Reproducing on A100 - MatrixMultiplication

[deepn27]

Hello,

I was reproducing the results on a A100 box on google colab. For level 1 -> Square Matrix Multiplication I got ~0.84 score. Below is the final eval. Is this expected? Do you have scores per gpu for each problem in KernelBench logged somewhere?

{
"avg_speedup": 0.8400684292895984,
"accuracy": 1.0,
"total_tokens_sum": 141823,
"num_tasks": 1,
"tasks": [
{
"task": "KernelBench/level1/1_Square_matrix_multiplication_.py",
"best_score": 0.8400684292895984,
"best_runnable": true,
"task_dir": "/content/CudaForge/run/20260123_223852_1_Square_matrix_multiplication__openai_o3/1_Square_matrix_multiplication_",
"figure": "/content/CudaForge/run/20260123_223852_1_Square_matrix_multiplication__openai_o3/1_Square_matrix_multiplication_/figures/1_Square_matrix_multiplication__score.png",
"input_tokens_sum": 64685,
"output_tokens_sum": 77138,
"total_tokens_sum": 141823
}
],
"timestamp": "2026-01-23 23:06:59"
}

[zzj1111]

Hello, it is normal to get this result, for CUDA optimization is highly random. You can try more to see if you can get a better result~