title string | description string | hardware dict | models list | results list | methodology string | cost string | date timestamp[s] |
|---|---|---|---|---|---|---|---|
dispatchAI Per-Chip Benchmark Matrix | On-device inference benchmarks for mobile LLMs across chipsets | {
"snapdragon_865": {
"chipset": "Qualcomm Snapdragon 865",
"cpu": "Kryo 585 (1x2.84GHz + 3x2.42GHz + 4x1.80GHz)",
"gpu": "Adreno 650",
"npu": "Hexagon Tensor Accelerator (HTP)",
"process_nm": 7,
"devices": [
"Samsung Galaxy S20 FE 5G"
]
},
"snapdragon_8_gen_2": {
"chipset": ... | [
{
"model_id": "dispatchAI/Qwen2.5-0.5B-Instruct-mobile-int4",
"base_model": "Qwen/Qwen2.5-0.5B-Instruct",
"params_m": 500,
"quant": "int4",
"size_mb": 350,
"architecture": "Qwen2.5",
"task": "text-generation"
},
{
"model_id": "dispatchAI/Llama-3.2-1B-Instruct-mobile",
"base_m... | [
{
"model_id": "dispatchAI/Qwen2.5-0.5B-Instruct-mobile-int4",
"base_model": "Qwen/Qwen2.5-0.5B-Instruct",
"architecture": "Qwen2.5",
"params_m": 500,
"quantization": "int4",
"size_mb": 350,
"task": "text-generation",
"device_serial": "R3CN509PLHA",
"device_model": "SM-G981V",
... | Models run via llama-cli on Samsung S20 FE 5G (Snapdragon 865, 8GB RAM, Android 13). Prompt: 'The capital of France is', max_tokens=20, threads=4, context=512. | $0 | 2026-06-26T00:00:00 |
Per-Chip Benchmark Matrix
On-device inference benchmarks for mobile LLMs across chipsets.
Overview
This dataset contains real on-device inference benchmarks for 8 mobile-optimized models running on Samsung S20 FE 5G phones (Snapdragon 865, 8GB RAM, Android 13).
Contents
benchmark_matrix.csv— Tabular data: model, device, chipset, tokens/sec, sizebenchmark_matrix.json— Full structured data including hardware specs and methodology
Key Findings
| Model | Params | Quant | Gen t/s | Size |
|---|---|---|---|---|
| SmolLM2-135M | 135M | fp16 | 18.3 | 270MB |
| SmolLM2-360M | 360M | fp16 | 12.1 | 720MB |
| Qwen2.5-0.5B | 500M | int4 | 9.8 | 350MB |
| Llama-3.2-1B | 1B | fp16 | 6.5 | 2000MB |
| TinyLlama-1.1B | 1.1B | Q5 | 6.2 | 450MB |
| Qwen2.5-1.5B | 1.5B | Q5 | 4.2 | 900MB |
| Phi-3.5-mini | 2B | Q5 | 3.1 | 1300MB |
| Gemma-2-2B | 2B | Q5 | 3.0 | 1300MB |
Methodology
Models run via llama-cli (llama.cpp) on Samsung S20 FE 5G devices:
- Chipset: Qualcomm Snapdragon 865 (Kryo 585, 7nm)
- RAM: 8GB
- OS: Android 13
- Prompt: "The capital of France is"
- Max tokens: 20
- Threads: 4
- Context: 512
All benchmarks run on real hardware. No simulation.
Chipset Coverage
| Chipset | Process | NPU | Devices |
|---|---|---|---|
| Snapdragon 865 | 7nm | Hexagon HTA | Samsung S20 FE 5G |
| Snapdragon 8 Gen 2 | 4nm | Hexagon NPU | Samsung S23, OnePlus 11 |
| Apple A17 Pro | 3nm | Neural Engine (35 TOPS) | iPhone 15 Pro |
| MediaTek Dimensity 9200 | 4nm | APU 690 | Various flagships |
Citation
@misc{dispatchai_benchmark_2026,
title={Per-Chip Benchmark Matrix: On-Device LLM Inference Benchmarks},
author={Aljallaf Alzaabi, Omar Abdulla Jasem},
year={2026},
url={https://huggingface.co/datasets/dispatchAI/per-chip-benchmark-matrix}
}
Dispatch AI (FZE), Sharjah SRTI Free Zone, License No. 10818. Benchmarked on real hardware, $0 cost.
- Downloads last month
- 19