Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,79 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
license: apache-2.0
|
| 4 |
+
tags:
|
| 5 |
+
- benchmark
|
| 6 |
+
- mobile-ai
|
| 7 |
+
- edge-computing
|
| 8 |
+
- snapdragon
|
| 9 |
+
- on-device
|
| 10 |
+
- llama-cpp
|
| 11 |
+
task_categories:
|
| 12 |
+
- text-generation
|
| 13 |
+
size_categories:
|
| 14 |
+
- n<1K
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Per-Chip Benchmark Matrix
|
| 18 |
+
|
| 19 |
+
On-device inference benchmarks for mobile LLMs across chipsets.
|
| 20 |
+
|
| 21 |
+
## Overview
|
| 22 |
+
|
| 23 |
+
This dataset contains real on-device inference benchmarks for 8 mobile-optimized
|
| 24 |
+
models running on Samsung S20 FE 5G phones (Snapdragon 865, 8GB RAM, Android 13).
|
| 25 |
+
|
| 26 |
+
## Contents
|
| 27 |
+
|
| 28 |
+
- `benchmark_matrix.csv` — Tabular data: model, device, chipset, tokens/sec, size
|
| 29 |
+
- `benchmark_matrix.json` — Full structured data including hardware specs and methodology
|
| 30 |
+
|
| 31 |
+
## Key Findings
|
| 32 |
+
|
| 33 |
+
| Model | Params | Quant | Gen t/s | Size |
|
| 34 |
+
|-------|--------|-------|---------|------|
|
| 35 |
+
| SmolLM2-135M | 135M | fp16 | 18.3 | 270MB |
|
| 36 |
+
| SmolLM2-360M | 360M | fp16 | 12.1 | 720MB |
|
| 37 |
+
| Qwen2.5-0.5B | 500M | int4 | 9.8 | 350MB |
|
| 38 |
+
| Llama-3.2-1B | 1B | fp16 | 6.5 | 2000MB |
|
| 39 |
+
| TinyLlama-1.1B | 1.1B | Q5 | 6.2 | 450MB |
|
| 40 |
+
| Qwen2.5-1.5B | 1.5B | Q5 | 4.2 | 900MB |
|
| 41 |
+
| Phi-3.5-mini | 2B | Q5 | 3.1 | 1300MB |
|
| 42 |
+
| Gemma-2-2B | 2B | Q5 | 3.0 | 1300MB |
|
| 43 |
+
|
| 44 |
+
## Methodology
|
| 45 |
+
|
| 46 |
+
Models run via `llama-cli` (llama.cpp) on Samsung S20 FE 5G devices:
|
| 47 |
+
- **Chipset**: Qualcomm Snapdragon 865 (Kryo 585, 7nm)
|
| 48 |
+
- **RAM**: 8GB
|
| 49 |
+
- **OS**: Android 13
|
| 50 |
+
- **Prompt**: "The capital of France is"
|
| 51 |
+
- **Max tokens**: 20
|
| 52 |
+
- **Threads**: 4
|
| 53 |
+
- **Context**: 512
|
| 54 |
+
|
| 55 |
+
All benchmarks run on real hardware. No simulation.
|
| 56 |
+
|
| 57 |
+
## Chipset Coverage
|
| 58 |
+
|
| 59 |
+
| Chipset | Process | NPU | Devices |
|
| 60 |
+
|---------|---------|-----|---------|
|
| 61 |
+
| Snapdragon 865 | 7nm | Hexagon HTA | Samsung S20 FE 5G |
|
| 62 |
+
| Snapdragon 8 Gen 2 | 4nm | Hexagon NPU | Samsung S23, OnePlus 11 |
|
| 63 |
+
| Apple A17 Pro | 3nm | Neural Engine (35 TOPS) | iPhone 15 Pro |
|
| 64 |
+
| MediaTek Dimensity 9200 | 4nm | APU 690 | Various flagships |
|
| 65 |
+
|
| 66 |
+
## Citation
|
| 67 |
+
|
| 68 |
+
```bibtex
|
| 69 |
+
@misc{dispatchai_benchmark_2026,
|
| 70 |
+
title={Per-Chip Benchmark Matrix: On-Device LLM Inference Benchmarks},
|
| 71 |
+
author={Aljallaf Alzaabi, Omar Abdulla Jasem},
|
| 72 |
+
year={2026},
|
| 73 |
+
url={https://huggingface.co/datasets/dispatchAI/per-chip-benchmark-matrix}
|
| 74 |
+
}
|
| 75 |
+
```
|
| 76 |
+
|
| 77 |
+
---
|
| 78 |
+
|
| 79 |
+
*Dispatch AI (FZE), Sharjah SRTI Free Zone, License No. 10818. Benchmarked on real hardware, $0 cost.*
|