[
  {
    "50.00 percentile latency (ns)": 61512754168,
    "90.00 percentile latency (ns)": 135167010149,
    "95.00 percentile latency (ns)": 146253866173,
    "97.00 percentile latency (ns)": 151081129728,
    "99.00 percentile latency (ns)": 156408465632,
    "99.90 percentile latency (ns)": 159500460274,
    "Max latency (ns)": 160241082999,
    "Mean latency (ns)": 68366915145,
    "Min duration satisfied": "Yes",
    "Min latency (ns)": 5643768565,
    "Min queries satisfied": "Yes",
    "Mode": "Performance",
    "Result is": "VALID",
    "SUT name": "PySUT",
    "Samples per second": 11981.9,
    "Scenario": "offline",
    "accelerator_memory_capacity": "16 GB per accelerator",
    "accelerator_memory_configuration": "HBM",
    "accelerator_model_name": "tpu-v3.8",
    "accelerators_per_node": 32,
    "accuracy_log_probability": 0,
    "accuracy_log_rng_seed": 0,
    "characteristics.blue": 24.0,
    "characteristics.samples_per_second": 11981.9,
    "characteristics.samples_per_second.normalized_per_core": 374.434375,
    "characteristics.samples_per_second.normalized_per_processor": 374.434375,
    "ck_system": "tpu-v3.8x4",
    "ck_used": true,
    "dataset": "WMT E-G",
    "dataset_link": "",
    "dim_x_default": "characteristics.samples_per_second",
    "dim_x_maximize": true,
    "dim_y_default": "characteristics.blue",
    "dim_y_maximize": true,
    "division": "closed",
    "formal_model": "nmt",
    "formal_model_accuracy": 99.0,
    "formal_model_link": "",
    "framework": "tf-nightly==1.15.0-dev20190730",
    "host_memory_capacity": "360 GB",
    "host_processor_core_count": 120,
    "host_processor_model_name": "AMD EPYC 7V13 64-Core Processor",
    "host_processors_per_node": 2,
    "host_storage_capacity": "200 GB + 50 GB",
    "host_storage_type": "SSD",
    "hw_notes": "'Number_of_nodes' is the number of user programmable nodes. Each GCP TPU v3-8 has a non-user programmable 'tensorflow host' node which provides direct network and storage access to the TPU.",
    "informal_model": "gnmt",
    "input_data_types": "bfloat16",
    "key.accuracy": "characteristics.blue",
    "max_async_queries": 1,
    "max_duration (ms)": 0,
    "max_query_count": 0,
    "min_duration (ms)": 60000,
    "min_query_count": 1,
    "mlperf_version": 0.5,
    "normalize_cores": 32,
    "normalize_processors": 32,
    "note_code": "https://github.com/mlcommons/inference_results_v0.5/tree/master/closed/Google/code",
    "note_details": "https://github.com/mlcommons/inference_results_v0.5/tree/master/closed/Google/results/tpu-v3.8x4",
    "number_of_nodes": 1,
    "operating_system": "Ubuntu 18.04.5 LTS (Linux-5.4.0-1055-azure-x86_64-with-Ubuntu-18.04-bionic)",
    "other_software_stack": "TPU version 1.15.dev20190821; GCC 7.5.0; Python 3.7.10",
    "performance_issue_same": true,
    "performance_issue_same_index": 0,
    "performance_issue_unique": true,
    "performance_sample_count": 3903900,
    "print_timestamps": true,
    "problem": false,
    "qsl_rng_seed": 3133965575612453542,
    "retraining": "None",
    "sample_index_rng_seed": 665484352860916858,
    "samples_per_query": 1920000,
    "schedule_rng_seed": 3622009729038561421,
    "starting_weights_filename": "https://zenodo.org/record/2530924/files/gnmt_model.zip",
    "status": "available",
    "submitter": "Google",
    "submitter_link": "https://github.com/ctuning/ck-mlperf-inference/tree/main/bench.mlperf.submitter/Google",
    "sw_notes": "Powered by CK v2.5.8 (https://github.com/ctuning/ck)",
    "system_link": "https://github.com/ctuning/ck-mlperf-inference/tree/main/bench.mlperf.system/tpu-v3.8x4",
    "system_name": "Microsoft Corporation 7.0 (Virtual Machine)",
    "target_latency (ns)": 0,
    "target_qps": 16000,
    "task": "NMT",
    "task2": "nmt",
    "total_cores": 240,
    "uid": "a31e3fe44ad1a628",
    "use_accelerator": true,
    "weight_data_types": "bfloat16",
    "weight_transformations": "Transpose the final projection layer"
  },
