MLPerf v3.0: Dell r7515, 4x Qualcomm Cloud AI 100 Pro, BERT-99.9 Offline 1487.12 samples/s
1487.12 samples/s aggregate, 4-card node, Offline scenario
Fields
id
benchmarks/qaic-ai100-4x-mlperf-v3-0-bert999-offline-samplesMetric
tps_aggregate
Value
1,487.12
Unit
samples/s aggregate, 4-card node, Offline scenario
Model
bert-large-squad (bert-99.9, high-accuracy)
Methodology
MLPerf Inference v3.0 CLOSED submission, Dell r7515 with 4x Qualcomm Cloud AI 100 Pro cards (q4_pro config, Qualcomm Apps SDK v1.8.3.7, 'aic100' backend). BERT-large SQuAD INT8, Offline scenario: throughput = completed samples/s read verbatim from performance/run_1/mlperf_log_summary.txt ('Result is : VALID ; Scheduled samples per second'). Non-LLM encoder workload; filed to give qualcomm-ai100 a verified vendor-submitted row since no public LLM tok/s exists for the base AI100 (all published LLM numbers are for the AI100 Ultra, filed under qualcomm-ai200).
Measured by
vendor
Reproducible
yes