Global ETD Search

Search theses and dissertations gathered from participating repositories worldwide. Every result links back to the library that holds it. No account is needed.

Results

Showing 1 to 18 of 18 for “"FP16"”.

  1. Quantized Lookup Table-Based Representation for Key-Value Cache

    … followed by Huffman coding, converting large FP16 tensors into compact bitstreams. Experimental results across three open-source LLMs and six long-context benchmarks from LongBench demonstrate that Quark achieves effective bitwidths of 3.07-3.32 bits, corresponding to compression ratios of …

    tamu Repository record for Quantized Lookup Table-Based Representation for Key-Value Cache (opens in a new tab) Embargoed until 2028-05-01

  2. Quantized Lookup Table-Based Representation for Key-Value Cache

    … followed by Huffman coding, converting large FP16 tensors into compact bitstreams. Experimental results across three open-source LLMs and six long-context benchmarks from LongBench demonstrate that Quark achieves effective bitwidths of 3.07-3.32 bits, corresponding to compression ratios of …

    tamu Repository record for Quantized Lookup Table-Based Representation for Key-Value Cache (opens in a new tab) Embargoed until 2028-05-01

  3. Quantized Lookup Table-Based Representation for Key-Value Cache

    … followed by Huffman coding, converting large FP16 tensors into compact bitstreams. Experimental results across three open-source LLMs and six long-context benchmarks from LongBench demonstrate that Quark achieves effective bitwidths of 3.07-3.32 bits, corresponding to compression ratios of …

    tamu Repository record for Quantized Lookup Table-Based Representation for Key-Value Cache (opens in a new tab) Embargoed until 2028-05-01

  4. Rychlá předdetekce objektů v obraze

    … formátu OpenVINO ve variantách přesnosti FP32, FP16 a INT8 a systematicky testovány na cílovém hardwaru z hlediska rychlosti inference i přesnosti detekce. Na základě výsledků byl jako finální model zvolen YOLOv8n INT8, který dosahuje nejlepšího kompromisu mezi rychlostí a přesností ze všech …

    brno-tech Repository record for Rychlá předdetekce objektů v obraze (opens in a new tab)

  5. ONBOARD SHIP CLASSIFICATION FOR MARITIME SURVEILLANCE: BENCHMARKING CONVOLUTIONAL NEURAL NETWORK-BASED TRANSFER LEARNING ON THE NVIDIA JETSON ORIN NANO 8GB

    … integer (INT8), half-precision floating-point (FP16), and full-precision floating-point (FP32) inference precision. The improved configuration—partial backbone unfreezing, weighted cross-entropy loss, and extended training—increased ResNet-50 accuracy to 86.8% and EfficientNet-B0 to 81.0%, with …

    nps Repository record for ONBOARD SHIP CLASSIFICATION FOR MARITIME SURVEILLANCE: BENCHMARKING CONVOLUTIONAL NEURAL NETWORK-BASED TRANSFER LEARNING ON THE NVIDIA JETSON ORIN NANO 8GB (opens in a new tab)

  6. Comparing the Performance of Small Word-Size Floating-Point Numerics to Fixed-Point Numerics in Neural Networks

    … floating-point formats (FP8, FP12, FP16) for dynamic, layer-wise quantization during training. APT monitors per-layer Quantization Error Measurement (QEM) to guide precision adjustments and incorporates a novel bit-shuffling mechanism to reallocate bits between exponent and mantissa …

    gatech Repository record for Comparing the Performance of Small Word-Size Floating-Point Numerics to Fixed-Point Numerics in Neural Networks (opens in a new tab)

  7. Efficient Mixed-Precision Inference for Vision Transformers

    … (GB) de memoria en punto flotante de 16 bits (FP16), lo cual requiere múltiples nodos de cómputo acelerado. Esta arquitectura fue transferida a la visión por computadora como Vision Transformer (ViT), donde aumentar el número de parámetros también incrementa el poder predictivo. Versiones …

    upv Repository record for Efficient Mixed-Precision Inference for Vision Transformers (opens in a new tab)

  8. Efficient And Numerically Stable GPU Implementations of The Fast Walsh-Hadamard Transform

    … up to 263 and the numerical formats: FP64, FP32, FP16 and BF16. Both implementations suffer heavily from being memory bound, only outperforming Nvidia’s implementation of the computationally heavier Fast Fourier Transform on L4/L40s GPUs when using FP64. We find no benefit of using Alman & Rao’s …

    chalmers Repository record for Efficient And Numerically Stable GPU Implementations of The Fast Walsh-Hadamard Transform (opens in a new tab)

