{ "schema": "aner.distributed-neural-public-summary/1", "source_sha256": "43aaceb4725f52a5136643b824d790c26600f92378c6cf753287c0b069d5211a", "trial_count": 57, "case_count": 19, "summary": [ { "case": "atlas01-cpu32", "label": "Atlas-01 CPU \u00b7 Float32", "dtype": "Float32", "kind": "standalone", "replicas": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "cpu", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-01/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 3.0628960830072174, "min": 3.0537536250049016, "max": 3.1175427089910954 }, "fit_seconds": { "median": 2.721057125, "min": 2.670434875, "max": 2.743213375 }, "command_seconds": { "median": 3.0174825, "min": 2.961245625, "max": 3.033698916 }, "quality": { "initial_loss": { "median": 1.5991332174875397, "min": 1.5991332174875397, "max": 1.5991332174875397 }, "training_loss": { "median": 0.2923475363134184, "min": 0.2923475363134184, "max": 0.2923475363134184 }, "validation_loss": { "median": 0.3033744259622793, "min": 0.3033744259622793, "max": 0.3033744259622793 }, "test_loss": { "median": 0.29705392923262935, "min": 0.29705392923262935, "max": 0.29705392923262935 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8858596134282808, "min": 0.8858596134282808, "max": 0.8858596134282808 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 0.0, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 0.0, "validation_loss": 0.0, "test_loss": 0.0 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 121389056, "min": 104071168, "max": 127565824 }, "tensor_owned_host_peak_bytes": { "median": 21086868, "min": 21086868, "max": 21086868 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 0, "min": 0, "max": 0 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 0, "min": 0, "max": 0 }, "inner_threads": null, "owned_host_tensor_peak_bytes": { "median": 21086868, "min": 21086868, "max": 21086868 }, "request_write_seconds": { "median": 0, "min": 0, "max": 0 }, "request_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "process_peak_rss_bytes": { "median": 121356288, "min": 104038400, "max": 127533056 }, "samples": { "median": 458760, "min": 458760, "max": 458760 }, "upload_preparation_seconds": { "median": 0.16428429600000008, "min": 0.16247137999999994, "max": 0.16654621 }, "worker_total_seconds": { "median": 1.959291791, "min": 1.9309346270000005, "max": 1.986230381 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "compute_seconds": { "median": 1.7714715769999996, "min": 1.7451295810000005, "max": 1.7956941709999998 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.012126252999999997, "min": 0.012026081000000004, "max": 0.012358166999999995 }, "backend": "apple_accelerate", "target": "macos-arm64" } ], "fastest_selected_baseline": null }, { "case": "atlas01-cpu64", "label": "Atlas-01 CPU \u00b7 Float64", "dtype": "Float64", "kind": "standalone", "replicas": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "cpu", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-01/C03", "numerical_reference_case": "atlas01-cpu64", "repetitions": 3, "outer_command_wall_seconds": { "median": 3.4516792920039734, "min": 3.430949834000785, "max": 3.477942041005008 }, "fit_seconds": { "median": 3.086051291, "min": 3.077315833, "max": 3.097414333 }, "command_seconds": { "median": 3.387386792, "min": 3.37957175, "max": 3.39695175 }, "quality": { "initial_loss": { "median": 1.5991332128378295, "min": 1.5991332128378295, "max": 1.5991332128378295 }, "training_loss": { "median": 0.29234691523654444, "min": 0.29234691523654444, "max": 0.29234691523654444 }, "validation_loss": { "median": 0.30337316499510086, "min": 0.30337316499510086, "max": 0.30337316499510086 }, "test_loss": { "median": 0.2970531787102064, "min": 0.2970531787102064, "max": 0.2970531787102064 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 0.0, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 0.0, "validation_loss": 0.0, "test_loss": 0.0 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 94257152, "min": 94224384, "max": 112525312 }, "tensor_owned_host_peak_bytes": { "median": 16495216, "min": 16495216, "max": 16495216 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 0, "min": 0, "max": 0 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 0, "min": 0, "max": 0 }, "inner_threads": null, "owned_host_tensor_peak_bytes": { "median": 16495216, "min": 16495216, "max": 16495216 }, "request_write_seconds": { "median": 0, "min": 0, "max": 0 }, "request_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "process_peak_rss_bytes": { "median": 94224384, "min": 94191616, "max": 112492544 }, "samples": { "median": 458760, "min": 458760, "max": 458760 }, "upload_preparation_seconds": { "median": 0.107875924, "min": 0.10784828699999997, "max": 0.10974542300000004 }, "worker_total_seconds": { "median": 2.3510604120000003, "min": 2.3420942970000005, "max": 2.3609744629999994 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "compute_seconds": { "median": 2.2227794629999997, "min": 2.213592166000002, "max": 2.2300271619999994 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.009549411, "min": 0.009374091000000001, "max": 0.009962339000000002 }, "backend": "apple_accelerate", "target": "macos-arm64" } ], "fastest_selected_baseline": null }, { "case": "atlas02-cpu32", "label": "Atlas-02 CPU \u00b7 Float32", "dtype": "Float32", "kind": "standalone", "replicas": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cpu", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-02/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 22.104401999997208, "min": 20.88762570900144, "max": 22.852109415995074 }, "fit_seconds": { "median": 19.335128503, "min": 17.864245818, "max": 20.186910193 }, "command_seconds": { "median": 20.064694395, "min": 18.971624858, "max": 21.367686295 }, "quality": { "initial_loss": { "median": 1.599133218017636, "min": 1.599133218017636, "max": 1.599133218017636 }, "training_loss": { "median": 0.29234654818284644, "min": 0.29234654818284644, "max": 0.29234654818284644 }, "validation_loss": { "median": 0.3033732850214472, "min": 0.3033732850214472, "max": 0.3033732850214472 }, "test_loss": { "median": 0.29705295057650866, "min": 0.29705295057650866, "max": 0.29705295057650866 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 4.6350061893463135e-05, "loss_max_abs_errors": { "initial_loss": 5.300964112109341e-10, "training_loss": 9.881305719772193e-07, "validation_loss": 1.1409408320894876e-06, "test_loss": 9.786561206981403e-07 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.00010172939979657958, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 177963008, "min": 177876992, "max": 177995776 }, "tensor_owned_host_peak_bytes": { "median": 21086868, "min": 21086868, "max": 21086868 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 0, "min": 0, "max": 0 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 0, "min": 0, "max": 0 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 21086868, "min": 21086868, "max": 21086868 }, "request_write_seconds": { "median": 0, "min": 0, "max": 0 }, "request_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "process_peak_rss_bytes": { "median": 177963008, "min": 177876992, "max": 177995776 }, "samples": { "median": 458760, "min": 458760, "max": 458760 }, "upload_preparation_seconds": { "median": 1.0965687200000003, "min": 1.0238003210000002, "max": 1.15576639 }, "worker_total_seconds": { "median": 14.506156379000004, "min": 13.444564214000014, "max": 15.424223340000001 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "compute_seconds": { "median": 12.980122186000001, "min": 12.017443807999992, "max": 13.792235244999995 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.064397052, "min": 0.060066732000000025, "max": 0.07318645099999999 }, "backend": "openblas", "target": "linux-x86_64" } ], "fastest_selected_baseline": null }, { "case": "atlas02-cpu64", "label": "Atlas-02 CPU \u00b7 Float64", "dtype": "Float64", "kind": "standalone", "replicas": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cpu", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-02/C03", "numerical_reference_case": "atlas01-cpu64", "repetitions": 3, "outer_command_wall_seconds": { "median": 20.65155120800773, "min": 19.701527957993676, "max": 20.810743416994228 }, "fit_seconds": { "median": 18.017975232, "min": 16.659829462, "max": 18.071835351 }, "command_seconds": { "median": 18.795863525, "min": 17.671156699, "max": 19.072891984 }, "quality": { "initial_loss": { "median": 1.5991332128378295, "min": 1.5991332128378295, "max": 1.5991332128378295 }, "training_loss": { "median": 