{ "comparison": { "left": "amd-mi355x", "production_qualified": false, "right": "nvidia-b300", "right_over_left_completed_request_rate": 1.075991431396243, "right_over_left_completed_token_rate": 1.0355815156646804, "scope": "these full submitted systems, software and benchmark quality/scenario only", "universal_hardware_ranking": false }, "local_gpu_execution": false, "observations": { "amd-mi355x": { "accuracy": "ROUGE1: 44.5266 ROUGE2: 22.1733 ROUGEL: 28.9289 TOKENS_PER_SAMPLE: 284.2", "availability": "available", "hardware": "AMD Instinct MI355X 288GB HBM3e", "inferred": false, "kind": "inference", "locally_measured": false, "measured_by": "AMD", "observed": { "completed_requests_per_second": 364.12, "completed_tokens_per_second": 103274.69, "tpot_constraint_ms": 200.0, "tpot_p99_ms": 149.535693, "ttft_constraint_ms": 2000.0, "ttft_p99_ms": 1175.800773 }, "query_unit_source": { "bytes": 268, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "amd-mi355x-query-size.txt", "local_sha256": "d7698b0c3d186825dfbefdd25ee0514f9ba03dd533baaad2674fbca1d02bd3d8", "path": "closed/AMD/results/8xMI355X_2xEPYC_9575F/llama2-70b-99.9/Server/performance/run_1/mlperf_log_detail.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/AMD/results/8xMI355X_2xEPYC_9575F/llama2-70b-99.9/Server/performance/run_1/mlperf_log_detail.txt", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "e5df1fadc362b6d15b8f548c74c9152199b154f6fbb16482fb23aedafc97cad5", "transformation": "selected exact generated_samples_per_query event line; trailing newline normalized" }, "replica_boundary": "entire submitted eight-accelerator system", "scope": { "Division": "closed", "MlperfModel": "llama2-70b-99.9", "Platform": "8xMI355X_2xEPYC_9575F", "Scenario": "Server", "framework": "PyTorch 2.10.0+git8514f05, ROCm 7.2.2", "total_accelerators": 8, "version": "v6.1", "weight_data_types": "fp4" }, "source": { "bytes": 3172, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "amd-mi355x-inference-summary.txt", "local_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "path": "closed/AMD/results/8xMI355X_2xEPYC_9575F/llama2-70b-99.9/Server/performance/run_1/mlperf_log_summary.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/AMD/results/8xMI355X_2xEPYC_9575F/llama2-70b-99.9/Server/performance/run_1/mlperf_log_summary.txt", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "transformation": "none" } }, "amd-mi355x-training": { "accelerator_count": 8, "benchmark": "llama2_70b_lora", "hardware_identity": { "hardware_specific_calibration_eligible": false, "limit": "A selected run is not hardware qualification; retain conflicting source labels without guessing a correction.", "logged_submission_platform": "MI355X", "status": "source_labels_agree", "system_declared_model": "AMD Instinct MI355X 288GB HBM3e" }, "kind": "training", "locally_measured": false, "observed": { "elapsed_seconds": 474.955, "final_quality": 0.9249749779701233, "logged_submission_platform": "MI355X", "official_aggregate_score": false, "quality_direction": "lower_is_better", "quality_metric": "validation_loss", "quality_target": 0.925, "run_status": "success" }, "run_index": 0, "source": { "bytes": 15825, "commit": "eabf23a07b2a0c60a289ff871dc3a46fff0d0421", "license": "Apache-2.0", "local": "amd-mi355x-training-run-0.txt", "local_sha256": "34aa01e24bb761b428b675b38654db4538d91c7557a35d9d5fb1f848ae42a074", "path": "AMD/results/MI355X_EPYC_9575F_primus/llama2_70b_lora/result_0.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/training_results_v6.0/eabf23a07b2a0c60a289ff871dc3a46fff0d0421/AMD/results/MI355X_EPYC_9575F_primus/llama2_70b_lora/result_0.txt", "repository": "https://github.com/mlcommons/training_results_v6.0", "source_sha256": "34aa01e24bb761b428b675b38654db4538d91c7557a35d9d5fb1f848ae42a074", "transformation": "none" }, "system": { "accelerator_frequency": "", "accelerator_host_interconnect": "PCIe Gen 5 x16", "accelerator_interconnect": "XGMI", "accelerator_interconnect_topology": "", "accelerator_memory_capacity": "288 GB", "accelerator_memory_configuration": "HBM3e", "accelerator_model_name": "AMD Instinct MI355X 288GB HBM3e", "accelerator_on-chip_memories": "", "accelerators_per_node": "8", "cooling": "Passive & Active", "division": "closed", "framework": "AMD Primus Framework 0.2.0", "framework_name": "v26.2_primus", "host_memory_capacity": "3.0 TB", "host_memory_configuration": "24 x 128GB Micron Technology MTC40F2047S1RC64BB1 UXCC", "host_networking": "2 x Ethernet Controller X710 for 10GBASE-T, 2 x Ethernet Controller 10G X550T", "host_networking_topology": "Ethernet on switching network", "host_processor_caches": "", "host_processor_core_count": "64", "host_processor_frequency": "", "host_processor_interconnect": "", "host_processor_model_name": "AMD EPYC 9575F", "host_processor_vcpu_count": "256", "host_processors_per_node": "2", "host_storage_capacity": "3.5T", "host_storage_type": "NVMe SSD", "hw_notes": "", "number_of_nodes": "1", "operating_system": "Ubuntu 24.04.4 LTS", "other_software_stack": { "amdgpu_driver_version": "30.30.0.0", "amdgpu_kernel_driver": "6.16.6", "hip_version": "7.2.26015", "hipblaslt_version": "1.2.0-de5c1aebb6~dirty", "kernel_version": "Linux 