Skip to content

GPU Benchmarking

Within the PoA’s health check, the drill test incorporates sophisticated benchmarking techniques such as MLPerf to evaluate machine performance comprehensively. By conducting benchmarking assessments, including MLPerf, the algorithm quantifies the machine’s efficiency. This quantitative measure serves as a reliable indicator of the machine’s condition, ensuring robustness and reliability in its operational capabilities.

Work flow of Drill Test

Here are some sample results of the drill test on Nvidia A100:

MLPerf Results Summary:

FieldValue
SUT nameBERT SERVER
ScenarioOffline
ModePerformanceOnly
Samples per second1532.17
ResultVALID
Min duration satisfiedYes
Min queries satisfiedYes
Early stopping satisfiedYes

Additional Stats:

MetricValue (ns)
Min latency3,559,383,281
Max latency1,292,280,950,807
Mean latency788,846,755,872
50.00 percentile latency840,201,049,914
90.00 percentile latency1,234,598,190,171
95.00 percentile latency1,268,998,116,410
97.00 percentile latency1,280,065,956,777
99.00 percentile latency1,289,280,826,440
99.90 percentile latency1,292,043,266,934

Test Parameters Used:

ParameterValue
samples_per_query1,980,000
target_qps3,000
target_latency (ns)0
max_async_queries1
min_duration (ms)600,000
max_duration (ms)0
min_query_count1
max_query_count0
qsl_rng_seed13,281,865,557,512,327,830
sample_index_rng_seed198,141,574,272,810,017
schedule_rng_seed7,575,108,116,881,280,410
accuracy_log_rng_seed0
accuracy_log_probability0
accuracy_log_sampling_target0
print_timestamps0
performance_issue_unique0
performance_issue_same0
performance_issue_same_index0
performance_sample_count10,833