Argonne LLM-Inference-Bench: MI250 x1, vLLM 0.6.1, Llama-3-8B FP16, 128in/128out, batch 64
3751.2 tok/s aggregate across 64 batched requests
Fields
id
benchmarks/argonne-mi250-vllm-llama3-8b-fp16-b64-128io-tpsMetric
tps_aggregate
Value
3,751.2
Unit
tok/s aggregate across 64 batched requests
Model
meta-llama/Meta-Llama-3-8B
Methodology
THIRD-PARTY MEASURED (Argonne LLM-Inference-Bench, arXiv:2411.00136, All_results.csv). vLLM benchmark_throughput.py, FP16/BF16, 128 in / 128 out. Throughput = batch*(in+out)/end-to-end latency = AGGREGATE over the batch, NOT per-request decode. Batch 64, 1x MI250. Latency 4.368s.
Measured by
third_party
Reproducible
yes
notes
Paper notes MI250 throughput for Llama-3-8B DROPS beyond batch 32 at longer input/output (NUMA page-fault waiter); bs64/128io is a peak point, not a plateau.
Notes
Paper notes MI250 throughput for Llama-3-8B DROPS beyond batch 32 at longer input/output (NUMA page-fault waiter); bs64/128io is a peak point, not a plateau.