  {
    "50.00 percentile latency (ns)": 71923697423,
    "90.00 percentile latency (ns)": 145546184401,
    "95.00 percentile latency (ns)": 156686945263,
    "97.00 percentile latency (ns)": 161423098990,
    "99.00 percentile latency (ns)": 166792421339,
    "99.90 percentile latency (ns)": 170045591808,
    "Max latency (ns)": 170597895701,
    "Mean latency (ns)": 78611652216,
    "Min duration satisfied": "Yes",
    "Min latency (ns)": 12074152385,
    "Min queries satisfied": "Yes",
    "Mode": "Performance",
    "Result is": "VALID",
    "SUT name": "PySUT",
    "Samples per second": 22509.1,
    "Scenario": "offline",
    "accelerator_memory_capacity": "16 GB per accelerator",
    "accelerator_memory_configuration": "HBM",
    "accelerator_model_name": "tpu-v3.8",
    "accelerators_per_node": 64,
    "accuracy_log_probability": 0,
    "accuracy_log_rng_seed": 0,
    "characteristics.blue": 24.0,
    "characteristics.samples_per_second": 22509.1,
    "characteristics.samples_per_second.normalized_per_core": 351.7046875,
    "characteristics.samples_per_second.normalized_per_processor": 351.7046875,
    "ck_system": "tpu-v3.8x8",
    "ck_used": true,
    "dataset": "WMT E-G",
    "dataset_link": "",
    "dim_x_default": "characteristics.samples_per_second",
    "dim_x_maximize": true,
    "dim_y_default": "characteristics.blue",
    "dim_y_maximize": true,
    "division": "closed",
    "formal_model": "nmt",
    "formal_model_accuracy": 99.0,
    "formal_model_link": "",
    "framework": "tf-nightly==1.15.0-dev20190730",
    "host_memory_capacity": "360 GB",
    "host_processor_core_count": 120,
    "host_processor_model_name": "AMD EPYC 7V13 64-Core Processor",
    "host_processors_per_node": 2,
    "host_storage_capacity": "200 GB + 50 GB",
    "host_storage_type": "SSD",
    "hw_notes": "'Number_of_nodes' is the number of user programmable nodes. Each GCP TPU v3-8 has a non-user programmable 'tensorflow host' node which provides direct network and storage access to the TPU.",
    "informal_model": "gnmt",
    "input_data_types": "bfloat16",
    "key.accuracy": "characteristics.blue",
    "max_async_queries": 1,
    "max_duration (ms)": 0,
    "max_query_count": 0,
    "min_duration (ms)": 60000,
    "min_query_count": 1,
    "mlperf_version": 0.5,
    "normalize_cores": 64,
    "normalize_processors": 64,
    "note_code": "https://github.com/mlcommons/inference_results_v0.5/tree/master/closed/Google/code",
    "note_details": "https://github.com/mlcommons/inference_results_v0.5/tree/master/closed/Google/results/tpu-v3.8x8",
    "number_of_nodes": 1,
    "operating_system": "Ubuntu 18.04.5 LTS (Linux-5.4.0-1055-azure-x86_64-with-Ubuntu-18.04-bionic)",
    "other_software_stack": "TPU version 1.15.dev20190821; GCC 7.5.0; Python 3.7.10",
    "performance_issue_same": true,
    "performance_issue_same_index": 0,
    "performance_issue_unique": true,
    "performance_sample_count": 3903900,
    "print_timestamps": true,
    "problem": false,
    "qsl_rng_seed": 3133965575612453542,
    "retraining": "None",
    "sample_index_rng_seed": 665484352860916858,
    "samples_per_query": 3840000,
    "schedule_rng_seed": 3622009729038561421,
    "starting_weights_filename": "https://zenodo.org/record/2530924/files/gnmt_model.zip",
    "status": "available",
    "submitter": "Google",
    "submitter_link": "https://github.com/ctuning/ck-mlperf-inference/tree/main/bench.mlperf.submitter/Google",
    "sw_notes": "Powered by CK v2.5.8 (https://github.com/ctuning/ck)",
    "system_link": "https://github.com/ctuning/ck-mlperf-inference/tree/main/bench.mlperf.system/tpu-v3.8x8",
    "system_name": "Microsoft Corporation 7.0 (Virtual Machine)",
    "target_latency (ns)": 0,
    "target_qps": 32000,
    "task": "NMT",
    "task2": "nmt",
    "total_cores": 240,
    "uid": "148a9b85b65c324c",
    "use_accelerator": true,
    "weight_data_types": "bfloat16",
    "weight_transformations": "Transpose the final projection layer"
