per-chip-benchmark-matrix / benchmark_matrix.json
3morixd's picture
Upload benchmark_matrix.json with huggingface_hub
a7a1e44 verified
Raw
History Blame Contribute Delete
12 kB
{
"title": "dispatchAI Per-Chip Benchmark Matrix",
"description": "On-device inference benchmarks for mobile LLMs across chipsets",
"hardware": {
"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": "Qualcomm Snapdragon 8 Gen 2",
"cpu": "Kryo CPU (1x3.2GHz + 2x2.8GHz + 2x3.0GHz + 3x2.0GHz)",
"gpu": "Adreno 740",
"npu": "Hexagon NPU ( Qualcomm AI Engine)",
"process_nm": 4,
"devices": [
"Samsung Galaxy S23",
"OnePlus 11"
]
},
"apple_a17_pro": {
"chipset": "Apple A17 Pro",
"cpu": "2x3.78GHz + 4x2.11GHz",
"gpu": "Apple GPU 6-core",
"npu": "Neural Engine 16-core (35 TOPS)",
"process_nm": 3,
"devices": [
"iPhone 15 Pro"
]
},
"mediatek_dimensity_9200": {
"chipset": "MediaTek Dimensity 9200",
"cpu": "1x3.05GHz + 3x2.85GHz + 4x1.8GHz",
"gpu": "Immortalis-G715",
"npu": "APU 690",
"process_nm": 4,
"devices": [
"Various 2023 flagships"
]
}
},
"models": [
{
"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_model": "meta-llama/Llama-3.2-1B-Instruct",
"params_m": 1000,
"quant": "fp16",
"size_mb": 2000,
"architecture": "Llama-3.2",
"task": "text-generation"
},
{
"model_id": "dispatchAI/SmolLM2-360M-Instruct-mobile",
"base_model": "HuggingFaceTB/SmolLM2-360M-Instruct",
"params_m": 360,
"quant": "fp16",
"size_mb": 720,
"architecture": "SmolLM2",
"task": "text-generation"
},
{
"model_id": "dispatchAI/SmolLM2-135M-Instruct-mobile",
"base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
"params_m": 135,
"quant": "fp16",
"size_mb": 270,
"architecture": "SmolLM2",
"task": "text-generation"
},
{
"model_id": "dispatchAI/TinyLlama-1.1B-Chat-Q5-mobile",
"base_model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
"params_m": 1100,
"quant": "Q5_K_M",
"size_mb": 450,
"architecture": "Llama-2",
"task": "text-generation"
},
{
"model_id": "dispatchAI/Qwen2.5-1.5B-Instruct-Q5-mobile",
"base_model": "Qwen/Qwen2.5-1.5B-Instruct",
"params_m": 1500,
"quant": "Q5_K_M",
"size_mb": 900,
"architecture": "Qwen2.5",
"task": "text-generation"
},
{
"model_id": "dispatchAI/Gemma-2-2B-IT-Q5-mobile",
"base_model": "google/gemma-2-2b-it",
"params_m": 2000,
"quant": "Q5_K_M",
"size_mb": 1300,
"architecture": "Gemma-2",
"task": "text-generation"
},
{
"model_id": "dispatchAI/Phi-3.5-mini-instruct-Q5-mobile",
"base_model": "microsoft/Phi-3.5-mini-instruct",
"params_m": 2000,
"quant": "Q5_K_M",
"size_mb": 1300,
"architecture": "Phi-3.5",
"task": "text-generation"
}
],
"results": [
{
"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",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 47.5,
"generation_tps": 9.8,
"tokens_per_second": 9.8,
"notes": "Stable, good quality"
},
{
"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": "R3CN50A0HSV",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 45.2,
"generation_tps": 9.5,
"tokens_per_second": 9.5,
"notes": "Stable"
},
{
"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": "R3CN50AXHWE",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 43.8,
"generation_tps": 9.2,
"tokens_per_second": 9.2,
"notes": "Stable"
},
{
"model_id": "dispatchAI/Llama-3.2-1B-Instruct-mobile",
"base_model": "meta-llama/Llama-3.2-1B-Instruct",
"architecture": "Llama-3.2",
"params_m": 1000,
"quantization": "fp16",
"size_mb": 2000,
"task": "text-generation",
"device_serial": "R3CN509PLHA",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 28.0,
"generation_tps": 6.5,
"tokens_per_second": 6.5,
"notes": "Usable, moderate speed"
},
{
"model_id": "dispatchAI/Llama-3.2-1B-Instruct-mobile",
"base_model": "meta-llama/Llama-3.2-1B-Instruct",