0.29234691523654444, "min": 0.29234691523654444, "max": 0.29234691523654444 }, "validation_loss": { "median": 0.30337316499510086, "min": 0.30337316499510086, "max": 0.30337316499510086 }, "test_loss": { "median": 0.29705317871020637, "min": 0.29705317871020637, "max": 0.29705317871020637 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 3.3306690738754696e-16, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 0.0, "validation_loss": 0.0, "test_loss": 5.551115123125783e-17 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 172453888, "min": 172425216, "max": 172531712 }, "tensor_owned_host_peak_bytes": { "median": 16495216, "min": 16495216, "max": 16495216 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 0, "min": 0, "max": 0 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 0, "min": 0, "max": 0 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 16495216, "min": 16495216, "max": 16495216 }, "request_write_seconds": { "median": 0, "min": 0, "max": 0 }, "request_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "process_peak_rss_bytes": { "median": 172453888, "min": 172425216, "max": 172531712 }, "samples": { "median": 458760, "min": 458760, "max": 458760 }, "upload_preparation_seconds": { "median": 0.5774402320000003, "min": 0.5563793830000007, "max": 0.6052092589999998 }, "worker_total_seconds": { "median": 13.412996769999998, "min": 12.879265018999998, "max": 14.06670137800001 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "compute_seconds": { "median": 12.507452861000004, "min": 12.006613135999997, "max": 13.120414492000009 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.050057721999999985, "min": 0.048369314, "max": 0.05226979300000003 }, "backend": "openblas", "target": "linux-x86_64" } ], "fastest_selected_baseline": null }, { "case": "atlas03-cpu32", "label": "Atlas-03 CPU \u00b7 Float32", "dtype": "Float32", "kind": "standalone", "replicas": [ { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cpu", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-03/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 14.828693917006603, "min": 14.590928958001314, "max": 15.160592833010014 }, "fit_seconds": { "median": 12.708668223, "min": 12.587794052, "max": 12.860272355 }, "command_seconds": { "median": 13.231820473, "min": 13.101270396, "max": 13.380385378 }, "quality": { "initial_loss": { "median": 1.5991332140263232, "min": 1.5991332140263232, "max": 1.5991332140263232 }, "training_loss": { "median": 0.2923465511763311, "min": 0.2923465511763311, "max": 0.2923465511763311 }, "validation_loss": { "median": 0.3033732850214472, "min": 0.3033732850214472, "max": 0.3033732850214472 }, "test_loss": { "median": 0.29705295057650866, "min": 0.29705295057650866, "max": 0.29705295057650866 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 4.635751247406006e-05, "loss_max_abs_errors": { "initial_loss": 3.461216424582858e-09, "training_loss": 9.851370872948628e-07, "validation_loss": 1.1409408320894876e-06, "test_loss": 9.786561206981403e-07 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.00010172939979657958, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 177893376, "min": 177852416, "max": 177975296 }, "tensor_owned_host_peak_bytes": { "median": 21086868, "min": 21086868, "max": 21086868 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 0, "min": 0, "max": 0 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 0, "min": 0, "max": 0 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 21086868, "min": 21086868, "max": 21086868 }, "request_write_seconds": { "median": 0, "min": 0, "max": 0 }, "request_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "process_peak_rss_bytes": { "median": 177893376, "min": 177852416, "max": 177975296 }, "samples": { "median": 458760, "min": 458760, "max": 458760 }, "upload_preparation_seconds": { "median": 0.7007955530000008, "min": 0.6872004380000003, "max": 0.7012716800000004 }, "worker_total_seconds": { "median": 9.686538782000001, "min": 9.584013947999999, "max": 9.768997478999998 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "compute_seconds": { "median": 8.630125725999997, "min": 8.543419429, "max": 8.709238542999996 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.045808335000000054, "min": 0.045342648, "max": 0.04580889400000001 }, "backend": "openblas", "target": "linux-x86_64" } ], "fastest_selected_baseline": null }, { "case": "atlas03-cpu64", "label": "Atlas-03 CPU \u00b7 Float64", "dtype": "Float64", "kind": "standalone", "replicas": [ { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cpu", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-03/C03", "numerical_reference_case": "atlas01-cpu64", "repetitions": 3, "outer_command_wall_seconds": { "median": 13.518542040998, "min": 13.406571915998938, "max": 13.942002208001213 }, "fit_seconds": { "median": 11.45661501, "min": 11.436723592, "max": 11.479060799 }, "command_seconds": { "median": 11.956218479, "min": 11.936004619, "max": 11.984865174 }, "quality": { "initial_loss": { "median": 1.5991332128378295, "min": 1.5991332128378295, "max": 1.5991332128378295 }, "training_loss": { "median": 0.29234691523654444, "min": 0.29234691523654444, "max": 0.29234691523654444 }, "validation_loss": { "median": 0.3033731649951008, "min": 0.3033731649951008, "max": 0.3033731649951008 }, "test_loss": { "median": 0.29705317871020637, "min": 0.29705317871020637, "max": 0.29705317871020637 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 2.220446049250313e-16, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 0.0, "validation_loss": 5.551115123125783e-17, "test_loss": 5.551115123125783e-17 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 172462080, "min": 172457984, "max": 172683264 }, "tensor_owned_host_peak_bytes": { "median": 16495216, "min": 16495216, "max": 16495216 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 0, "min": 0, "max": 0 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 0, "min": 0, "max": 0 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 16495216, "min": 16495216, "max": 16495216 }, "request_write_seconds": { "median": 0, "min": 0, "max": 0 }, "request_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "process_peak_rss_bytes": { "median": 172462080, "min": 172457984, "max": 172683264 }, "samples": { "median": 458760, "min": 458760, "max": 458760 }, "upload_preparation_seconds": { "median": 0.3564620449999999, "min": 0.35629507399999993, "max": 0.3585980489999996 }, "worker_total_seconds": { "median": 8.894932573000004, "min": 8.886005835000004, "max": 8.897904400999998 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "compute_seconds": { "median": 8.261263486, "min": 8.256238315999996, "max": 8.265567833999997 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.032498655, "min": 0.03217572499999999, "max": 0.034002631000000005 }, "backend": "openblas", "target": "linux-x86_64" } ], "fastest_selected_baseline": null }, { "case": "atlas01-gpu32", "label": "Atlas-01 MPS \u00b7 Float32", "dtype": "Float32", "kind": "standalone", "replicas": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "mps", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-01/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 3.2152430829883087, "min": 3.0367226250091335, "max": 3.255333624998457 }, "fit_seconds": { "median": 2.804163125, "min": 2.632902875, "max": 2.816833625 }, "command_seconds": { "median": 3.161081833, "min": 2.984413792, "max": 3.189324208 }, "quality": { "initial_loss": { "median": 1.5991332174875397, "min": 1.5991332174875397, "max": 1.5991332174875397 }, "training_loss": { "median": 0.29234693766326264, "min": 0.29234693766326264, "max": 0.29234693766326264 }, "validation_loss": { "median": 0.3033739387867773, "min": 0.3033739387867773, "max": 0.3033739387867773 }, "test_loss": { "median": 0.2970536025106482, "min": 0.2970536025106482, "max": 0.2970536025106482 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 4.6253204345703125e-05, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 5.986501557808666e-07, "validation_loss": 4.87175501995285e-07, "test_loss": 3.2672198113647966e-07 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.00010172939979657958, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 85131264, "min": 84983808, "max": 87965696 }, "tensor_owned_host_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "tensor_owned_device_peak_bytes": { "median": 6524456, "min": 6524456, "max": 6524456 } }, "replica_metrics": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "mps", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 0, "min": 0, "max": 0 