  },
  {
    "50.00 percentile latency (ns)": 58012269749,
    "90.00 percentile latency (ns)": 131676014330,
    "95.00 percentile latency (ns)": 142839885020,
    "97.00 percentile latency (ns)": 147490890922,
    "99.00 percentile latency (ns)": 152947601439,
    "99.90 percentile latency (ns)": 156069300441,
    "Max latency (ns)": 156665642104,
    "Mean latency (ns)": 64863482958,
    "Min duration satisfied": "Yes",
    "Min latency (ns)": 2775534622,
    "Min queries satisfied": "Yes",
    "Mode": "Performance",
    "Result is": "VALID",
    "SUT name": "PySUT",
    "Samples per second": 6127.7,
    "Scenario": "offline",
    "accelerator_memory_capacity": "16 GB per accelerator",
    "accelerator_memory_configuration": "HBM",
    "accelerator_model_name": "tpu-v3.8",
    "accelerators_per_node": 16,
    "accuracy_log_probability": 0,
    "accuracy_log_rng_seed": 0,
    "characteristics.blue": 24.0,
    "characteristics.samples_per_second": 6127.7,
    "characteristics.samples_per_second.normalized_per_core": 382.98125,
    "characteristics.samples_per_second.normalized_per_processor": 382.98125,
    "ck_system": "tpu-v3.8x2",
    "ck_used": true,
    "dataset": "WMT E-G",
    "dataset_link": "",
    "dim_x_default": "characteristics.samples_per_second",
    "dim_x_maximize": true,
    "dim_y_default": "characteristics.blue",
    "dim_y_maximize": true,
    "division": "closed",
    "formal_model": "nmt",
    "formal_model_accuracy": 99.0,
    "formal_model_link": "",
    "framework": "tf-nightly==1.15.0-dev20190730",
    "host_memory_capacity": "360 GB",
    "host_processor_core_count": 120,
    "host_processor_model_name": "AMD EPYC 7V13 64-Core Processor",
    "host_processors_per_node": 2,
    "host_storage_capacity": "200 GB + 50 GB",
    "host_storage_type": "SSD",
    "hw_notes": "'Number_of_nodes' is the number of user programmable nodes. Each GCP TPU v3-8 has a non-user programmable 'tensorflow host' node which provides direct network and storage access to the TPU.",
    "informal_model": "gnmt",
    "input_data_types": "bfloat16",
    "key.accuracy": "characteristics.blue",
    "max_async_queries": 1,
    "max_duration (ms)": 0,
    "max_query_count": 0,
    "min_duration (ms)": 60000,
    "min_query_count": 1,
    "mlperf_version": 0.5,
    "normalize_cores": 16,
    "normalize_processors": 16,
    "note_code": "https://github.com/mlcommons/inference_results_v0.5/tree/master/closed/Google/code",
    "note_details": "https://github.com/mlcommons/inference_results_v0.5/tree/master/closed/Google/results/tpu-v3.8x2",
    "number_of_nodes": 1,
    "operating_system": "Ubuntu 18.04.5 LTS (Linux-5.4.0-1055-azure-x86_64-with-Ubuntu-18.04-bionic)",
    "other_software_stack": "TPU version 1.15.dev20190821; GCC 7.5.0; Python 3.7.10",
    "performance_issue_same": true,
    "performance_issue_same_index": 0,
    "performance_issue_unique": true,
    "performance_sample_count": 3903900,
    "print_timestamps": true,
    "problem": false,
    "qsl_rng_seed": 3133965575612453542,
    "retraining": "None",
    "sample_index_rng_seed": 665484352860916858,
    "samples_per_query": 960000,
    "schedule_rng_seed": 3622009729038561421,
    "starting_weights_filename": "https://zenodo.org/record/2530924/files/gnmt_model.zip",
    "status": "available",
    "submitter": "Google",
    "submitter_link": "https://github.com/ctuning/ck-mlperf-inference/tree/main/bench.mlperf.submitter/Google",
    "sw_notes": "Powered by CK v2.5.8 (https://github.com/ctuning/ck)",
    "system_link": "https://github.com/ctuning/ck-mlperf-inference/tree/main/bench.mlperf.system/tpu-v3.8x2",
    "system_name": "Microsoft Corporation 7.0 (Virtual Machine)",
    "target_latency (ns)": 0,
    "target_qps": 8000,
    "task": "NMT",
    "task2": "nmt",
    "total_cores": 240,
    "uid": "9cb615966b9a4850",
    "use_accelerator": true,
    "weight_data_types": "bfloat16",
    "weight_transformations": "Transpose the final projection layer"
  },
