Hooshaai/neuralops-results

NeuralOps Benchmark Results & Artifacts Repository

Author: Taha Majlesi
Release Version: 1.4.0
Hardware Environment: NVIDIA GeForce RTX 4070 (AD104, 12GB VRAM, Compute Capability 8.9)
Software Stack: PyTorch 2.5+, Triton 3.0+, CUDA 12.4+

---

1. Overview

This directory contains the canonical benchmark results, empirical evaluation matrices, hardware profiling traces, and verification logs for the NeuralOps library at release version v1.4.0.

All evaluation data points reflect real physical hardware executions under identical reproducible seeds and thermal conditions. The directory is sanitized to maintain zero build clutter (.err, .log, .tmp, and zero-byte files are barred).

---

2. Canonical Benchmark Artifacts

The following canonical benchmark files represent the single source of truth for all published metrics in the thesis and research paper:

File	Format	Description
MASTER_RESULTS_MATRIX.csv	CSV	Master results matrix spanning all attention mechanisms and compressors, recording latency, memory, throughput, and GLUE downstream accuracies.
MASTER_RESULTS_MATRIX.md	Markdown	Rendered markdown summary table of the global master results matrix.
triton_all_26_kernels_benchmark.json	JSON	Hardware benchmark report for all 26 Triton kernel modules, measuring speedup vs. native PyTorch baselines.
master_108_cuda_modules_benchmark.json	JSON	Microbenchmark measurements and speedup figures for all native CUDA modules.
master_108_cuda_modules_benchmark.csv	CSV	Tabular counterpart of the 108 CUDA modules microbenchmark report.
master_paper_categorized_inventory.json	JSON	Structured architectural taxonomy and mathematical classifications for all modules.
master_paper_categorized_inventory.csv	CSV	Tabular taxonomy inventory for paper generation and appendices.
triton_all_131_attentions_benchmark.json	JSON	Empirical validation and latency metrics for all 131 canonical subquadratic attention mechanisms.
triton_all_72_compressors_benchmark.json	JSON	Empirical compression metrics, recovery scores, and memory footprints for all 72 compression engines.
SEQLEN_SCALING_MASTER.csv	CSV	Throughput and memory consumption scaled from sequence lengths $T = 512$ up to $T = 131,072$.
HARDWARE_ENVIRONMENT_MANIFEST.json	JSON	Full snapshot of hardware topology, GPU compute capability, clocks, and software toolchains.
RELEASE_CHECKSUMS_v1.4.0.txt	Text	SHA-256 cryptographic checksums for all canonical result artifacts.
RESULTS_MASTER_INDEX.json	JSON	Master JSON index indexing all experiment categories and suite paths.
INDEX.md	Markdown	Detailed browsable index across all 268 evaluation suites.

---

3. Directory Layout

The benchmark repository is partitioned into modular directories:

results/
├── README.md                                  # This document
├── INDEX.md                                   # Master suite index
├── RESULTS_MASTER_INDEX.json                  # Machine-readable suite index
├── RELEASE_CHECKSUMS_v1.4.0.txt               # Cryptographic integrity manifest
├── HARDWARE_ENVIRONMENT_MANIFEST.json         # Hardware & environment specification
│
├── [Canonical Matrices]
│   ├── MASTER_RESULTS_MATRIX.csv              # Global cross-evaluation matrix
│   ├── master_108_cuda_modules_benchmark.json # 108 CUDA kernels benchmark
│   ├── triton_all_26_kernels_benchmark.json   # 26 Triton kernels benchmark
│   ├── triton_all_131_attentions_benchmark.json# 131 Attentions benchmark
│   └── triton_all_72_compressors_benchmark.json# 72 Compressors benchmark
│
├── plots/                                     # Publication-grade figures & curves
├── paper_figures/                             # Rendered vector diagrams for paper/thesis
├── profiling/                                 # Kernel traces, roofline analysis & Nsight profiles
├── kd_runs/                                   # Knowledge distillation fine-tuning logs & weights
├── roberta_benchmark_sweep/                   # RoBERTa GLUE downstream evaluation sweeps
└── 00_archive/ ... 12_archive/                # Preserved historical experimental suites

---

4. Data Hygiene and Integrity Assurance

1. Zero Clutter: Ephemeral .log, .err, .tmp, and zero-byte files are automatically pruned by hygiene scripts.
2. Reproducibility: Experiments are tagged with exact git commit SHAs in MASTER_RESULTS_MATRIX.csv.
3. Cryptographic Validation: Any modified benchmark file must have its SHA-256 updated in RELEASE_CHECKSUMS_v1.4.0.txt.