5.15.0-157-generic", "miopen_version": "3.5.1", "openmpi_version": "4.1.6", "rccl_version": "2.27.7", "rocm_version": "7.2.0" }, "status": "Available on-premise", "submitter": "AMD", "sw_notes": "", "system_name": "smci355-ccs-aus-n12-05 (AS -4126GS-NMR-LCC)" }, "version": "v6.0" }, "nvidia-b300": { "accuracy": "ROUGE1: 44.8452 ROUGE2: 22.4177 ROUGEL: 29.2149 TOKENS_PER_SAMPLE: 272.7", "availability": "available", "hardware": "NVIDIA B300-SXM-270GB", "inferred": false, "kind": "inference", "locally_measured": false, "measured_by": "NVIDIA", "observed": { "completed_requests_per_second": 391.79, "completed_tokens_per_second": 106949.36, "tpot_constraint_ms": 200.0, "tpot_p99_ms": 155.357236, "ttft_constraint_ms": 2000.0, "ttft_p99_ms": 480.656211 }, "query_unit_source": { "bytes": 268, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-b300-query-size.txt", "local_sha256": "f55cf197d112e7fa72d16af2a7928006dbb591cf6d24d9505b6f58d4e91db968", "path": "closed/NVIDIA/results/B300-SXM-270GBx8_TRT/llama2-70b-99.9/Server/performance/run_1/mlperf_log_detail.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/NVIDIA/results/B300-SXM-270GBx8_TRT/llama2-70b-99.9/Server/performance/run_1/mlperf_log_detail.txt", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "5b4fb373a78a4affac1a75e78bb89605b407c81f8247c5f4a63ce53ded751553", "transformation": "selected exact generated_samples_per_query event line; trailing newline normalized" }, "replica_boundary": "entire submitted eight-accelerator system", "scope": { "Division": "closed", "MlperfModel": "llama2-70b-99.9", "Platform": "B300-SXM-270GBx8_TRT", "Scenario": "Server", "framework": "TensorRT 10.14, CUDA 13.1, cuDNN 9.17, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.1", "total_accelerators": 8, "version": "v6.1", "weight_data_types": "fp4" }, "source": { "bytes": 3172, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-b300-inference-summary.txt", "local_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "path": "closed/NVIDIA/results/B300-SXM-270GBx8_TRT/llama2-70b-99.9/Server/performance/run_1/mlperf_log_summary.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/NVIDIA/results/B300-SXM-270GBx8_TRT/llama2-70b-99.9/Server/performance/run_1/mlperf_log_summary.txt", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "transformation": "none" } }, "nvidia-gb300-training": { "accelerator_count": 8, "benchmark": "llama2_70b_lora", "hardware_identity": { "hardware_specific_calibration_eligible": false, "limit": "A selected run is not hardware qualification; retain conflicting source labels without guessing a correction.", "logged_submission_platform": "2xNVIDIA GB200", "status": "unresolved_conflicting_source_labels", "system_declared_model": "NVIDIA Blackwell Ultra GPU (GB300)" }, "kind": "training", "locally_measured": false, "observed": { "elapsed_seconds": 333.08, "final_quality": 0.9244517683982849, "logged_submission_platform": "2xNVIDIA GB200", "official_aggregate_score": false, "quality_direction": "lower_is_better", "quality_metric": "validation_loss", "quality_target": 0.925, "run_status": "success" }, "run_index": 0, "source": { "bytes": 24480, "commit": "eabf23a07b2a0c60a289ff871dc3a46fff0d0421", "license": "Apache-2.0", "local": "nvidia-gb300-training-run-0.txt", "local_sha256": "d413dff2892c33e477b2aed90c88d57e79bd444bd2ba512eab70f9d2bcf96e14", "path": "NVIDIA/results/theia_ngpu8_ngc26.04_nemo/llama2_70b_lora/result_0.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/training_results_v6.0/eabf23a07b2a0c60a289ff871dc3a46fff0d0421/NVIDIA/results/theia_ngpu8_ngc26.04_nemo/llama2_70b_lora/result_0.txt", "repository": "https://github.com/mlcommons/training_results_v6.0", "source_sha256": "d413dff2892c33e477b2aed90c88d57e79bd444bd2ba512eab70f9d2bcf96e14", "transformation": "none" }, "system": { "accelerator_frequency": "", "accelerator_host_interconnect": "", "accelerator_interconnect": "NVLINK Gen5 1800 GB/s + NVSWITCH Gen5", "accelerator_interconnect_topology": "", "accelerator_memory_capacity": "284208 MiB", "accelerator_memory_configuration": "HBM3e", "accelerator_model_name": "NVIDIA Blackwell Ultra GPU (GB300)", "accelerator_on-chip_memories": "", "accelerators_per_node": "4", "cooling": "", "division": "closed", "framework": "NVIDIA NeMo Framework Release 26.04", "framework_name": "ngc26.04_nemo", "host_memory_capacity": "2.0 TB", "host_memory_configuration": "", "host_networking": "4x ConnectX-8 IB XDR 800 Gb/s", "host_networking_topology": "", "host_processor_caches": "", "host_processor_core_count": "72", "host_processor_frequency": "", "host_processor_interconnect": "", "host_processor_model_name": "Neoverse-V2", "host_processor_vcpu_count": "", "host_processors_per_node": "2", "host_storage_capacity": "8x 3.5TB NVMe E1.s", "host_storage_type": "NVMe SSD", "hw_notes": "", "number_of_nodes": "2", "operating_system": "Ubuntu 24.04.4 LTS", "other_software_stack": { "cublas_version": "13.4.0.1", "cuda_driver_version": "595.58.03", "cuda_version": "13.2.1.009", "cudnn_version": "9.21.1.3", "dali_version": "2.0.0", "kernel_version": "Linux 