  9. Evaluating Design Choices for GPU-Accelerated Finite-Difference Time-Domain Simulation on Resource-Constrained Devices

    … for a given computation. Floating point 16 (FP16) is the fastest and most energy-efficient precision choice, but up to 1000 times less accurate. Fixed-point integer formats draw less instantaneous power but run much longer, consuming more total energy, and are tens of thousands of times less …

    calpoly Repository record for Evaluating Design Choices for GPU-Accelerated Finite-Difference Time-Domain Simulation on Resource-Constrained Devices (opens in a new tab)

  10. OS INSPIRED COMPLETE KERNEL FUSION

    … FlashDMoE using FP32 while baselines use FP16.

    cornell Repository record for OS INSPIRED COMPLETE KERNEL FUSION (opens in a new tab)

  11. Vision Perception Optimization and Adaptive Control for Resource-Constrained Platform: A Ping-Pong Ball Pickup & Place System

    … studies reveal that combining TensorRT with FP16 precision yields a 592.6% performance improvement with negligible accuracy loss. The spiral search strategy demonstrates effective target recovery capabilities when objects temporarily leave the field of view. Most notably, visual servoing …

    auckland-tech Repository record for Vision Perception Optimization and Adaptive Control for Resource-Constrained Platform: A Ping-Pong Ball Pickup & Place System (opens in a new tab)

  12. Fast and Energy-Efficient Inference for Attention-Based Natural Language Processing Models

    … that focus on the bulk of values to map FP32 or FP16 numbers into 3- or 4-bit indexes to dictionaries of representative values. This means forcing most of the values in a layer to be one of 8 or 16 unique values (instead of 2^32 in single-precision floating-point, FP32, format) while maintaining …

    toronto-retro Repository record for Fast and Energy-Efficient Inference for Attention-Based Natural Language Processing Models (opens in a new tab)

  13. Statistical Analysis of Outlier Patterns in Low-bit Quantization Strategies in Large Language Models

    … forming a controlled ablation study: FP16 baseline, weight-only rotation (W4A16KV16), W4A4KV4 without prefix, PrefixQuant-O2 ( η1 = 64), and PrefixQuant-EVT ( θEVT = 3.47). Our anal- ysis reveals four key findings. First, Hadamard rotation and prefix isolation address dif- ferent …

    ttu Repository record for Statistical Analysis of Outlier Patterns in Low-bit Quantization Strategies in Large Language Models (opens in a new tab)

  14. Mixed-Precision Numerical Linear Algebra Algorithms: Integer Arithmetic Based LU Factorization and Iterative Refinement for Hermitian Eigenvalue Problem

    … low-precision formats including half precision (FP16), Bfloat16 and integer operations for quantized integers, which uses integers with a shared scalar to represent a set of equally spaced numbers. As new hardware architectures focus on bringing performance in these formats, the mixed-precision …

    utk-theses Repository record for Mixed-Precision Numerical Linear Algebra Algorithms: Integer Arithmetic Based LU Factorization and Iterative Refinement for Hermitian Eigenvalue Problem (opens in a new tab)

  15. A Custom RTL Domain-Specific GEMM Accelerator on Cyclone V SoC: Dynamic Precision Selection and Tiling, from RTL Design to Linux Bring-Up for Edge AI Inference and Predictive Control

    … the feed path avoids almost all of that cost: BFP16 converts to fixed point before the array, reaches all 64 elements rather than the 16 in the corner, and matches INT16 cycle for cycle to within 0.014%. The conversion costs 514 logic cells, zero DSP slices, and roughly 1% of accelerator power. …

    washington Repository record for A Custom RTL Domain-Specific GEMM Accelerator on Cyclone V SoC: Dynamic Precision Selection and Tiling, from RTL Design to Linux Bring-Up for Edge AI Inference and Predictive Control (opens in a new tab)

  16. Automatic Lyrics Transcription for Greek Songs

    … Transformers, batch size 4-8, μικτή ακρίβεια FP16 και optimizer AdamW (learning rate 5×10⁻⁵). Η αξιολόγηση της απόδοσης βασίστηκε στη μετρική normalized Word Error Rate (WER), η οποία εξουδετερώνει την επίδραση της στίξης και των μικρο-ορθογραφικών διαφορών. Πειραματικά, το zero-shot …

    athens Repository record for Automatic Lyrics Transcription for Greek Songs (opens in a new tab)