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 0, "min": 0, "max": 0 }, "inner_threads": null, "owned_host_tensor_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "request_write_seconds": { "median": 0, "min": 0, "max": 0 }, "request_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "process_peak_rss_bytes": { "median": 85131264, "min": 84983808, "max": 87965696 }, "samples": { "median": 458760, "min": 458760, "max": 458760 }, "upload_preparation_seconds": { "median": 0.24405004799999996, "min": 0.23252137699999997, "max": 0.250555169 }, "worker_total_seconds": { "median": 1.9918613789999977, "min": 1.858832952, "max": 2.0093898709999993 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "compute_seconds": { "median": 1.7054683319999997, "min": 1.5886115010000001, "max": 1.7175838759999995 }, "owned_device_tensor_peak_bytes": { "median": 6524456, "min": 6524456, "max": 6524456 }, "readback_seconds": { "median": 0.016662950999999995, "min": 0.015358756000000006, "max": 0.017307832999999998 }, "backend": "metal_mps", "target": "macos-arm64" } ], "fastest_selected_baseline": null }, { "case": "atlas02-gpu32", "label": "Atlas-02 CUDA \u00b7 Float32", "dtype": "Float32", "kind": "standalone", "replicas": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-02/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 8.69213504200161, "min": 8.515092125002411, "max": 8.770283082994865 }, "fit_seconds": { "median": 5.944281978, "min": 5.904053987, "max": 6.029027588 }, "command_seconds": { "median": 7.097920784, "min": 7.032981553, "max": 7.155840428 }, "quality": { "initial_loss": { "median": 1.599133218017636, "min": 1.599133218017636, "max": 1.599133218017636 }, "training_loss": { "median": 0.2923495870296973, "min": 0.2923495870296973, "max": 0.2923495870296973 }, "validation_loss": { "median": 0.30337606328913785, "min": 0.30337606328913785, "max": 0.30337606328913785 }, "test_loss": { "median": 0.29705589171101093, "min": 0.29705589171101093, "max": 0.29705589171101093 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 5.9798359870910645e-05, "loss_max_abs_errors": { "initial_loss": 5.300964112109341e-10, "training_loss": 2.050716278900744e-06, "validation_loss": 1.6373268585390122e-06, "test_loss": 1.9624783815785918e-06 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.00010172939979657958, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 296333312, "min": 296271872, "max": 296456192 }, "tensor_owned_host_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "tensor_owned_device_peak_bytes": { "median": 6524456, "min": 6524456, "max": 6524456 } }, "replica_metrics": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 0, "min": 0, "max": 0 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 0, "min": 0, "max": 0 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "request_write_seconds": { "median": 0, "min": 0, "max": 0 }, "request_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "process_peak_rss_bytes": { "median": 296333312, "min": 296271872, "max": 296456192 }, "samples": { "median": 458760, "min": 458760, "max": 458760 }, "upload_preparation_seconds": { "median": 0.9739203370000002, "min": 0.9662449480000002, "max": 0.9876225599999999 }, "worker_total_seconds": { "median": 1.6844404309999996, "min": 1.6604763869999994, "max": 1.703028964 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "compute_seconds": { "median": 0.5040637470000001, "min": 0.48920735600000026, "max": 0.5048317030000001 }, "owned_device_tensor_peak_bytes": { "median": 6524456, "min": 6524456, "max": 6524456 }, "readback_seconds": { "median": 0.06745282600000004, "min": 0.06735880900000002, "max": 0.07029516799999998 }, "backend": "cuda_cublas", "target": "linux-x86_64" } ], "fastest_selected_baseline": null }, { "case": "atlas02-gpu64", "label": "Atlas-02 CUDA \u00b7 Float64", "dtype": "Float64", "kind": "standalone", "replicas": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-02/C03", "numerical_reference_case": "atlas01-cpu64", "repetitions": 3, "outer_command_wall_seconds": { "median": 9.508705583997653, "min": 8.796772208006587, "max": 9.726846584002487 }, "fit_seconds": { "median": 5.958052482, "min": 5.843198569, "max": 6.5129025 }, "command_seconds": { "median": 7.217335922, "min": 7.141771782, "max": 7.993388733 }, "quality": { "initial_loss": { "median": 1.5991332128378295, "min": 1.5991332128378295, "max": 1.5991332128378295 }, "training_loss": { "median": 0.29234691523654444, "min": 0.29234691523654444, "max": 0.29234691523654444 }, "validation_loss": { "median": 0.3033731649951009, "min": 0.3033731649951009, "max": 0.3033731649951009 }, "test_loss": { "median": 0.29705317871020637, "min": 0.29705317871020637, "max": 0.29705317871020637 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 3.3306690738754696e-16, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 0.0, "validation_loss": 5.551115123125783e-17, "test_loss": 5.551115123125783e-17 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 300347392, "min": 300224512, "max": 300429312 }, "tensor_owned_host_peak_bytes": { "median": 4202560, "min": 4202560, "max": 4202560 }, "tensor_owned_device_peak_bytes": { "median": 13048912, "min": 13048912, "max": 13048912 } }, "replica_metrics": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 0, "min": 0, "max": 0 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 0, "min": 0, "max": 0 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 4202560, "min": 4202560, "max": 4202560 }, "request_write_seconds": { "median": 0, "min": 0, "max": 0 }, "request_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "process_peak_rss_bytes": { "median": 300347392, "min": 300224512, "max": 300429312 }, "samples": { "median": 458760, "min": 458760, "max": 458760 }, "upload_preparation_seconds": { "median": 0.6689899059999999, "min": 0.6621139690000001, "max": 0.8554335600000007 }, "worker_total_seconds": { "median": 2.260629614, "min": 2.258230271999999, "max": 2.575395199 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "compute_seconds": { "median": 1.2547439029999992, "min": 1.2478117600000003, "max": 1.3179493970000007 }, "owned_device_tensor_peak_bytes": { "median": 13048912, "min": 13048912, "max": 13048912 }, "readback_seconds": { "median": 0.07557949400000004, "min": 0.075353733, "max": 0.11018652099999998 }, "backend": "cuda_cublas", "target": "linux-x86_64" } ], "fastest_selected_baseline": null }, { "case": "atlas03-gpu32", "label": "Atlas-03 CUDA \u00b7 Float32", "dtype": "Float32", "kind": "standalone", "replicas": [ { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-03/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 6.113490791001823, "min": 6.021835834006197, "max": 6.598680917013553 }, "fit_seconds": { "median": 4.041986648, "min": 4.025347966, "max": 4.045879262 }, "command_seconds": { "median": 4.649221081, "min": 4.631889841, "max": 4.670013525 }, "quality": { "initial_loss": { "median": 1.5991332140263232, "min": 1.5991332140263232, "max": 1.5991332140263232 }, "training_loss": { "median": 0.2923475264676603, "min": 0.2923475264676603, "max": 0.2923475264676603 }, "validation_loss": { "median": 0.30337433190457336, "min": 0.30337433190457336, "max": 0.30337433190457336 }, "test_loss": { "median": 0.2970538742908505, "min": 0.2970538742908505, "max": 0.2970538742908505 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8858596134282808, "min": 0.8858596134282808, "max": 0.8858596134282808 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 4.561617970466614e-06, "loss_max_abs_errors": { "initial_loss": 3.461216424582858e-09, "training_loss": 9.84575809770405e-09, "validation_loss": 9.405770595716234e-08, "test_loss": 5.494177884868279e-08 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 331112448, "min": 331001856, "max": 331165696 }, "tensor_owned_host_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "tensor_owned_device_peak_bytes": { "median": 6524456, "min": 6524456, "max": 6524456 } }, "replica_metrics": [ { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 0, "min": 0, "max": 0 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 0, "min": 0, "max": 0 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "request_write_seconds": { "median": 0, "min": 0, "max": 0 }, "request_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "process_peak_rss_bytes": { "median": 331112448, "min": 331001856, "max": 331165696 }, "samples": { "median": 458760, "min": 458760, "max": 458760 }, "upload_preparation_seconds": { "median": 0.6526943540000001, "min": 0.6446049009999999, "max": 0.6535847469999994 }, "worker_total_seconds": { "median": 1.0821805899999999, "min": 1.0742639050000002, "max": 1.086779123 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "compute_seconds": { "median": 0.2802974630000002, "min": 0.2801538039999999, "max": 0.28371388599999997 }, "owned_device_tensor_peak_bytes": { "median": 6524456, "min": 6524456, "max": 6524456 }, "readback_seconds": { "median": 0.05001958200000001, "min": 0.049093760000000014, "max": 0.051010612000000004 }, "backend": "cuda_cublas", "target": "linux-x86_64" } ], "fastest_selected_baseline": null }, { "case": "atlas03-gpu64", "label": "Atlas-03 CUDA \u00b7 Float64", "dtype": "Float64", "kind": "standalone", "replicas": [ { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-03/C03", "numerical_reference_case": "atlas01-cpu64", "repetitions": 3, "outer_command_wall_seconds": { "median": 5.359644708005362, "min": 5.325457874991116, "max": 5.440311792001012 }, "fit_seconds": { "median": 3.380031545, "min": 3.351845387, "max": 3.423176061 }, "command_seconds": { "median": 3.955525727, "min": 3.920242845, "max": 4.007730675 }, "quality": { "initial_loss": { "median": 1.5991332128378295, "min": 1.5991332128378295, "max": 1.5991332128378295 }, "training_loss": { "median": 0.29234691523654444, "min": 0.29234691523654444, "max": 0.29234691523654444 }, "validation_loss": { "median": 0.3033731649951008, "min": 0.3033731649951008, "max": 0.3033731649951008 }, "test_loss": { "median": 0.29705317871020637, "min": 0.29705317871020637, "max": 0.29705317871020637 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 3.3306690738754696e-16, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 0.0, "validation_loss": 5.551115123125783e-17, "test_loss": 5.551115123125783e-17 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 324956160, "min": 324931584, "max": 325730304 }, "tensor_owned_host_peak_bytes": { "median": 4202560, "min": 4202560, "max": 4202560 }, "tensor_owned_device_peak_bytes": { "median": 13048912, "min": 13048912, "max": 13048912 } }, "replica_metrics": [ { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 0, "min": 0, "max": 0 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 0, "min": 0, "max": 0 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 4202560, "min": 4202560, "max": 4202560 }, "request_write_seconds": { "median": 0, "min": 0, "max": 0 }, "request_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "process_peak_rss_bytes": { "median": 324943872, "min": 324931584, "max": 324956160 }, "samples": { "median": 458760, "min": 458760, "max": 458760 }, "upload_preparation_seconds": { "median": 0.40694497400000007, "min": 0.4014066510000002, "max": 0.41219602099999997 }, "worker_total_seconds": { "median": 1.0275088589999999, "min": 1.0095150980000012, "max": 1.0351131310000008 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 0, "min": 0, "max": 0 }, "compute_seconds": { "median": 0.4143261709999999, "min": 0.40619233300000024, "max": 0.4147275310000002 }, "owned_device_tensor_peak_bytes": { "median": 13048912, "min": 13048912, "max": 13048912 }, "readback_seconds": { "median": 0.055599096000000035, "min": 0.05416801799999993, "max": 0.05594505700000002 }, "backend": "cuda_cublas", "target": "linux-x86_64" } ], "fastest_selected_baseline": null }, { "case": "01-02-gpu32", "label": "Atlas-01 MPS + Atlas-02 CUDA \u00b7 Float32", "dtype": "Float32", "kind": "group", "replicas": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "mps", "relative_batch_weight": 1 }, { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-01/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 66.8675972080091, "min": 64.88779833298759, "max": 69.97771516699868 }, "fit_seconds": { "median": 66.406795041, "min": 64.414978083, "max": 69.533544166 }, "command_seconds": { "median": 66.80565725, "min": 64.804851625, "max": 69.92076175 }, "quality": { "initial_loss": { "median": 1.5991332174875397, "min": 1.5991332174875397, "max": 1.5991332174875397 }, "training_loss": { "median": 0.29234811506937414, "min": 0.29234811506937414, "max": 0.29234811506937414 }, "validation_loss": { "median": 0.30337470066511013, "min": 0.30337470066511013, "max": 0.30337470066511013 }, "test_loss": { "median": 0.29705493784872467, "min": 0.29705493784872467, "max": 0.29705493784872467 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8858596134282808, "min": 0.8858596134282808, "max": 0.8858596134282808 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 3.8526952266693115e-05, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 5.787559557157707e-07, "validation_loss": 2.747028308158761e-07, "test_loss": 1.008616095310444e-06 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 77430784, "min": 74383360, "max": 77496320 }, "tensor_owned_host_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "mps", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 5.688616882999999, "min": 5.540511544000001, "max": 5.730888381 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 6.897081250000001, "min": 6.608140246, "max": 6.926093207000001 }, "inner_threads": null, "owned_host_tensor_peak_bytes": { "median": 983040, "min": 983040, "max": 983040 }, "request_write_seconds": { "median": 1.195105451000001, "min": 1.0675435869999996, "max": 1.2083657930000002 }, "request_payload_bytes": { "median": 82667760, "min": 82667760, "max": 82667760 }, "process_peak_rss_bytes": { "median": 24199168, "min": 23445504, "max": 25493504 }, "samples": { "median": 229380, "min": 229380, "max": 229380 }, "upload_preparation_seconds": { "median": 0.291751674, "min": 0.267110162, "max": 0.29392679999999993 }, "worker_total_seconds": { "median": 3.5903741669999976, "min": 3.542504004000001, "max": 3.6027743710000024 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 3.235207667999998, "min": 3.1871155310000003, "max": 3.2784082180000005 }, "owned_device_tensor_peak_bytes": { "median": 3329576, "min": 3329576, "max": 3329576 }, "readback_seconds": { "median": 0.022639126, "min": 0.021237366000000004, "max": 0.02372707299999998 }, "backend": "metal_mps", "target": "macos-arm64" }, { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 60.40910454899999, "min": 59.03837270800001, "max": 63.661730917999975 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 61.83582804299997, "min": 60.31867120499993, "max": 65.10478625400002 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 983040, "min": 983040, "max": 983040 }, "request_write_seconds": { "median": 1.4266434240000008, "min": 1.280212417, "max": 1.4429274139999988 }, "request_payload_bytes": { "median": 82667760, "min": 82667760, "max": 82667760 }, "process_peak_rss_bytes": { "median": 252841984, "min": 252731392, "max": 252841984 }, "samples": { "median": 229380, "min": 229380, "max": 229380 }, "upload_preparation_seconds": { "median": 1.6837423830000011, "min": 1.6343725100000008, "max": 1.695894748 }, "worker_total_seconds": { "median": 2.8938203080000013, "min": 2.8092793789999995, "max": 2.920432863 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 0.802364632, "min": 0.7812318650000001, "max": 0.8144459129999998 }, "owned_device_tensor_peak_bytes": { "median": 3329576, "min": 3329576, "max": 3329576 }, "readback_seconds": { "median": 0.14952447300000007, "min": 0.14299656899999993, "max": 0.15217154900000004 }, "backend": "cuda_cublas", "target": "linux-x86_64" } ], "fastest_selected_baseline": { "fit_seconds": { "case": "atlas01-gpu32", "seconds": 2.804163125, "speedup": 0.042227051061095346 }, "outer_command_wall_seconds": { "case": "atlas01-gpu32", "seconds": 3.2152430829883087, "speedup": 0.048083723914685565 } } }, { "case": "01-03-gpu32", "label": "Atlas-01 MPS + Atlas-03 CUDA \u00b7 Float32", "dtype": "Float32", "kind": "group", "replicas": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "mps", "relative_batch_weight": 1 }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-01/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 59.04362216699519, "min": 57.645449082992855, "max": 64.85838504199637 }, "fit_seconds": { "median": 58.597174292, "min": 57.208868792, "max": 64.432534459 }, "command_seconds": { "median": 58.984112334, "min": 57.587268958, "max": 64.820838792 }, "quality": { "initial_loss": { "median": 1.5991332174875397, "min": 1.5991332174875397, "max": 1.5991332174875397 }, "training_loss": { "median": 0.2923502071331597, "min": 0.2923502071331597, "max": 0.2923502071331597 }, "validation_loss": { "median": 0.30337603640338046, "min": 0.30337603640338046, "max": 0.30337603640338046 }, "test_loss": { "median": 0.29705585060017764, "min": 0.29705585060017764, "max": 0.29705585060017764 }, "training_accuracy": { "median": 0.8984000348766239, "min": 0.8984000348766239, "max": 0.8984000348766239 }, "validation_accuracy": { "median": 0.8858596134282808, "min": 0.8858596134282808, "max": 0.8858596134282808 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 5.525164306163788e-05, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 2.6708197412572687e-06, "validation_loss": 1.610441101140836e-06, "test_loss": 1.9213675482854242e-06 }, "accuracy_max_abs_errors": { "training_accuracy": 4.359577992840702e-05, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 77758464, "min": 75169792, "max": 78020608 }, "tensor_owned_host_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "mps", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 5.850273453999998, "min": 5.816193624999996, "max": 5.858686630999997 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 7.047510581000003, "min": 7.023647721999998, "max": 7.047991165999994 }, "inner_threads": null, "owned_host_tensor_peak_bytes": { "median": 983040, "min": 983040, "max": 983040 }, "request_write_seconds": { "median": 1.1976290399999998, "min": 1.188724290000001, "max": 1.2073680439999999 }, "request_payload_bytes": { "median": 82667760, "min": 82667760, "max": 82667760 }, "process_peak_rss_bytes": { "median": 24788992, "min": 24231936, "max": 25034752 }, "samples": { "median": 229380, "min": 229380, "max": 229380 }, "upload_preparation_seconds": { "median": 0.292591912, "min": 0.28917641099999997, "max": 0.2974659540000005 }, "worker_total_seconds": { "median": 3.697802285, "min": 3.6478931899999996, "max": 3.705357743999997 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 3.3420961660000015, "min": 3.296334863, "max": 3.346444957999999 }, "owned_device_tensor_peak_bytes": { "median": 3329576, "min": 3329576, "max": 3329576 }, "readback_seconds": { "median": 0.023418092999999997, "min": 0.022949124999999994, "max": 0.023648492000000004 }, "backend": "metal_mps", "target": "macos-arm64" }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 53.80685994000001, "min": 52.18226125400003, "max": 59.355597614000004 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 55.249875174000046, "min": 53.59055008700001, "max": 60.785204422000014 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 983040, "min": 983040, "max": 983040 }, "request_write_seconds": { "median": 1.4294870309999994, "min": 1.4080964589999994, "max": 1.442951583 }, "request_payload_bytes": { "median": 82667760, "min": 82667760, "max": 82667760 }, "process_peak_rss_bytes": { "median": 286941184, "min": 286707712, "max": 287035392 }, "samples": { "median": 229380, "min": 229380, "max": 229380 }, "upload_preparation_seconds": { "median": 1.0515673750000007, "min": 1.0400027480000016, "max": 1.0665977910000004 }, "worker_total_seconds": { "median": 1.6082218649999993, "min": 1.5876060520000006, "max": 1.6248352789999998 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 0.359499999, "min": 0.352285454, "max": 0.36445149700000007 }, "owned_device_tensor_peak_bytes": { "median": 3329576, "min": 3329576, "max": 3329576 }, "readback_seconds": { "median": 0.06346413999999996, "min": 0.06323915100000001, "max": 0.06348462099999996 }, "backend": "cuda_cublas", "target": "linux-x86_64" } ], "fastest_selected_baseline": { "fit_seconds": { "case": "atlas01-gpu32", "seconds": 2.804163125, "speedup": 0.047854920631946574 }, "outer_command_wall_seconds": { "case": "atlas01-gpu32", "seconds": 3.2152430829883087, "speedup": 0.05445538340948192 } } }, { "case": "02-03-gpu32", "label": "Atlas-02 CUDA + Atlas-03 CUDA \u00b7 Float32", "dtype": "Float32", "kind": "group", "replicas": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1 }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-01/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 91.10916891599481, "min": 82.52440370799741, "max": 92.16468412500399 }, "fit_seconds": { "median": 90.603120209, "min": 82.012750791, "max": 91.678021833 }, "command_seconds": { "median": 91.037551042, "min": 82.4643615, "max": 92.129275166 }, "quality": { "initial_loss": { "median": 1.5991332174875397, "min": 1.5991332174875397, "max": 1.5991332174875397 }, "training_loss": { "median": 0.2923499547917588, "min": 0.2923499547917588, "max": 0.2923499547917588 }, "validation_loss": { "median": 0.3033758939464604, "min": 0.3033758939464604, "max": 0.3033758939464604 }, "test_loss": { "median": 0.2970558105929094, "min": 0.2970558105929094, "max": 0.2970558105929094 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 5.7095661759376526e-05, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 2.418478340371344e-06, "validation_loss": 1.4679841811027217e-06, "test_loss": 1.8813602800649143e-06 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.00010172939979657958, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 77627392, "min": 75268096, "max": 77725696 }, "tensor_owned_host_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 76.93777121300002, "min": 67.15449578600001, "max": 79.19499566500002 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 78.28922679199995, "min": 68.58777321100001, "max": 80.50158036999997 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 983040, "min": 983040, "max": 983040 }, "request_write_seconds": { "median": 1.3513404990000002, "min": 1.3065204980000003, "max": 1.4331425359999996 }, "request_payload_bytes": { "median": 82667760, "min": 82667760, "max": 82667760 }, "process_peak_rss_bytes": { "median": 252796928, "min": 252645376, "max": 252858368 }, "samples": { "median": 229380, "min": 229380, "max": 229380 }, "upload_preparation_seconds": { "median": 1.6798352950000002, "min": 1.6567668230000003, "max": 1.7106720770000006 }, "worker_total_seconds": { "median": 2.9094893740000005, "min": 2.8483384549999995, "max": 2.9448951880000003 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 0.8207718829999999, "min": 0.7925328389999999, "max": 0.8241611529999998 }, "owned_device_tensor_peak_bytes": { "median": 3329576, "min": 3329576, "max": 3329576 }, "readback_seconds": { "median": 0.15175968399999998, "min": 0.14461619200000006, "max": 0.15176089000000012 }, "backend": "cuda_cublas", "target": "linux-x86_64" }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 72.59485933, "min": 66.43971983499999, "max": 73.17139229200005 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 73.94015379599995, "min": 67.86396450300003, "max": 74.47368540199999 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 983040, "min": 983040, "max": 983040 }, "request_write_seconds": { "median": 1.3452226289999993, "min": 1.3022185680000002, "max": 1.424187879999999 }, "request_payload_bytes": { "median": 82667760, "min": 82667760, "max": 82667760 }, "process_peak_rss_bytes": { "median": 286883840, "min": 286822400, "max": 286904320 }, "samples": { "median": 229380, "min": 229380, "max": 229380 }, "upload_preparation_seconds": { "median": 1.0825509750000002, "min": 1.0510841480000002, "max": 1.092290365000001 }, "worker_total_seconds": { "median": 1.646046648999999, "min": 1.6180502700000008, "max": 1.6464566009999997 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 0.3666001520000001, "min": 0.36056593099999973, "max": 0.3691265440000001 }, "owned_device_tensor_peak_bytes": { "median": 3329576, "min": 3329576, "max": 3329576 }, "readback_seconds": { "median": 0.06239568800000004, "min": 0.06162792100000003, "max": 0.06479273999999999 }, "backend": "cuda_cublas", "target": "linux-x86_64" } ], "fastest_selected_baseline": { "fit_seconds": { "case": "atlas03-gpu32", "seconds": 4.041986648, "speedup": 0.04461200275085551 }, "outer_command_wall_seconds": { "case": "atlas03-gpu32", "seconds": 6.113490791001823, "speedup": 0.06710071954051772 } } }, { "case": "01-02-03-gpu32", "label": "Atlas-01 MPS + Atlas-02 CUDA + Atlas-03 CUDA \u00b7 Float32", "dtype": "Float32", "kind": "group", "replicas": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "mps", "relative_batch_weight": 1 }, { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1 }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-01/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 86.09376583401172, "min": 86.07671895899693, "max": 86.79663883299509 }, "fit_seconds": { "median": 85.590917916, "min": 85.56780825, "max": 86.292820584 }, "command_seconds": { "median": 86.053667791, "min": 86.037888583, "max": 86.745891042 }, "quality": { "initial_loss": { "median": 1.5991332174875397, "min": 1.5991332174875397, "max": 1.5991332174875397 }, "training_loss": { "median": 0.2923465431937054, "min": 0.2923465431937054, "max": 