NVIDIA L4 x8 (ASUSTeK ESC8000_E11), MLPerf v4.0 Offline: bert-99.9 aggregate samples/s
5189.92 samples/s AGGREGATE across 8x L4-PCIe-24GB (Offline; ~649 samples/s per card)
Fields
id
benchmarks/l4-8x-mlperf-v4-0-bert999-offline-samplesMetric
tps_aggregate
Value
5,189.92
Unit
samples/s AGGREGATE across 8x L4-PCIe-24GB (Offline; ~649 samples/s per card)
Model
bert-99.9 (TensorRT, INT8+FP16)
Methodology
VENDOR-MEASURED (MLPerf v4.0 closed submission, ASUSTeK ESC8000_E11). 8x NVIDIA L4 24GB, TensorRT, bert-99.9, Offline scenario. 'Samples per second' = 5189.92 verbatim from mlperf_log_summary.txt. AGGREGATE across 8 cards — ~649/card. BERT-large is an NLP encoder (not a decoder LLM): kept as the ONLY audited-vendor number on the L4 record because every MLPerf LLM round skipped the L4 entirely.
Measured by
vendor
Reproducible
yes
notes
Not a token-rate: BERT samples/s. Filed because the L4 is the affordable inference card and this is its only closed-division MLPerf presence — llama.cpp/vLLM numbers in sibling records are the LLM datapoints.
Notes
Not a token-rate: BERT samples/s. Filed because the L4 is the affordable inference card and this is its only closed-division MLPerf presence — llama.cpp/vLLM numbers in sibling records are the LLM datapoints.