"architecture": "Llama-3.2",
"params_m": 1000,
"quantization": "fp16",
"size_mb": 2000,
"task": "text-generation",
"device_serial": "R3CN50A0HSV",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 27.5,
"generation_tps": 6.3,
"tokens_per_second": 6.3,
"notes": "Usable"
},
{
"model_id": "dispatchAI/SmolLM2-360M-Instruct-mobile",
"base_model": "HuggingFaceTB/SmolLM2-360M-Instruct",
"architecture": "SmolLM2",
"params_m": 360,
"quantization": "fp16",
"size_mb": 720,
"task": "text-generation",
"device_serial": "R3CN509PLHA",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 55.0,
"generation_tps": 12.1,
"tokens_per_second": 12.1,
"notes": "Good speed/quality balance"
},
{
"model_id": "dispatchAI/SmolLM2-360M-Instruct-mobile",
"base_model": "HuggingFaceTB/SmolLM2-360M-Instruct",
"architecture": "SmolLM2",
"params_m": 360,
"quantization": "fp16",
"size_mb": 720,
"task": "text-generation",
"device_serial": "R3CN50A0HSV",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 53.2,
"generation_tps": 11.8,
"tokens_per_second": 11.8,
"notes": "Good balance"
},
{
"model_id": "dispatchAI/SmolLM2-135M-Instruct-mobile",
"base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
"architecture": "SmolLM2",
"params_m": 135,
"quantization": "fp16",
"size_mb": 270,
"task": "text-generation",
"device_serial": "R3CN509PLHA",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 85.0,
"generation_tps": 18.3,
"tokens_per_second": 18.3,
"notes": "Very fast, limited quality"
},
{
"model_id": "dispatchAI/SmolLM2-135M-Instruct-mobile",
"base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
"architecture": "SmolLM2",
"params_m": 135,
"quantization": "fp16",
"size_mb": 270,
"task": "text-generation",
"device_serial": "R3CN50A0HSV",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 82.5,
"generation_tps": 17.8,
"tokens_per_second": 17.8,
"notes": "Very fast"
},
{
"model_id": "dispatchAI/TinyLlama-1.1B-Chat-Q5-mobile",
"base_model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
"architecture": "Llama-2",
"params_m": 1100,
"quantization": "Q5_K_M",
"size_mb": 450,
"task": "text-generation",
"device_serial": "R3CN509PLHA",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 26.5,
"generation_tps": 6.2,
"tokens_per_second": 6.2,
"notes": "Q5 quant, stable"
},
{
"model_id": "dispatchAI/Qwen2.5-1.5B-Instruct-Q5-mobile",
"base_model": "Qwen/Qwen2.5-1.5B-Instruct",
"architecture": "Qwen2.5",
"params_m": 1500,
"quantization": "Q5_K_M",
"size_mb": 900,
"task": "text-generation",
"device_serial": "R3CN509PLHA",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 18.0,
"generation_tps": 4.2,
"tokens_per_second": 4.2,
"notes": "Slower but higher quality"
},
{
"model_id": "dispatchAI/Gemma-2-2B-IT-Q5-mobile",
"base_model": "google/gemma-2-2b-it",
"architecture": "Gemma-2",
"params_m": 2000,
"quantization": "Q5_K_M",
"size_mb": 1300,
"task": "text-generation",
"device_serial": "R3CN509PLHA",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 12.5,
"generation_tps": 3.0,
"tokens_per_second": 3.0,
"notes": "Slow but best quality"
},
{
"model_id": "dispatchAI/Phi-3.5-mini-instruct-Q5-mobile",
"base_model": "microsoft/Phi-3.5-mini-instruct",
"architecture": "Phi-3.5",
"params_m": 2000,
"quantization": "Q5_K_M",
"size_mb": 1300,
"task": "text-generation",
"device_serial": "R3CN509PLHA",
"device_model": "SM-G981V",
"chipset": "Snapdragon 865",
"chipset_family": "Qualcomm",
"cpu_cores": 8,
"ram_gb": 8,
"android_version": 13,
"prompt_tps": 13.0,
"generation_tps": 3.1,
"tokens_per_second": 3.1,
"notes": "Comparable to Gemma"
}
],
"methodology": "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.",
"cost": "$0",
"date": "2026-06-26"
}