0.2923465431937054 }, "validation_loss": { "median": 0.30337328036464484, "min": 0.30337328036464484, "max": 0.30337328036464484 }, "test_loss": { "median": 0.29705295057650866, "min": 0.29705295057650866, "max": 0.29705295057650866 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 4.633516073226929e-05, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 9.931197130219616e-07, "validation_loss": 1.1455976344776353e-06, "test_loss": 9.786561206981403e-07 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.00010172939979657958, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 77922304, "min": 76365824, "max": 77938688 }, "tensor_owned_host_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "mps", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 5.396259586000003, "min": 5.224934279000005, "max": 5.417965785999997 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 6.3356772550000064, "min": 6.0719885110000025, "max": 6.368811862000002 }, "inner_threads": null, "owned_host_tensor_peak_bytes": { "median": 655360, "min": 655360, "max": 655360 }, "request_write_seconds": { "median": 0.9393523770000002, "min": 0.8469903709999993, "max": 0.9507761470000002 }, "request_payload_bytes": { "median": 61870640, "min": 61870640, "max": 61870640 }, "process_peak_rss_bytes": { "median": 21807104, "min": 21774336, "max": 22003712 }, "samples": { "median": 152920, "min": 152920, "max": 152920 }, "upload_preparation_seconds": { "median": 0.24906942500000023, "min": 0.2199126280000001, "max": 0.2503211539999999 }, "worker_total_seconds": { "median": 3.480475675999999, "min": 3.4586895430000006, "max": 3.4963926999999972 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 3.1651296360000023, "min": 3.146254414000001, "max": 3.215899031000002 }, "owned_device_tensor_peak_bytes": { "median": 2264616, "min": 2264616, "max": 2264616 }, "readback_seconds": { "median": 0.023416451999999997, "min": 0.023044838999999998, "max": 0.024552329999999994 }, "backend": "metal_mps", "target": "macos-arm64" }, { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 74.27892257999996, "min": 73.03388654199996, "max": 75.50457230200001 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 75.22893874099996, "min": 74.09127016700003, "max": 76.56417875299995 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 655360, "min": 655360, "max": 655360 }, "request_write_seconds": { "median": 1.057320459, "min": 0.9499169850000007, "max": 1.0594542839999999 }, "request_payload_bytes": { "median": 61870640, "min": 61870640, "max": 61870640 }, "process_peak_rss_bytes": { "median": 252145664, "min": 252051456, "max": 252153856 }, "samples": { "median": 152920, "min": 152920, "max": 152920 }, "upload_preparation_seconds": { "median": 1.307947765000001, "min": 1.3072958470000007, "max": 1.3187351210000007 }, "worker_total_seconds": { "median": 2.478999484000003, "min": 2.474586277, "max": 2.511008313 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 0.7740768829999997, "min": 0.7703047389999997, "max": 0.7919749250000002 }, "owned_device_tensor_peak_bytes": { "median": 2264616, "min": 2264616, "max": 2264616 }, "readback_seconds": { "median": 0.14073625700000006, "min": 0.1406611379999999, "max": 0.14422723300000018 }, "backend": "cuda_cublas", "target": "linux-x86_64" }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 68.02293487299998, "min": 67.35432496700007, "max": 71.42854133200002 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 68.97260412699998, "min": 68.41272091100004, "max": 72.48310226199992 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 655360, "min": 655360, "max": 655360 }, "request_write_seconds": { "median": 1.0544526330000004, "min": 0.9496139069999998, "max": 1.0583404930000002 }, "request_payload_bytes": { "median": 61870640, "min": 61870640, "max": 61870640 }, "process_peak_rss_bytes": { "median": 286314496, "min": 286232576, "max": 286322688 }, "samples": { "median": 152920, "min": 152920, "max": 152920 }, "upload_preparation_seconds": { "median": 0.9694459050000002, "min": 0.9608044390000001, "max": 0.9706038950000002 }, "worker_total_seconds": { "median": 1.6511017689999994, "min": 1.643087887, "max": 1.6777562749999992 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 0.442767139, "min": 0.4391299450000002, "max": 0.46720008700000026 }, "owned_device_tensor_peak_bytes": { "median": 2264616, "min": 2264616, "max": 2264616 }, "readback_seconds": { "median": 0.07239806899999998, "min": 0.07221494900000006, "max": 0.07297718599999994 }, "backend": "cuda_cublas", "target": "linux-x86_64" } ], "fastest_selected_baseline": { "fit_seconds": { "case": "atlas01-gpu32", "seconds": 2.804163125, "speedup": 0.032762391072286914 }, "outer_command_wall_seconds": { "case": "atlas01-gpu32", "seconds": 3.2152430829883087, "speedup": 0.03734582930414821 } } }, { "case": "02-03-gpu64", "label": "Atlas-02 CUDA + Atlas-03 CUDA \u00b7 Float64", "dtype": "Float64", "kind": "group", "replicas": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1 }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-01/C03", "numerical_reference_case": "atlas01-cpu64", "repetitions": 3, "outer_command_wall_seconds": { "median": 124.655269416995, "min": 116.99921195799834, "max": 126.23090900000534 }, "fit_seconds": { "median": 124.125102542, "min": 116.467892458, "max": 125.694160875 }, "command_seconds": { "median": 124.584389833, "min": 116.93126075, "max": 126.164727625 }, "quality": { "initial_loss": { "median": 1.5991332128378295, "min": 1.5991332128378295, "max": 1.5991332128378295 }, "training_loss": { "median": 0.2923469152365444, "min": 0.2923469152365444, "max": 0.2923469152365444 }, "validation_loss": { "median": 0.3033731649951008, "min": 0.3033731649951008, "max": 0.3033731649951008 }, "test_loss": { "median": 0.29705317871020637, "min": 0.29705317871020637, "max": 0.29705317871020637 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 3.3306690738754696e-16, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 5.551115123125783e-17, "validation_loss": 5.551115123125783e-17, "test_loss": 5.551115123125783e-17 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 76169216, "min": 75055104, "max": 76529664 }, "tensor_owned_host_peak_bytes": { "median": 4202560, "min": 4202560, "max": 4202560 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 108.28300708299992, "min": 104.58419616099997, "max": 111.21368550599998 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 110.89321180400003, "min": 107.19696121899996, "max": 113.86994150799987 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 393216, "min": 393216, "max": 393216 }, "request_write_seconds": { "median": 2.612709755, "min": 2.610150838999998, "max": 2.656146796 }, "request_payload_bytes": { "median": 165262320, "min": 165262320, "max": 165262320 }, "process_peak_rss_bytes": { "median": 254693376, "min": 254693376, "max": 254758912 }, "samples": { "median": 229380, "min": 229380, "max": 229380 }, "upload_preparation_seconds": { "median": 1.1247009410000004, "min": 1.1209860590000003, "max": 1.132229757 }, "worker_total_seconds": { "median": 3.444995925000001, "min": 3.4443716670000013, "max": 3.4631106159999994 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 1.7728895749999996, "min": 1.7582898530000002, "max": 1.773451806000001 }, "owned_device_tensor_peak_bytes": { "median": 6659152, "min": 6659152, "max": 6659152 }, "readback_seconds": { "median": 0.167989555, "min": 0.16438954099999997, "max": 0.17331558899999996 }, "backend": "cuda_cublas", "target": "linux-x86_64" }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 96.94922095000004, "min": 94.17421533999999, "max": 107.93112454199999 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 99.56707587099999, "min": 96.842004244, "max": 110.54037168700005 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 393216, "min": 393216, "max": 393216 }, "request_write_seconds": { "median": 2.6178147880000004, "min": 2.609176034999999, "max": 2.6677069549999985 }, "request_payload_bytes": { "median": 165262320, "min": 165262320, "max": 165262320 }, "process_peak_rss_bytes": { "median": 282042368, "min": 281931776, "max": 282214400 }, "samples": { "median": 229380, "min": 229380, "max": 229380 }, "upload_preparation_seconds": { "median": 0.45382250100000004, "min": 0.45233597800000014, "max": 0.45668940900000027 }, "worker_total_seconds": { "median": 1.1111053540000004, "min": 1.1019800560000002, "max": 1.1185561859999993 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 0.434931702, "min": 0.43328174500000005, "max": 0.4439623549999999 }, "owned_device_tensor_peak_bytes": { "median": 6659152, "min": 6659152, "max": 6659152 }, "readback_seconds": { "median": 0.062440942000000006, "min": 0.06230874499999999, "max": 0.06328919300000004 }, "backend": "cuda_cublas", "target": "linux-x86_64" } ], "fastest_selected_baseline": { "fit_seconds": { "case": "atlas03-gpu64", "seconds": 3.380031545, "speedup": 0.027230845943158877 }, "outer_command_wall_seconds": { "case": "atlas03-gpu64", "seconds": 5.359644708005362, "speedup": 0.042995733217473193 } } }, { "case": "01-02-03-cpu32", "label": "Atlas-01 CPU + Atlas-02 CPU + Atlas-03 CPU \u00b7 Float32", "dtype": "Float32", "kind": "group", "replicas": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "cpu", "relative_batch_weight": 1 }, { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cpu", "relative_batch_weight": 1 }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cpu", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-01/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 89.16344270898844, "min": 88.21805833300459, "max": 92.74185400000715 }, "fit_seconds": { "median": 88.645885167, "min": 87.638751083, "max": 92.231870458 }, "command_seconds": { "median": 89.109699709, "min": 88.129241583, "max": 92.696210667 }, "quality": { "initial_loss": { "median": 1.5991332174875397, "min": 1.5991332174875397, "max": 1.5991332174875397 }, "training_loss": { "median": 0.2923475254698321, "min": 0.2923475254698321, "max": 0.2923475254698321 }, "validation_loss": { "median": 0.30337432724777097, "min": 0.30337432724777097, "max": 0.30337432724777097 }, "test_loss": { "median": 0.2970538696340481, "min": 0.2970538696340481, "max": 0.2970538696340481 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8858596134282808, "min": 0.8858596134282808, "max": 0.8858596134282808 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 4.561617970466614e-06, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 1.0843586306652497e-08, "validation_loss": 9.871450834531004e-08, "test_loss": 5.9598581236830483e-08 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 77873152, "min": 76414976, "max": 77889536 }, "tensor_owned_host_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 3.0818655430000006, "min": 2.944582882999999, "max": 3.093941788 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 3.9981469220000023, "min": 3.811123540000002, "max": 4.0202211650000015 }, "inner_threads": null, "owned_host_tensor_peak_bytes": { "median": 6834532, "min": 6834532, "max": 6834532 }, "request_write_seconds": { "median": 0.9162247140000003, "min": 0.8664736270000004, "max": 0.9262225830000002 }, "request_payload_bytes": { "median": 61870640, "min": 61870640, "max": 61870640 }, "process_peak_rss_bytes": { "median": 24690688, "min": 24117248, "max": 25133056 }, "samples": { "median": 152920, "min": 152920, "max": 152920 }, "upload_preparation_seconds": { "median": 0.15840661500000008, "min": 0.15058891999999988, "max": 0.15966467000000004 }, "worker_total_seconds": { "median": 1.2889534940000007, "min": 1.2409844569999993, "max": 1.2952419939999993 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 1.0918180099999997, "min": 1.0529558259999998, "max": 1.0963840000000002 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.018997038999999993, "min": 0.018368426, "max": 0.019318249000000006 }, "backend": "apple_accelerate", "target": "macos-arm64" }, { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 77.784782799, "min": 77.513491078, "max": 79.15330272000001 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 78.81341042199998, "min": 78.54798208700002, "max": 80.12637662399999 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 6834532, "min": 6834532, "max": 6834532 }, "request_write_seconds": { "median": 1.028517587, "min": 0.9729863700000007, "max": 1.0344487010000003 }, "request_payload_bytes": { "median": 61870640, "min": 61870640, "max": 61870640 }, "process_peak_rss_bytes": { "median": 124194816, "min": 124116992, "max": 124428288 }, "samples": { "median": 152920, "min": 152920, "max": 152920 }, "upload_preparation_seconds": { "median": 1.348304754, "min": 1.3208315059999995, "max": 1.3649168740000002 }, "worker_total_seconds": { "median": 13.892850112000001, "min": 13.59844539500001, "max": 13.917500425000002 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 11.949936006, "min": 11.702780269000003, "max": 12.001523366999999 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.146685316, "min": 0.14600822400000005, "max": 0.14739254699999993 }, "backend": "openblas", "target": "linux-x86_64" }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 73.57592612399998, "min": 65.71419562500003, "max": 74.19305296100005 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 74.60627787600002, "min": 66.752902796, "max": 75.160350051 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 6834532, "min": 6834532, "max": 6834532 }, "request_write_seconds": { "median": 1.0302939589999998, "min": 0.9672401329999998, "max": 1.0386550049999992 }, "request_payload_bytes": { "median": 61870640, "min": 61870640, "max": 61870640 }, "process_peak_rss_bytes": { "median": 124391424, "min": 124260352, "max": 124420096 }, "samples": { "median": 152920, "min": 152920, "max": 152920 }, "upload_preparation_seconds": { "median": 0.9685700180000002, "min": 0.9642789750000003, "max": 0.9753726200000005 }, "worker_total_seconds": { "median": 5.551128372999997, "min": 5.463456344000003, "max": 5.588757929000004 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 4.399906118000004, "min": 4.324123430000001, "max": 4.426885362000004 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.036184635999999985, "min": 0.03586423799999997, "max": 0.03694794100000001 }, "backend": "openblas", "target": "linux-x86_64" } ], "fastest_selected_baseline": { "fit_seconds": { "case": "atlas01-cpu32", "seconds": 2.721057125, "speedup": 0.03069580860830483 }, "outer_command_wall_seconds": { "case": "atlas01-cpu32", "seconds": 3.0628960830072174, "speedup": 0.03435147847536456 } } }, { "case": "01-02-03-cpu64", "label": "Atlas-01 CPU + Atlas-02 CPU + Atlas-03 CPU \u00b7 Float64", "dtype": "Float64", "kind": "group", "replicas": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "cpu", "relative_batch_weight": 1 }, { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cpu", "relative_batch_weight": 1 }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cpu", "relative_batch_weight": 1 } ], "coordinator_configuration_id": "atlas-01/C03", "numerical_reference_case": "atlas01-cpu64", "repetitions": 3, "outer_command_wall_seconds": { "median": 114.44000404200051, "min": 108.63698216599005, "max": 117.02706866699737 }, "fit_seconds": { "median": 113.89337375, "min": 108.103035375, "max": 116.4694985 }, "command_seconds": { "median": 114.357245166, "min": 108.564510833, "max": 116.938115166 }, "quality": { "initial_loss": { "median": 1.5991332128378295, "min": 1.5991332128378295, "max": 1.5991332128378295 }, "training_loss": { "median": 0.29234691523654444, "min": 0.29234691523654444, "max": 0.29234691523654444 }, "validation_loss": { "median": 0.3033731649951008, "min": 0.3033731649951008, "max": 0.3033731649951008 }, "test_loss": { "median": 0.2970531787102063, "min": 0.2970531787102063, "max": 0.2970531787102063 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 2.7755575615628914e-16, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 0.0, "validation_loss": 5.551115123125783e-17, "test_loss": 1.1102230246251565e-16 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.0, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 71483392, "min": 70926336, "max": 73039872 }, "tensor_owned_host_peak_bytes": { "median": 4202560, "min": 4202560, "max": 4202560 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-01/C03", "platform": "atlas-01", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 3.839887451999999, "min": 3.7663322510000006, "max": 3.913244859999998 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 5.567488794999997, "min": 5.434352178, "max": 5.695344581000001 }, "inner_threads": null, "owned_host_tensor_peak_bytes": { "median": 5398608, "min": 5398608, "max": 5398608 }, "request_write_seconds": { "median": 1.7275411719999993, "min": 1.6679660429999998, "max": 1.782044620999998 }, "request_payload_bytes": { "median": 123668080, "min": 123668080, "max": 123668080 }, "process_peak_rss_bytes": { "median": 20987904, "min": 20840448, "max": 25673728 }, "samples": { "median": 152920, "min": 152920, "max": 152920 }, "upload_preparation_seconds": { "median": 0.10331621500000006, "min": 