150 lines
3.3 KiB
JSON
150 lines
3.3 KiB
JSON
{
|
|
"metadata": {
|
|
"generated": "2025-12-09",
|
|
"commit": "2250e35",
|
|
"framework": "RustyTorch++",
|
|
"benchmark_framework": "Criterion.rs"
|
|
},
|
|
"system": {
|
|
"os": "Ubuntu 24.04.3 LTS",
|
|
"kernel": "6.8.0-88-generic",
|
|
"cpu": {
|
|
"model": "12th Gen Intel Core i7-12650H",
|
|
"cores": 10,
|
|
"threads": 16,
|
|
"max_freq_mhz": 4700
|
|
},
|
|
"gpu": {
|
|
"model": "NVIDIA GeForce RTX 3050 Ti Laptop GPU",
|
|
"memory_mib": 4096,
|
|
"driver": "580.105.08",
|
|
"compute_capability": "8.6"
|
|
},
|
|
"memory_gb": 32
|
|
},
|
|
"benchmarks": {
|
|
"forward_pass": {
|
|
"lffn_mlp": {
|
|
"200": {
|
|
"time_ms": 3.9889,
|
|
"throughput_kelem_s": 50.139
|
|
},
|
|
"1000": {
|
|
"time_ms": 17.291,
|
|
"throughput_kelem_s": 57.833
|
|
},
|
|
"10000": {
|
|
"time_ms": 167.02,
|
|
"throughput_kelem_s": 59.872
|
|
}
|
|
}
|
|
},
|
|
"forward_pass_optimized": {
|
|
"lffn_mlp_workspace": {
|
|
"200": {
|
|
"time_us": 261.63,
|
|
"throughput_kelem_s": 764.43
|
|
},
|
|
"1000": {
|
|
"time_us": 281.99,
|
|
"throughput_melem_s": 3.5463
|
|
},
|
|
"10000": {
|
|
"time_ms": 1.1822,
|
|
"throughput_melem_s": 8.4587
|
|
},
|
|
"100000": {
|
|
"time_ms": 10.356,
|
|
"throughput_melem_s": 9.6564
|
|
}
|
|
}
|
|
},
|
|
"training_step": {
|
|
"single_step": {
|
|
"200": {
|
|
"time_ms": 4.6902,
|
|
"throughput_kelem_s": 42.642
|
|
},
|
|
"1000": {
|
|
"time_ms": 19.517,
|
|
"throughput_kelem_s": 51.237
|
|
}
|
|
}
|
|
},
|
|
"training_step_optimized": {
|
|
"single_step_workspace": {
|
|
"200": {
|
|
"time_us": 831.88,
|
|
"throughput_kelem_s": 240.42
|
|
},
|
|
"1000": {
|
|
"time_ms": 1.0573,
|
|
"throughput_kelem_s": 945.77
|
|
}
|
|
}
|
|
},
|
|
"training_step_cached": {
|
|
"single_step_cached": {
|
|
"200": {
|
|
"time_ms": 4.7216,
|
|
"throughput_kelem_s": 42.359
|
|
},
|
|
"1000": {
|
|
"time_ms": 19.209,
|
|
"throughput_kelem_s": 52.058
|
|
}
|
|
}
|
|
},
|
|
"training_step_fully_optimized": {
|
|
"single_step_fully_opt": {
|
|
"200": {
|
|
"time_us": 792.81,
|
|
"throughput_kelem_s": 252.27
|
|
},
|
|
"1000": {
|
|
"time_us": 997.76,
|
|
"throughput_melem_s": 1.0022
|
|
}
|
|
}
|
|
},
|
|
"training_100_epochs": {
|
|
"train_100": {
|
|
"200": {
|
|
"time_ms": 472.04,
|
|
"throughput_elem_s": 211.85
|
|
}
|
|
}
|
|
},
|
|
"training_100_epochs_cached": {
|
|
"train_100_cached": {
|
|
"200": {
|
|
"time_ms": 473.43,
|
|
"throughput_elem_s": 211.23
|
|
}
|
|
}
|
|
},
|
|
"training_100_epochs_fully_optimized": {
|
|
"train_100_fully_opt": {
|
|
"200": {
|
|
"time_ms": 79.883,
|
|
"throughput_kelem_s": 1.2518
|
|
}
|
|
}
|
|
}
|
|
},
|
|
"analysis": {
|
|
"forward_pass_speedup": {
|
|
"200_points": "15x",
|
|
"1000_points": "61x",
|
|
"10000_points": "141x"
|
|
},
|
|
"training_step_speedup": {
|
|
"200_points": "5.9x",
|
|
"1000_points": "19.5x"
|
|
},
|
|
"training_100_epochs_speedup": "5.9x",
|
|
"peak_throughput_melem_s": 9.6564,
|
|
"best_training_time_ms": 79.883
|
|
}
|
|
}
|