6.14.0-1015-nvidia-64k", "mofed_version": "5.4-rdmacore59.1", "nccl_version": "v2.30.4-1", "nvidia_kernel_driver": "580.159.04", "openmpi_version": "4.1.7", "trt_version": "10.16.1.11" }, "status": "Available on-premise", "submitter": "NVIDIA", "sw_notes": "", "system_name": "Theia (1x NVIDIA GB300 NVL72)" }, "version": "v6.0" }, "nvidia-rubin-preview": { "availability": "preview", "inferred": true, "kind": "inference", "locally_measured": false, "summary": { "Accuracy": "81.42661804922516", "Availability": "preview", "Division": "closed", "Location": "./closed/NVIDIA/results/VR-NVL72_VR-288GB_aarch64x72_TRT/deepseek-r1/Server/performance/run_1", "MlperfModel": "deepseek-r1", "Model": "deepseek-r1", "Organization": "NVIDIA", "Platform": "VR-NVL72_VR-288GB_aarch64x72_TRT", "PrivateID": null, "Result": 1175890.220816824, "Scenario": "Server", "SystemName": "NVIDIA Vera Rubin NVL72 (72x VR-288GB_aarch64)", "SystemType": "datacenter", "Units": "Tokens/s", "accelerator_model_name": "NVIDIA Vera Rubin", "accelerators_per_node": 4, "compliance": "closed", "design_power_watts": null, "errors": 0, "framework": "TRTLLM-rubin, vLLM-CentML:mlperf-inf-mm-q3vl-v6.1, Dynamo", "has_power": false, "host_processor_core_count": "144", "host_processor_model_name": "NVIDIA Vera CPU", "host_processors_per_node": 2, "inferred": 1, "notes": "NVIDIA Vera Rubin NVL72, 2300W TGP Per GPU", "number_of_nodes": 18, "operating_system": "Ubuntu 24.04", "total_accelerators": 72, "version": "v6.1", "weight_data_types": "fp4" } } }, "placement_sensitivity": { "assumptions": [ "0.8 goodput fraction; unchanged benchmark full-system replica", "Synthetic DE locations, independent labels, capacity inventory, RTT and scenario clock", "Two warm full-system replicas and four available per synthetic pool; 0.7 utilization", "Resident demand 0.1 times assumed goodput; one workload and one failover event" ], "cases": { "amd-mi355x": { "cases": { "expired": { "candidates": [ { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [], "pool_id": "synthetic-a", "reasons": [ "stale_or_future_evidence" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 378.6848 }, { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [], "pool_id": "synthetic-b", "reasons": [ "stale_or_future_evidence" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 378.6848 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 120, "execution_authority": false, "input_sha256": "ddb32089ddfad7427859321c9fcdb35e1823d022e60e6e7b89c53a2857b2ed3d", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "no_feasible_candidate" }, "load_0.5": { "candidates": [ { "candidate_for_review": true, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [ "synthetic-b" ], "pool_id": "synthetic-a", "reasons": [], "required_total_replicas": 1, "scale_up_replicas": 0, "warm_spillover_rps": 378.6848 }, { "candidate_for_review": true, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [ "synthetic-a" ], "pool_id": "synthetic-b", "reasons": [], "required_total_replicas": 1, "scale_up_replicas": 0, "warm_spillover_rps": 378.6848 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 100, "execution_authority": false, "input_sha256": "4a8bf4efd642cb1a6f9913400ec456e6efd73e761b8e107013deb076fecd2136", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "candidate_for_review" }, "load_1.0": { "candidates": [ { "candidate_for_review": true, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [ "synthetic-b" ], "pool_id": "synthetic-a", "reasons": [], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 378.6848 }, { "candidate_for_review": true, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [ "synthetic-a" ], "pool_id": "synthetic-b", "reasons": [], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 378.6848 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 100, "execution_authority": false, "input_sha256": "6dc225861bb8aa5bbb586372937b0ada8ad32307bac0edcb8e7443d16f2e901f", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "candidate_for_review" }, "load_1.5": { "candidates": [ { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [], "pool_id": "synthetic-a", "reasons": [ "critical_requires_warm_capacity", "cold_start_exceeds_deadline" ], "required_total_replicas": 3, "scale_up_replicas": 1, "warm_spillover_rps": 378.6848 }, { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [], "pool_id": "synthetic-b", "reasons": [ "critical_requires_warm_capacity", "cold_start_exceeds_deadline" ], "required_total_replicas": 3, "scale_up_replicas": 1, "warm_spillover_rps": 378.6848 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 100, "execution_authority": false, "input_sha256": "b03bb8a49bde524d9048fd44b89ec64c3bde6b76ac02e7c0523e6013a9b004eb", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "no_feasible_candidate" }, "profile_mismatch": { "candidates": [ { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [], "pool_id": "synthetic-a", "reasons": [ "workload_profile_mismatch" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 378.6848 }, { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [], "pool_id": "synthetic-b", "reasons": [ "no_independent_warm_failover" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 378.6848 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 100, "execution_authority": false, "input_sha256": "b4b08a4885682ee83943a6fbe178f2f0400a71f3ef4b1fde94a8cd11e1a32165", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "no_feasible_candidate" }, "stricter_latency": { "candidates": [ { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [], "pool_id": "synthetic-a", "reasons": [ "latency:ttft", "latency:tpot" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 378.6848 }, { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200", "failover_pool_ids": [], "pool_id": "synthetic-b", "reasons": [ "latency:ttft", "latency:tpot" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 378.6848 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 100, "execution_authority": false, "input_sha256": "5d428d0bb445c090f32c5057123fd5fa381af0e35cd050b3886552d8381156f2", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "no_feasible_candidate" } }, "scope": { "Division": "closed", "MlperfModel": "llama2-70b-99.9", "Platform": "8xMI355X_2xEPYC_9575F", "Scenario": "Server", "framework": "PyTorch 2.10.0+git8514f05, ROCm 7.2.2", "total_accelerators": 8, "version": "v6.1", "weight_data_types": "fp4" }, "source_sha256": "c5e25bfc61295aae5f94eba085d49f75eb59d823e8bd64491027abaaba03e200" }, "nvidia-b300": { "cases": { "expired": { "candidates": [ { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [], "pool_id": "synthetic-a", "reasons": [ "stale_or_future_evidence" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 407.4616 }, { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [], "pool_id": "synthetic-b", "reasons": [ "stale_or_future_evidence" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 407.4616 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 120, "execution_authority": false, "input_sha256": "75896147b148ff521093c056650d89146d5c8cb5df28cb9c788f3f2d9085dcbe", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "no_feasible_candidate" }, "load_0.5": { "candidates": [ { "candidate_for_review": true, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [ "synthetic-b" ], "pool_id": "synthetic-a", "reasons": [], "required_total_replicas": 1, "scale_up_replicas": 0, "warm_spillover_rps": 407.4616 }, { "candidate_for_review": true, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [ "synthetic-a" ], "pool_id": "synthetic-b", "reasons": [], "required_total_replicas": 1, "scale_up_replicas": 0, "warm_spillover_rps": 407.4616 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 100, "execution_authority": false, "input_sha256": "de7cd9e33eb3973081e1e73e718204e25730ddc4058fe0fba3773d45f811c645", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "candidate_for_review" }, "load_1.0": { "candidates": [ { "candidate_for_review": true, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [ "synthetic-b" ], "pool_id": "synthetic-a", "reasons": [], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 407.4616 }, { "candidate_for_review": true, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [ "synthetic-a" ], "pool_id": "synthetic-b", "reasons": [], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 407.4616 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 100, "execution_authority": false, "input_sha256": "391e70dc47886ae824a399f3a392327559137802c552fe1331e74f0218ff1b29", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "candidate_for_review" }, "load_1.5": { "candidates": [ { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [], "pool_id": "synthetic-a", "reasons": [ "critical_requires_warm_capacity", "cold_start_exceeds_deadline" ], "required_total_replicas": 3, "scale_up_replicas": 1, "warm_spillover_rps": 407.4616 }, { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [], "pool_id": "synthetic-b", "reasons": [ "critical_requires_warm_capacity", "cold_start_exceeds_deadline" ], "required_total_replicas": 3, "scale_up_replicas": 1, "warm_spillover_rps": 407.4616 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 100, "execution_authority": false, "input_sha256": "c25f26a6f2e4cca139da350e5d9bbafb94ff91bce00f48299b9dd828d2f4c1d9", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "no_feasible_candidate" }, "profile_mismatch": { "candidates": [ { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [], "pool_id": "synthetic-a", "reasons": [ "workload_profile_mismatch" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 407.4616 }, { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [], "pool_id": "synthetic-b", "reasons": [ "no_independent_warm_failover" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 407.4616 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 100, "execution_authority": false, "input_sha256": "bb2b003e4ddb32bcf410e1137a36fcee8aa83614238185e671b70be0d4649713", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "no_feasible_candidate" }, "stricter_latency": { "candidates": [ { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [], "pool_id": "synthetic-a", "reasons": [ "latency:tpot" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 407.4616 }, { "candidate_for_review": false, "evidence_kind": "synthetic", "evidence_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c", "failover_pool_ids": [], "pool_id": "synthetic-b", "reasons": [ "latency:tpot" ], "required_total_replicas": 2, "scale_up_replicas": 0, "warm_spillover_rps": 407.4616 } ], "country_registry_sha256": "fe17c79d024cba8a172a7bef6ef6276a7f1247af042738612697ca8fe3d5ba2f", "evaluation_time_s": 100, "execution_authority": false, "input_sha256": "8d45894609cd6a5367b19a5c30f980d6030febdb3f19c9eb2f2ee9c8f9c96895", "limits": [ "Caller-supplied evidence is not authenticated; no Tool Guard call was made.", "Independent labels and linear replica scaling require deployment evidence.", "One workload increment and one failover event; concurrent shared reserve is not allocated.", "Locality lists express policy constraints, not legal compliance or sovereign assurance." ], "policy_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "production_qualified": false, "schema": "dimaggi-regional-plan/v1", "status": "no_feasible_candidate" } }, "scope": { "Division": "closed", "MlperfModel": "llama2-70b-99.9", "Platform": "B300-SXM-270GBx8_TRT", "Scenario": "Server", "framework": "TensorRT 10.14, CUDA 13.1, cuDNN 9.17, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.1", "total_accelerators": 8, "version": "v6.1", "weight_data_types": "fp4" }, "source_sha256": "d2e1cc9d53686ea09f4731c1a8b08dd4fd3ee67cff7ad7985b46909a7c1fc15c" } }, "evidence_kind": "mixed_public_observations_and_synthetic_scenario", "execution_authority": false, "production_qualified": false, "public_data_does_not_establish": [ "geographic availability", "fresh deployment evidence", "customer latency", "cost", "hardware qualification" ] }, "production_qualified": false, "size_comparison": { "benchmark_scope": { "Division": "closed", "MlperfModel": "deepseek-r1", "Scenario": "Server", "framework": "TensorRT 10.14, CUDA 13.1, cuDNN 9.17, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.1", "tpot_constraint_ms": 80.0, "ttft_constraint_ms": 2000.0, "version": "v6.1", "weight_data_types": "fp4" }, "comparison_kind": "published_system_size_observations", "confounders": [ "Different achieved quality and output lengths; common benchmark category does not mean identical outputs.", "288-device system uses Driver 580.67; 8/72-device systems list 580.105, with different host-memory declarations.", "288 source system_name says 72x while node/device fields and summary say 288; hardware identity is unresolved.", "Declared fabric labels lack complete measured topology/collective data; software configuration and replica layout are not controlled." ], "execution_authority": false, "limits": [ "Per-device normalization is an efficiency diagnostic, not independently serviceable per-GPU capacity.", "No interpolation/extrapolation or automatic placement/pricing calibration is supported.", "No local load, fault, region, operator, governance or recovery execution was performed." ], "observations": [ { "accelerators_per_node_field": 4, "accuracy": "exact_match: 80.78851412944394 TOKENS_PER_SAMPLE: 3829.014357338195", "capacity_conversion_allowed": false, "derived_accelerator_count": 8, "diagnostic_normalization": { "request": { "normalized_throughput_ratio_vs_8": 1.0, "throughput_per_declared_accelerator": 2.2125, "throughput_ratio_vs_8": 1.0 }, "token": { "normalized_throughput_ratio_vs_8": 1.0, "throughput_per_declared_accelerator": 8447.29375, "throughput_ratio_vs_8": 1.0 } }, "hardware_identity_status": "declared_counts_agree", "id": "nvidia-gb300-8-size", "locally_measured": false, "node_field_count": 2, "observed": { "completed_requests_per_second": 17.7, "completed_tokens_per_second": 67578.35, "tpot_constraint_ms": 80.0, "tpot_p99_ms": 73.885145, "ttft_constraint_ms": 2000.0, "ttft_p99_ms": 985.577332 }, "platform": "GB300-NVL72_GB300-288GB_aarch64x8_TRT", "profile_sha256": "ec410b9d0ac7ae6fe4d05e66869d72d8dae282acc52b250b08307ba3b8d1c860", "size_ratio_vs_8": 1.0, "sources": { "nvidia-gb300-8-inference-summary.txt": { "bytes": 3174, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-8-inference-summary.txt", "local_sha256": "64af50929a3c734d9dd3f6f902542afce71eb77886df704afbc889f19cad3f97", "path": "closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x8_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_summary.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x8_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_summary.txt", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "64af50929a3c734d9dd3f6f902542afce71eb77886df704afbc889f19cad3f97", "transformation": "none" }, "nvidia-gb300-8-query-size.txt": { "bytes": 268, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-8-query-size.txt", "local_sha256": "83eb68be98020bbb3017f4b737130010cc7f635e1eaffeb3dc6a32b635786346", "path": "closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x8_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_detail.