[ { "50.00 percentile latency (ns)": 106628969, "90.00 percentile latency (ns)": 171239758, "95.00 percentile latency (ns)": 201599005, "97.00 percentile latency (ns)": 235499533, "99.00 percentile latency (ns)": 347804998, "99.90 percentile latency (ns)": 1020102147, "Completed samples per second": 1503.45, "Max latency (ns)": 5190820937, "Mean latency (ns)": 118990524, "Min duration satisfied": "Yes", "Min latency (ns)": 29287110, "Min queries satisfied": "Yes", "Mode": "Performance", "Performance constraints satisfied": "Yes", "Result is": "VALID", "SUT name": "PySUT", "Scenario": "server", "Scheduled samples per second": 1506.76, "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.scheduled_queries_per_second": 1506.76, "characteristics.scheduled_queries_per_second.normalized_per_core": 188.345, "characteristics.scheduled_queries_per_second.normalized_per_processor": 188.345, "ck_system": "tpu-rdi-v3.8x1", "ck_used": true, "dataset": "WMT E-G", "dataset_link": "", "dim_x_default": "characteristics.scheduled_queries_per_second", "dim_y_default": "characteristics.blue", "dim_y_maximize": true, "division": "closed", "formal_model": "nmt", "formal_model_accuracy": 99.0, "formal_model_link": "", "framework": "TF at the internal CL listed in the code dir", "host_memory_capacity": "383 GB", "host_processor_core_count": 120, "host_processor_model_name": "AMD EPYC 7V13 64-Core Processor", "host_processors_per_node": 1, "host_storage_capacity": "21.42 GB", "host_storage_type": "HDD", "informal_model": "gnmt", "input_data_types": "bfloat16", "key.accuracy": "characteristics.blue", "max_async_queries": 0, "max_duration (ms)": 0, "max_query_count": 0, "min_duration (ms)": 60000, "min_query_count": 90000, "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-rdi-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": "TF; 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": 1, "schedule_rng_seed": 3622009729038561421, "starting_weights_filename": "https://zenodo.org/record/2530924/files/gnmt_model.zip", "status": "rdi", "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-rdi-v3.8x1", "system_name": "Microsoft Corporation 7.0 (Virtual Machine)", "target_latency (ns)": 250000000, "target_qps": 1500, "task": "NMT", "task2": "nmt", "total_cores": 120, "uid": "b40de47f483d5466", "use_accelerator": true, "weight_data_types": "bfloat16", "weight_transformations": "Transpose the final projection layer" }, { "50.00 percentile latency (ns)": 107574687, "90.00 percentile latency (ns)": 173976299, "95.00 percentile latency (ns)": 197135106, "97.00 percentile latency (ns)": 221696575, "99.00 percentile latency (ns)": 343692588, "99.90 percentile latency (ns)": 1172310680, "Completed samples per second": 1399.59, "Max latency (ns)": 5140964575, "Mean latency (ns)": 121145216, "Min duration satisfied": "Yes", "Min latency (ns)": 32187557, "Min queries satisfied": "Yes", "Mode": "Performance", "Performance constraints satisfied": "Yes", "Result is": "VALID", "SUT name": "PySUT", "Scenario": "server", "Scheduled samples per second": 1401.38, "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.scheduled_queries_per_second": 1401.38, "characteristics.scheduled_queries_per_second.normalized_per_core": 175.1725, "characteristics.scheduled_queries_per_second.normalized_per_processor": 175.1725, "ck_system": "tpu-v3.8x1", "ck_used": true, "dataset": "WMT E-G", "dataset_link": "", "dim_x_default": "characteristics.scheduled_queries_per_second", "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": 0, "max_duration (ms)": 0, "max_query_count": 0, "min_duration (ms)": 60000, "min_query_count": 168000, "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": 1, "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)": 250000000, "target_qps": 1400, "task": "NMT", "task2": "nmt", "total_cores": 240, "uid": "080e79be1cb4006b", "use_accelerator": true, "weight_data_types": "bfloat16", "weight_transformations": "Transpose the final projection layer" } ]