  {
    "50.00 percentile latency (ns)": 137125919118,
    "90.00 percentile latency (ns)": 235707472344,
    "95.00 percentile latency (ns)": 248988834548,
    "97.00 percentile latency (ns)": 254602785367,
    "99.00 percentile latency (ns)": 260477290262,
    "99.90 percentile latency (ns)": 263323450226,
    "Max latency (ns)": 264270496771,
    "Mean latency (ns)": 139523969851,
    "Min duration satisfied": "Yes",
    "Min latency (ns)": 25583739069,
    "Min queries satisfied": "Yes",
    "Mode": "Performance",
    "Result is": "VALID",
    "SUT name": "PySUT",
    "Samples per second": 29061.1,
    "Scenario": "offline",
    "accelerator_memory_capacity": "16 GB per accelerator",
    "accelerator_memory_configuration": "HBM",
    "accelerator_model_name": "tpu-v3.8",
    "accelerators_per_node": 128,
    "accuracy_log_probability": 0,
    "accuracy_log_rng_seed": 0,
    "characteristics.blue": 24.0,
    "characteristics.samples_per_second": 29061.1,
    "characteristics.samples_per_second.normalized_per_core": 227.03984375,
    "characteristics.samples_per_second.normalized_per_processor": 227.03984375,
    "ck_system": "tpu-v3.8x16",
    "ck_used": true,
    "dataset": "WMT E-G",
    "dataset_link": "",
    "dim_x_default": "characteristics.samples_per_second",
    "dim_x_maximize": true,
    "dim_y_default": "characteristics.blue",
    "dim_y_maximize": true,
    "division": "closed",
    "formal_model": "nmt",
    "formal_model_accuracy": 99.0,
    "formal_model_link": "",
    "framework": "tf-nightly==1.15.0-dev20190730",
    "host_memory_capacity": "360 GB",
    "host_processor_core_count": 120,
    "host_processor_model_name": "AMD EPYC 7V13 64-Core Processor",
    "host_processors_per_node": 2,
    "host_storage_capacity": "200 GB + 50 GB",
    "host_storage_type": "SSD",
    "hw_notes": "'Number_of_nodes' is the number of user programmable nodes. Each GCP TPU v3-8 has a non-user programmable 'tensorflow host' node which provides direct network and storage access to the TPU.",
    "informal_model": "gnmt",
    "input_data_types": "bfloat16",
    "key.accuracy": "characteristics.blue",
    "max_async_queries": 1,
    "max_duration (ms)": 0,
    "max_query_count": 0,
    "min_duration (ms)": 60000,
    "min_query_count": 1,
    "mlperf_version": 0.5,
    "normalize_cores": 128,
    "normalize_processors": 128,
    "note_code": "https://github.com/mlcommons/inference_results_v0.5/tree/master/closed/Google/code",
    "note_details": "https://github.com/mlcommons/inference_results_v0.5/tree/master/closed/Google/results/tpu-v3.8x16",
    "number_of_nodes": 1,
    "operating_system": "Ubuntu 18.04.5 LTS (Linux-5.4.0-1055-azure-x86_64-with-Ubuntu-18.04-bionic)",
    "other_software_stack": "TPU version 1.15.dev20190821; GCC 7.5.0; Python 3.7.10",
    "performance_issue_same": true,
    "performance_issue_same_index": 0,
    "performance_issue_unique": true,
    "performance_sample_count": 3903900,
    "print_timestamps": true,
    "problem": false,
    "qsl_rng_seed": 3133965575612453542,
    "retraining": "None",
    "sample_index_rng_seed": 665484352860916858,
    "samples_per_query": 7680000,
    "schedule_rng_seed": 3622009729038561421,
    "starting_weights_filename": "https://zenodo.org/record/2530924/files/gnmt_model.zip",
    "status": "available",
    "submitter": "Google",
    "submitter_link": "https://github.com/ctuning/ck-mlperf-inference/tree/main/bench.mlperf.submitter/Google",
    "sw_notes": "Powered by CK v2.5.8 (https://github.com/ctuning/ck)",
    "system_link": "https://github.com/ctuning/ck-mlperf-inference/tree/main/bench.mlperf.system/tpu-v3.8x16",
    "system_name": "Microsoft Corporation 7.0 (Virtual Machine)",
    "target_latency (ns)": 0,
    "target_qps": 64000,
    "task": "NMT",
    "task2": "nmt",
    "total_cores": 240,
    "uid": "8bb4d089a3c5bca3",
    "use_accelerator": true,
    "weight_data_types": "bfloat16",
    "weight_transformations": "Transpose the final projection layer"
