tahamajs/neuralops-results-2

NeuralOps Empirical Hardware & Distillation Benchmark Results

This repository contains the canonical, empirically grounded benchmark results, hardware telemetry, and knowledge distillation ledgers produced by the NeuralOps research framework.

All data in this repository was generated through physical hardware execution on dedicated computing resources (NVIDIA GeForce RTX 3090 24GB, SM 8.6, and Intel Xeon E5-2698R v4 CPUs) on the Sharif University HPC cluster. Zero synthetic or hallucinated metrics are included.

---

Repository Structure

neuralops-results-2/
├── master_ledger.sqlite             # Canonical master SQLite ledger containing all empirical runs
├── BENCHMARK_EXECUTIVE_SUMMARY.md  # Detailed executive benchmark summary
├── benchmarks_203_all_modules.json  # Comprehensive JSON export of all 203 tested modules
├── benchmarks_latest.jsonl         # Line-delimited JSON stream of benchmark executions
├── qwen_wikitext_kd_ablation.csv    # Qwen-2.5-0.5B attention surgery & WikiText-2 perplexity
├── csv/                             # Tabular CSV exports of every evaluation matrix
│   ├── per_module_kd_benchmarks.csv      # Complete 203 module KD runs (131 Attn + 72 Comp)
│   ├── hardware_kernel_shape_matrix.csv  # 36 hardware sweep configurations (HGRN2, GDN, GLA, Scan)
│   ├── cpu_module_benchmarks.csv         # Complete 187 module CPU inference sweep
│   ├── cpu_multithread_scaling.csv       # Multi-threaded OpenMP/MKL scalability (1 to 6 threads)
│   ├── cpu_svd_whitening_benchmarks.csv  # High-dimensional SVD & Covariance Whitening benchmarks
│   ├── cpu_prefix_scan_benchmarks.csv    # Associative scan vs linear recurrence on CPU
│   ├── competitor_benchmarks.csv         # Comparative analysis against baseline implementations
│   ├── cross_architecture_benchmarks.csv # Cross-architecture evaluation across model backbones
│   └── ...
├── tables/                          # Publication-grade LaTeX tables
│   ├── attention_zoo_summary.tex
│   ├── compressor_zoo_summary.tex
│   ├── serving_matrix_summary.tex
│   └── benchmark_summary.tex
├── plots/                           # Publication vector (PDF) and raster (PNG) charts
│   ├── 10_latency_scaling_32k.pdf
│   ├── 16_multiprecision_roofline.png
│   ├── 18_p_value_heatmap.png
│   └── ...
└── reports/                         # Executive dashboards and markdown summaries

---

Core Benchmark Highlights

1. 100% Complete Module Coverage (per_module_kd_benchmarks)

• Attentions: 131 / 131 PASS (100.0%) verified with gradient updates and real token distillation.
• Compressors: 72 / 72 PASS (100.0%) verified across low-rank, pruning, quantization, and tensor train architectures.
• Grand Total: 203 / 203 PASS (100.0% completion) with zero NaN / zero Inf.

2. Multi-Shape Hardware Kernel Matrix (hardware_kernel_shape_matrix)

• Swept 4 native CUDA C++ and Triton JIT kernel families across 9 shape configurations ($B \in {1, 4, 8}$, $S \in {256, 512, 1024, 2048}$, $D=64$):
  • HGRN2 Scan (Triton): Average 1,185.5x speedup (peak 2,195.3x), max throughput 79,840,323 tok/s.
  • GDN Scan (CUDA C++): Average 248.1x speedup (peak 589.9x), max throughput 497,011 tok/s.
  • Selective Scan (CUDA C++): Average 136.2x speedup (peak 235.8x), max throughput 4,767,012 tok/s.
  • GLA Scan (Triton): Average 35.3x speedup (peak 52.5x), max throughput 1,196,923 tok/s.

3. CPU-Bound Mathematical & Scalability Benchmark (cpu_multithread_scaling)

• High-dimensional GEMM ($2048 \times 2048$): Scaled from 75.6 GFLOPS (1 thread) to 396.0 GFLOPS (6 threads) — 5.24x speedup with 87.3% parallel efficiency.
• Truncated SVD ($2048 \times 2048$, rank 64): Accelerated by 3.06x on multi-threaded CPU.
• Covariance Whitening Pullback: Accelerated by 3.29x.
• Chunkwise Vectorized Scan: Reached 7,596,865 tokens/sec on CPU.

---

Quickstart: Querying the Master Ledger in Python

import sqlite3
import pandas as pd

# Connect to the master SQLite ledger
conn = sqlite3.connect("master_ledger.sqlite")

# Query module coverage
df_modules = pd.read_sql_query("SELECT module_type, count(*), sum(case when status='PASS' then 1 else 0 end) as passed FROM per_module_kd_benchmarks GROUP BY module_type", conn)
print(df_modules)

# Query hardware kernel speedups
df_kernels = pd.read_sql_query("SELECT kernel_name, shape, speedup_vs_pytorch, custom_throughput_tok_s FROM hardware_kernel_shape_matrix ORDER BY speedup_vs_pytorch DESC LIMIT 10", conn)
print(df_kernels)

---

License

MIT License. Copyright (c) 2024-2026, Taha Majlesi.