Networks,

    S. M. Cherati, M. Barzegar, and L. Sousa, “MSDF-Based MAC for Energy-Efficient Neural Networks,” ���� ������ �� ���� ����� ����� ����������� ������ �������, pp. 1–12, July 2025

    2025

  25. [33]

    Marsellus: A Heterogeneous RISC-V AI-IoT End-Node SoC With 2–8 b DNN Acceleration and 30%- Boost Adaptive Body Biasing,

    F. Conti, G. Paulin, A. Garofalo, �� ���, “Marsellus: A Heterogeneous RISC-V AI-IoT End-Node SoC With 2–8 b DNN Acceleration and 30%- Boost Adaptive Body Biasing,” ���� ������� �� ����������� ��������, vol. 59, no. 1, pp. 128–142, 2024

    2024

  26. [34]

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

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

    2022

  27. [35]

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

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

    2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.