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x8_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_detail.txt", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "54fca75876ee233af6e32d01a2b0f921524386e14bed7d7fbb0af1db4a0d3e8f", "transformation": "selected exact generated_samples_per_query event line; trailing newline normalized" }, "nvidia-gb300-8-system.json": { "bytes": 2454, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-8-system.json", "local_sha256": "ccaca25d6b532e8e648b4b9543b933ce8d032471c89d2f8a8d97411117240ad8", "path": "closed/NVIDIA/systems/GB300-NVL72_GB300-288GB_aarch64x8_TRT.json", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/NVIDIA/systems/GB300-NVL72_GB300-288GB_aarch64x8_TRT.json", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "ccaca25d6b532e8e648b4b9543b933ce8d032471c89d2f8a8d97411117240ad8", "transformation": "none" }, "nvidia-gb300-size-summary-selection.json": { "bytes": 3988, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-size-summary-selection.json", "local_sha256": "64a7934cc03a11bd88dc0418289a289b1cd7c1f5a1db5b946c9083372449f6b2", "path": "summary.json", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/summary.json", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "f40d71a728a8c1e647d5fef9869ec1fcbb54550433efbc12e96f90b09a2dd561", "transformation": "selected records 318, 308, 305; JSON whitespace normalized" } }, "summary_declared_accelerators": 8, "system": { "accelerator_frequency": "", "accelerator_host_interconnect": "NVLink-C2C", "accelerator_interconnect": "5th-gen NVLink andNVLink Switches", "accelerator_interconnect_topology": "", "accelerator_memory_capacity": "288 GB", "accelerator_memory_configuration": "HBM3e", "accelerator_model_name": "NVIDIA GB300", "accelerator_on-chip_memories": "", "accelerators_per_node": 4, "boot_firmware_version": "", "cooling": "Mix of Liquid-cooled (for CPU, GPU, CX7, NVLink Switches) and Air-cooled (for the rest of the components)", "disk_controllers": "NVMe", "disk_drives": "SSD", "division": "closed", "filesystem": "", "framework": "TensorRT 10.14, CUDA 13.1, cuDNN 9.17, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.1", "host_memory_capacity": "2.2 TB", "host_memory_configuration": "for each 4-GPU node: 2.2TB LPDDR5x", "host_network_card_count": "for each 4-GPU node: 4x Mellanox MT43244 BlueField-3 [ConnectX-7 Lx], 1x Intel I210 Gigabit Ethernet", "host_networking": "Ethernet(IPoIB)", "host_networking_topology": "Ethernet(IPoIB)/Infiniband on switching network", "host_processor_caches": "", "host_processor_core_count": 144, "host_processor_frequency": "", "host_processor_interconnect": "", "host_processor_model_name": "NVIDIA Grace CPU", "host_processors_per_node": 2, "host_storage_capacity": "for each 4-GPU node: 2 TB SSD, 5 TB CIFS", "host_storage_type": "NVMe SSD, CIFS mounted disk storage", "hw_notes": "NVIDIA GB300 NVL72, 1400W TGP Per GPU", "management_firmware_version": "", "network_speed_mbit": "", "nics_enabled_connected": "", "nics_enabled_firmware": "", "nics_enabled_os": "", "number_of_nodes": 2, "number_of_type_nics_installed": "", "operating_system": "Ubuntu 24.04", "other_hardware": "", "other_software_stack": "TensorRT 10.14, CUDA 13.1, cuDNN 9.17, Driver 580.105, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.0", "power_management": "", "power_supply_details": "", "power_supply_quantity_and_rating_watts": "", "status": "available", "submitter": "NVIDIA", "sw_notes": "", "system_name": "NVIDIA GB300 NVL72 (8x GB300-288GB_aarch64, TensorRT)", "system_type": "datacenter", "system_type_detail": "N/A" }, "system_name_declared_count": 8 }, { "accelerators_per_node_field": 4, "accuracy": "exact_match: 81.42661804922516 TOKENS_PER_SAMPLE: 3855.550136736554", "capacity_conversion_allowed": false, "derived_accelerator_count": 72, "diagnostic_normalization": { "request": { "normalized_throughput_ratio_vs_8": 0.9807281858129315, "throughput_per_declared_accelerator": 2.169861111111111, "throughput_ratio_vs_8": 8.826553672316384 }, "token": { "normalized_throughput_ratio_vs_8": 0.9814840929906627, "throughput_per_declared_accelerator": 8290.884444444444, "throughput_ratio_vs_8": 8.833356836915964 } }, "hardware_identity_status": "declared_counts_agree", "id": "nvidia-gb300-72-size", "locally_measured": false, "node_field_count": 18, "observed": { "completed_requests_per_second": 156.23, "completed_tokens_per_second": 596943.68, "tpot_constraint_ms": 80.0, "tpot_p99_ms": 76.846424, "ttft_constraint_ms": 2000.0, "ttft_p99_ms": 1006.052508 }, "platform": "GB300-NVL72_GB300-288GB_aarch64x72_TRT", "profile_sha256": "bcc2bf1a25cc3766517707451507c1e4ca531211fa35ef08cb6791a114269ea0", "size_ratio_vs_8": 9.0, "sources": { "nvidia-gb300-72-inference-summary.txt": { "bytes": 3177, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-72-inference-summary.txt", "local_sha256": "626ae678d15d83d1179beaea8ded96d544056405c08998b3773dee30edff709d", "path": "closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x72_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_summary.