  },
  {
    "50.00 percentile latency (ns)": 56743844607,
    "90.00 percentile latency (ns)": 130425725597,
    "95.00 percentile latency (ns)": 141574015327,
    "97.00 percentile latency (ns)": 146200569539,
    "99.00 percentile latency (ns)": 151728735052,
    "99.90 percentile latency (ns)": 154707893206,
    "Max latency (ns)": 155739115625,
    "Mean latency (ns)": 63568184121,
    "Min duration satisfied": "Yes",
    "Min latency (ns)": 1464472265,
    "Min queries satisfied": "Yes",
    "Mode": "Performance",
    "Result is": "VALID",
    "SUT name": "PySUT",
    "Samples per second": 3082.08,
    "Scenario": "offline",
    "accelerator_memory_capacity": "16 GB per accelerator",
    "accelerator_memory_configuration": "HBM",
    "accelerator_model_name": "tpu-v3.8",
    "accelerators_per_node": 8,
    "accuracy_log_probability": 0,
    "accuracy_log_rng_seed": 0,
    "characteristics.blue": 24.0,
    "characteristics.samples_per_second": 3082.08,
    "characteristics.samples_per_second.normalized_per_core": 385.26,
    "characteristics.samples_per_second.normalized_per_processor": 385.26,
    "ck_system": "tpu-v3.8x1",
    "ck_used": true,
    "dataset": "WMT E-G",
    "dataset_link": "",
    "dim_x_default": "characteristics.samples_per_second",
    "dim_x_maximize": true,
    "dim_y_default": "characteristics.blue",
    "dim_y_maximize": true,
    "division": "closed",
    "formal_model": "nmt",
    "formal_model_accuracy": 99.0,
    "formal_model_link": "",
    "framework": "tf-nightly==1.15.0-dev20190730",
    "host_memory_capacity": "360 GB",
    "host_processor_core_count": 120,
    "host_processor_model_name": "AMD EPYC 7V13 64-Core Processor",
    "host_processors_per_node": 2,
    "host_storage_capacity": "200 GB + 50 GB",
    "host_storage_type": "SSD",
    "hw_notes": "'Number_of_nodes' is the number of user programmable nodes. Each GCP TPU v3-8 has a non-user programmable 'tensorflow host' node which provides direct network and storage access to the TPU.",
    "informal_model": "gnmt",
    "input_data_types": "bfloat16",
    "key.accuracy": "characteristics.blue",
    "max_async_queries": 1,
    "max_duration (ms)": 0,
    "max_query_count": 0,
    "min_duration (ms)": 60000,
    "min_query_count": 1,
    "mlperf_version": 0.5,
    "normalize_cores": 8,
    "normalize_processors": 8,
    "note_code": "https://github.com/mlcommons/inference_results_v0.5/tree/master/closed/Google/code",
    "note_details": "https://github.com/mlcommons/inference_results_v0.5/tree/master/closed/Google/results/tpu-v3.8x1",
    "number_of_nodes": 1,
    "operating_system": "Ubuntu 18.04.5 LTS (Linux-5.4.0-1055-azure-x86_64-with-Ubuntu-18.04-bionic)",
    "other_software_stack": "TPU version 1.15.dev20190821; GCC 7.5.0; Python 3.7.10",
    "performance_issue_same": true,
    "performance_issue_same_index": 0,
    "performance_issue_unique": true,
    "performance_sample_count": 3903900,
    "print_timestamps": true,
    "problem": false,
    "qsl_rng_seed": 3133965575612453542,
    "retraining": "None",
    "sample_index_rng_seed": 665484352860916858,
    "samples_per_query": 480000,
    "schedule_rng_seed": 3622009729038561421,
    "starting_weights_filename": "https://zenodo.org/record/2530924/files/gnmt_model.zip",
    "status": "available",
    "submitter": "Google",
    "submitter_link": "https://github.com/ctuning/ck-mlperf-inference/tree/main/bench.mlperf.submitter/Google",
    "sw_notes": "Powered by CK v2.5.8 (https://github.com/ctuning/ck)",
    "system_link": "https://github.com/ctuning/ck-mlperf-inference/tree/main/bench.mlperf.system/tpu-v3.8x1",
    "system_name": "Microsoft Corporation 7.0 (Virtual Machine)",
    "target_latency (ns)": 0,
    "target_qps": 4000,
    "task": "NMT",
    "task2": "nmt",
    "total_cores": 240,
    "uid": "a0aab24e44a36243",
    "use_accelerator": true,
    "weight_data_types": "bfloat16",
    "weight_transformations": "Transpose the final projection layer"
  }
]