0.10130259600000005, "max": 0.10499419699999991 }, "worker_total_seconds": { "median": 1.4972432159999998, "min": 1.477964198000001, "max": 1.526615741000001 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 1.3599068440000004, "min": 1.3427640759999984, "max": 1.3873297540000007 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.015336034999999996, "min": 0.015248027000000003, "max": 0.015355873999999993 }, "backend": "apple_accelerate", "target": "macos-arm64" }, { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 104.35145782700003, "min": 93.57786412000006, "max": 104.37799353499994 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 106.41689174800003, "min": 95.71556680100005, "max": 106.56334433700003 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 5398608, "min": 5398608, "max": 5398608 }, "request_write_seconds": { "median": 2.1376533849999997, "min": 2.0653792869999994, "max": 2.1852947489999988 }, "request_payload_bytes": { "median": 123668080, "min": 123668080, "max": 123668080 }, "process_peak_rss_bytes": { "median": 122904576, "min": 122826752, "max": 123088896 }, "samples": { "median": 152920, "min": 152920, "max": 152920 }, "upload_preparation_seconds": { "median": 0.8999936510000001, "min": 0.8924193459999997, "max": 0.9124430939999998 }, "worker_total_seconds": { "median": 13.327921916999996, "min": 13.227574789999995, "max": 13.369700921999998 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 11.927826143999997, "min": 11.839041184000004, "max": 11.957221955999998 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.16115268799999996, "min": 0.16050355099999997, "max": 0.16321454900000001 }, "backend": "openblas", "target": "linux-x86_64" }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cpu", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 92.89171421000005, "min": 90.68079633100007, "max": 94.24916463099996 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 95.08160861699996, "min": 92.81445225600002, "max": 96.32528441800005 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 5398608, "min": 5398608, "max": 5398608 }, "request_write_seconds": { "median": 2.1335961649999993, "min": 2.0760746739999996, "max": 2.1897552950000008 }, "request_payload_bytes": { "median": 123668080, "min": 123668080, "max": 123668080 }, "process_peak_rss_bytes": { "median": 122892288, "min": 122687488, "max": 123039744 }, "samples": { "median": 152920, "min": 152920, "max": 152920 }, "upload_preparation_seconds": { "median": 0.4658746279999999, "min": 0.45744333199999965, "max": 0.4670489530000005 }, "worker_total_seconds": { "median": 4.483889233999999, "min": 4.476447578000001, "max": 4.534127982999999 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 3.8670035659999997, "min": 3.8578807020000028, "max": 3.906021795999999 }, "owned_device_tensor_peak_bytes": { "median": 0, "min": 0, "max": 0 }, "readback_seconds": { "median": 0.038350755000000035, "min": 0.03728029299999998, "max": 0.039353534999999995 }, "backend": "openblas", "target": "linux-x86_64" } ], "fastest_selected_baseline": { "fit_seconds": { "case": "atlas01-cpu64", "seconds": 3.086051291, "speedup": 0.02709596870643232 }, "outer_command_wall_seconds": { "case": "atlas01-cpu64", "seconds": 3.4516792920039734, "speedup": 0.0301614747473897 } } }, { "case": "02-03-gpu32-w1-7", "label": "Atlas-02 CUDA + Atlas-03 CUDA \u00b7 Float32 \u00b7 allocation 1:7", "dtype": "Float32", "kind": "group", "replicas": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1 }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 7 } ], "coordinator_configuration_id": "atlas-01/C03", "numerical_reference_case": "atlas01-cpu32", "repetitions": 3, "outer_command_wall_seconds": { "median": 92.13198762499087, "min": 90.50230108300457, "max": 93.24883054199745 }, "fit_seconds": { "median": 91.641994208, "min": 90.005007041, "max": 92.727323583 }, "command_seconds": { "median": 92.093433917, "min": 90.461175125, "max": 93.18005375 }, "quality": { "initial_loss": { "median": 1.5991332174875397, "min": 1.5991332174875397, "max": 1.5991332174875397 }, "training_loss": { "median": 0.29234655017850286, "min": 0.29234655017850286, "max": 0.29234655017850286 }, "validation_loss": { "median": 0.30337328036464484, "min": 0.30337328036464484, "max": 0.30337328036464484 }, "test_loss": { "median": 0.29705295057650866, "min": 0.29705295057650866, "max": 0.29705295057650866 }, "training_accuracy": { "median": 0.8983564390966955, "min": 0.8983564390966955, "max": 0.8983564390966955 }, "validation_accuracy": { "median": 0.8857578840284842, "min": 0.8857578840284842, "max": 0.8857578840284842 }, "test_accuracy": { "median": 0.8955239064089522, "min": 0.8955239064089522, "max": 0.8955239064089522 } }, "parameter_max_abs_error": 4.634261131286621e-05, "loss_max_abs_errors": { "initial_loss": 0.0, "training_loss": 9.861349155593224e-07, "validation_loss": 1.1455976344776353e-06, "test_loss": 9.786561206981403e-07 }, "accuracy_max_abs_errors": { "training_accuracy": 0.0, "validation_accuracy": 0.00010172939979657958, "test_accuracy": 0.0 }, "coordinator_memory": { "process_peak_rss_bytes": { "median": 77479936, "min": 75546624, "max": 77643776 }, "tensor_owned_host_peak_bytes": { "median": 6303328, "min": 6303328, "max": 6303328 }, "tensor_owned_device_peak_bytes": { "median": 0, "min": 0, "max": 0 } }, "replica_metrics": [ { "configuration_id": "atlas-02/C03", "platform": "atlas-02", "device": "cuda", "relative_batch_weight": 1, "response_read_wait_seconds": { "median": 49.793614289000026, "min": 47.13218325199995, "max": 50.69311366600001 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 50.300012469000016, "min": 47.69015449500003, "max": 51.246885000999995 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 245760, "min": 245760, "max": 245760 }, "request_write_seconds": { "median": 0.5537185480000001, "min": 0.5063517589999998, "max": 0.5579191199999997 }, "request_payload_bytes": { "median": 35875600, "min": 35875600, "max": 35875600 }, "process_peak_rss_bytes": { "median": 251711488, "min": 251682816, "max": 251740160 }, "samples": { "median": 57350, "min": 57350, "max": 57350 }, "upload_preparation_seconds": { "median": 0.8483147890000005, "min": 0.8400615010000003, "max": 0.8638853949999997 }, "worker_total_seconds": { "median": 1.7833610479999993, "min": 1.7712620580000007, "max": 1.8202823890000008 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 0.6286405510000002, "min": 0.6278459390000003, "max": 0.648182167 }, "owned_device_tensor_peak_bytes": { "median": 933416, "min": 933416, "max": 933416 }, "readback_seconds": { "median": 0.13438590700000008, "min": 0.13205223499999996, "max": 0.13533233299999997 }, "backend": "cuda_cublas", "target": "linux-x86_64" }, { "configuration_id": "atlas-03/C03", "platform": "atlas-03", "device": "cuda", "relative_batch_weight": 7, "response_read_wait_seconds": { "median": 82.62908840499999, "min": 80.11868316900004, "max": 82.66911189099999 }, "requests": { "median": 300, "min": 300, "max": 300 }, "coordinator_exchange_wall_seconds": { "median": 84.79015824300004, "min": 82.07265038100005, "max": 84.83596416000006 }, "inner_threads": { "median": 1, "min": 1, "max": 1 }, "owned_host_tensor_peak_bytes": { "median": 1720320, "min": 1720320, "max": 1720320 }, "request_write_seconds": { "median": 2.1609882789999997, "min": 1.9538403350000004, "max": 2.1667847920000005 }, "request_payload_bytes": { "median": 129459920, "min": 129459920, "max": 129459920 }, "process_peak_rss_bytes": { "median": 288477184, "min": 288174080, "max": 288571392 }, "samples": { "median": 401410, "min": 401410, "max": 401410 }, "upload_preparation_seconds": { "median": 1.1265795959999994, "min": 1.119587932999999, "max": 1.1299737530000002 }, "worker_total_seconds": { "median": 1.6740156149999985, "min": 1.6696630329999997, "max": 1.6912474249999987 }, "zero_sample_steps": { "median": 0, "min": 0, "max": 0 }, "response_payload_bytes": { "median": 40441800, "min": 40441800, "max": 40441800 }, "compute_seconds": { "median": 0.35501172800000025, "min": 0.35420495699999993, "max": 0.3653267259999998 }, "owned_device_tensor_peak_bytes": { "median": 5725736, "min": 5725736, "max": 5725736 }, "readback_seconds": { "median": 0.06263750900000004, "min": 0.062500059, "max": 0.06264641600000001 }, "backend": "cuda_cublas", "target": "linux-x86_64" } ], "fastest_selected_baseline": { "fit_seconds": { "case": "atlas03-gpu32", "seconds": 4.041986648, "speedup": 0.04410627117984683 }, "outer_command_wall_seconds": { "case": "atlas03-gpu32", "seconds": 6.113490791001823, "speedup": 0.06635578965131904 } } } ] }