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x72_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_summary.txt", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "626ae678d15d83d1179beaea8ded96d544056405c08998b3773dee30edff709d", "transformation": "none" }, "nvidia-gb300-72-query-size.txt": { "bytes": 269, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-72-query-size.txt", "local_sha256": "910b35f53262b7dfb8c3efb46f8b40d1e80914f8b9dabf869efa8c006efc413a", "path": "closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x72_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_detail.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x72_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_detail.txt", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "62216134472417f8b0fd0992e7c4b3a5adce606a3af6f30ed2d466d0a29f3f2f", "transformation": "selected exact generated_samples_per_query event line; trailing newline normalized" }, "nvidia-gb300-72-system.json": { "bytes": 2455, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-72-system.json", "local_sha256": "b8941c89d09204cbbc45de5748868f8d5735a95c83ba34c71c2593e1dbc1c619", "path": "closed/NVIDIA/systems/GB300-NVL72_GB300-288GB_aarch64x72_TRT.json", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/NVIDIA/systems/GB300-NVL72_GB300-288GB_aarch64x72_TRT.json", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "b8941c89d09204cbbc45de5748868f8d5735a95c83ba34c71c2593e1dbc1c619", "transformation": "none" }, "nvidia-gb300-size-summary-selection.json": { "bytes": 3988, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-size-summary-selection.json", "local_sha256": "64a7934cc03a11bd88dc0418289a289b1cd7c1f5a1db5b946c9083372449f6b2", "path": "summary.json", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/summary.json", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "f40d71a728a8c1e647d5fef9869ec1fcbb54550433efbc12e96f90b09a2dd561", "transformation": "selected records 318, 308, 305; JSON whitespace normalized" } }, "summary_declared_accelerators": 72, "system": { "accelerator_frequency": "", "accelerator_host_interconnect": "NVLink-C2C", "accelerator_interconnect": "5th-gen NVLink andNVLink Switches", "accelerator_interconnect_topology": "", "accelerator_memory_capacity": "288 GB", "accelerator_memory_configuration": "HBM3e", "accelerator_model_name": "NVIDIA GB300", "accelerator_on-chip_memories": "", "accelerators_per_node": 4, "boot_firmware_version": "", "cooling": "Mix of Liquid-cooled (for CPU, GPU, CX7, NVLink Switches) and Air-cooled (for the rest of the components)", "disk_controllers": "NVMe", "disk_drives": "SSD", "division": "closed", "filesystem": "", "framework": "TensorRT 10.14, CUDA 13.1, cuDNN 9.17, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.1", "host_memory_capacity": "2.2 TB", "host_memory_configuration": "for each 4-GPU node: 2.2TB LPDDR5x", "host_network_card_count": "for each 4-GPU node: 4x Mellanox MT43244 BlueField-3 [ConnectX-7 Lx], 1x Intel I210 Gigabit Ethernet", "host_networking": "Ethernet(IPoIB)", "host_networking_topology": "Ethernet(IPoIB)/Infiniband on switching network", "host_processor_caches": "", "host_processor_core_count": 144, "host_processor_frequency": "", "host_processor_interconnect": "", "host_processor_model_name": "NVIDIA Grace CPU", "host_processors_per_node": 2, "host_storage_capacity": "for each 4-GPU node: 2 TB SSD, 5 TB CIFS", "host_storage_type": "NVMe SSD, CIFS mounted disk storage", "hw_notes": "NVIDIA GB300 NVL72, 1400W TGP Per GPU", "management_firmware_version": "", "network_speed_mbit": "", "nics_enabled_connected": "", "nics_enabled_firmware": "", "nics_enabled_os": "", "number_of_nodes": 18, "number_of_type_nics_installed": "", "operating_system": "Ubuntu 24.04", "other_hardware": "", "other_software_stack": "TensorRT 10.14, CUDA 13.1, cuDNN 9.17, Driver 580.105, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.0", "power_management": "", "power_supply_details": "", "power_supply_quantity_and_rating_watts": "", "status": "available", "submitter": "NVIDIA", "sw_notes": "", "system_name": "NVIDIA GB300 NVL72 (72x GB300-288GB_aarch64, TensorRT)", "system_type": "datacenter", "system_type_detail": "N/A" }, "system_name_declared_count": 72 }, { "accelerators_per_node_field": 4, "accuracy": "exact_match: 81.13035551504102 TOKENS_PER_SAMPLE: 3861.4316317228804", "capacity_conversion_allowed": false, "derived_accelerator_count": 288, "diagnostic_normalization": { "request": { "normalized_throughput_ratio_vs_8": 0.8324231010671689, "throughput_per_declared_accelerator": 1.841736111111111, "throughput_ratio_vs_8": 29.967231638418077 }, "token": { "normalized_throughput_ratio_vs_8": 0.8336131155745722, "throughput_per_declared_accelerator": 7041.774861111111, "throughput_ratio_vs_8": 30.010072160684597 } }, "hardware_identity_status": "unresolved_conflicting_source_labels", "id": "nvidia-gb300-288-size", "locally_measured": false, "node_field_count": 72, "observed": { "completed_requests_per_second": 530.42, "completed_tokens_per_second": 2028031.16, "tpot_constraint_ms": 80.0, "tpot_p99_ms": 76.564557, "ttft_constraint_ms": 2000.0, "ttft_p99_ms": 969.502207 }, "platform": "GB300-NVL72_GB300-288GB_aarch64x288_TRT", "profile_sha256": "0acff970e269ed46ba20bcb8421b49ea365562f5a82dcb24cadc91cf723e9e7c", "size_ratio_vs_8": 36.0, "sources": { "nvidia-gb300-288-inference-summary.txt": { "bytes": 3179, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-288-inference-summary.txt", "local_sha256": "206fd16d3a59091ba35324dfbbf381ee4b0d2aa93e7566873c6d29d317c60ee2", "path": "closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x288_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_summary.