Skip to content

README_GPU.md

Latest commit

 

History

History
118 lines (102 loc) · 4.09 KB

README_GPU.md

File metadata and controls

118 lines (102 loc) · 4.09 KB
Raw

GPU Benchmark Reports

Latest benchmark snapshots written directly by the scripts in this folder.

Flat Arrays

Source: benchmark_flat_arrays.py

Settings:

  • scenario: flat_arrays
  • device: cuda
  • gpu: NVIDIA H100 PCIe
  • torch: 2.5.1+cu121
  • dtype: float32
  • method: simplex
  • tp: 0
  • exclusion_window: 5
  • library_size: all points
  • sample_size: all points
  • batch_size: auto
  • memory_budget_gb: 16.0
  • xtwx_precompute: True
  • xtwy_precompute: False
  • attempts: 3
  • warmups: 1
  • benchmark_cases: [(1000, 1, 1000), (5000, 1, 1000), (20000, 1, 1000), (20000, 1, 2000), (1, 1000, 1000), (1, 5000, 1000), (1, 20000, 1000), (1, 20000, 2000)]
  • x_embedding_dim: 5
  • torch_num_threads: 10
  • torch_num_interop_threads: 1

Results:

n_x n_y ts_length ex ey library_size sample_size exclusion_window attempts avg_sec min_sec max_sec
1000 1 1000 5 1 1000 1000 5 3 0.045312 0.044795 0.046033
5000 1 1000 5 1 1000 1000 5 3 0.245397 0.242002 0.251935
20000 1 1000 5 1 1000 1000 5 3 0.971647 0.969720 0.973339
20000 1 2000 5 1 2000 2000 5 3 2.682948 2.664168 2.720310
1 1000 1000 5 1 1000 1000 5 3 0.005361 0.004885 0.005914
1 5000 1000 5 1 1000 1000 5 3 0.021152 0.019954 0.022942
1 20000 1000 5 1 1000 1000 5 3 0.083439 0.077775 0.094189
1 20000 2000 5 1 2000 2000 5 3 0.112280 0.103756 0.127150

Matrix Performance

Source: benchmark_performance.py

Settings:

  • device: cuda
  • gpu: NVIDIA H100 PCIe
  • torch: 2.5.1+cu121
  • dtype: float32
  • method: simplex
  • tp: 0
  • exclusion_window: 5
  • library_size: all points
  • sample_size: all points
  • batch_size: auto
  • memory_budget_gb: 16.0
  • xtwx_precompute: True
  • xtwy_precompute: False
  • attempts: 3
  • warmups: 1
  • matrix_time_pairs: [(100, 1000), (200, 1000), (800, 500), (100, 8000)]
  • x_embedding_dim: 5
  • torch_num_threads: 10
  • torch_num_interop_threads: 1

Results:

matrix_size ts_length ex ey library_size sample_size exclusion_window attempts avg_sec min_sec max_sec
100 1000 5 1 1000 1000 5 3 0.006838 0.006599 0.007241
200 1000 5 1 1000 1000 5 3 0.013074 0.012861 0.013468
800 500 5 1 500 500 5 3 0.036624 0.036386 0.037025
100 8000 5 1 8000 8000 5 3 0.083923 0.083674 0.084260

Single-Series Self-Prediction

Source: benchmark_single_series_self_prediction.py

Settings:

  • scenario: single-series self-prediction
  • device: cuda
  • gpu: NVIDIA H100 PCIe
  • torch: 2.5.1+cu121
  • dtype: float32
  • method: smap
  • E: 20
  • tau: 1
  • tp: 1
  • exclusion_window: 10
  • library_size: all valid points
  • sample_size: all valid points
  • batch_size: auto
  • memory_budget_gb: 10.0
  • xtwx_precompute: True
  • xtwy_precompute: True
  • attempts: 3
  • warmups: 1
  • lengths: [2000, 8000, 32000, 128000]
  • torch_num_threads: 10
  • torch_num_interop_threads: 1

Results:

length embedded_length library_size sample_size exclusion_window attempts avg_sec min_sec max_sec
2000 1981 1980 1980 10 3 0.001595 0.001465 0.001842
8000 7981 7980 7980 10 3 0.006792 0.006741 0.006818
32000 31981 31980 31980 10 3 0.082072 0.081815 0.082525
128000 127981 127980 127980 10 3 1.391188 1.385473 1.395771