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x288_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_summary.txt", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "206fd16d3a59091ba35324dfbbf381ee4b0d2aa93e7566873c6d29d317c60ee2", "transformation": "none" }, "nvidia-gb300-288-query-size.txt": { "bytes": 270, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-288-query-size.txt", "local_sha256": "56289f58ef458ee797ff7cf403838a8ed4d7c138c94242075ad16531c91528da", "path": "closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x288_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_detail.txt", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/NVIDIA/results/GB300-NVL72_GB300-288GB_aarch64x288_TRT/deepseek-r1/Server/performance/run_1/mlperf_log_detail.txt", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "ff427f33bd3d5565715c75fb2444d87df58dc36ccb68ce86cfbd6ec02147e78f", "transformation": "selected exact generated_samples_per_query event line; trailing newline normalized" }, "nvidia-gb300-288-system.json": { "bytes": 2436, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-288-system.json", "local_sha256": "8617d3dd082440dad510a9e83c7e0e2041d8e4b3d56b2e7b4995d75eb5fb6856", "path": "closed/NVIDIA/systems/GB300-NVL72_GB300-288GB_aarch64x288_TRT.json", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/closed/NVIDIA/systems/GB300-NVL72_GB300-288GB_aarch64x288_TRT.json", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "8617d3dd082440dad510a9e83c7e0e2041d8e4b3d56b2e7b4995d75eb5fb6856", "transformation": "none" }, "nvidia-gb300-size-summary-selection.json": { "bytes": 3988, "commit": "4bb63cd28eb136be221bc801a9e38c66c8a7553f", "license": "Apache-2.0", "local": "nvidia-gb300-size-summary-selection.json", "local_sha256": "64a7934cc03a11bd88dc0418289a289b1cd7c1f5a1db5b946c9083372449f6b2", "path": "summary.json", "raw_url": "https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/4bb63cd28eb136be221bc801a9e38c66c8a7553f/summary.json", "repository": "https://github.com/mlcommons/inference_results_v6.1", "source_sha256": "f40d71a728a8c1e647d5fef9869ec1fcbb54550433efbc12e96f90b09a2dd561", "transformation": "selected records 318, 308, 305; JSON whitespace normalized" } }, "summary_declared_accelerators": 288, "system": { "accelerator_frequency": "", "accelerator_host_interconnect": "NVLink-C2C", "accelerator_interconnect": "5th-gen NVLink andNVLink Switches", "accelerator_interconnect_topology": "", "accelerator_memory_capacity": "288 GB", "accelerator_memory_configuration": "HBM3e", "accelerator_model_name": "NVIDIA GB300", "accelerator_on-chip_memories": "", "accelerators_per_node": 4, "boot_firmware_version": "", "cooling": "Mix of Liquid-cooled (for CPU, GPU, CX7, NVLink Switches) and Air-cooled (for the rest of the components)", "disk_controllers": "NVMe", "disk_drives": "SSD", "division": "closed", "filesystem": "", "framework": "TensorRT 10.14, CUDA 13.1, cuDNN 9.17, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.1", "host_memory_capacity": "960 GB", "host_memory_configuration": "for each 4-GPU node: 1.3TB LPDDR5x", "host_network_card_count": "for each 4-GPU node: 4x Mellanox MT43244 BlueField-3 [ConnectX-7 Lx], 1x Intel I210 Gigabit Ethernet", "host_networking": "Ethernet(IPoIB)", "host_networking_topology": "Ethernet(IPoIB)/Infiniband on switching network", "host_processor_caches": "", "host_processor_core_count": 144, "host_processor_frequency": "", "host_processor_interconnect": "", "host_processor_model_name": "NVIDIA Grace CPU", "host_processors_per_node": 2, "host_storage_capacity": "for each 4-GPU node: 2 TB SSD, 5 TB CIFS", "host_storage_type": "NVMe SSD, CIFS mounted disk storage", "hw_notes": "NVIDIA GB300 NVL72", "management_firmware_version": "", "network_speed_mbit": "", "nics_enabled_connected": "", "nics_enabled_firmware": "", "nics_enabled_os": "", "number_of_nodes": 72, "number_of_type_nics_installed": "", "operating_system": "Ubuntu 24.04", "other_hardware": "", "other_software_stack": "TensorRT 10.14, CUDA 13.1, cuDNN 9.17, Driver 580.67, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.0", "power_management": "", "power_supply_details": "", "power_supply_quantity_and_rating_watts": "", "status": "available", "submitter": "NVIDIA", "sw_notes": "", "system_name": "NVIDIA GB300 NVL72 (72x GB300-288GB_aarch64, TensorRT)", "system_type": "datacenter", "system_type_detail": "N/A" }, "system_name_declared_count": 72 } ], "production_qualified": false, "pure_hardware_scaling": false } }