Argonne LLM-Inference-Bench: MI250 x1, vLLM 0.6.1, Llama-2-7B FP16, 128in/128out, batch 1
139.9 tok/s aggregate (batch 1 = single request; in+out tokens / latency)
Fields
id
benchmarks/argonne-mi250-vllm-llama2-7b-fp16-b1-128io-tpsMetric
tps_aggregate
Value
139.9
Unit
tok/s aggregate (batch 1 = single request; in+out tokens / latency)
Model
meta-llama/Llama-2-7b-hf
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 1, 1x MI250 (128GB HBM2e, 2 GCDs). Latency 1.830s.
Measured by
third_party
Reproducible
yes
notes
MI250 is the dual-GCD 128GB part on Polaris. At batch 1 the 128-out decode phase is bandwidth-bound; 139.9 tok/s aggregate for 1 request approximates ~22 tok/s/user effective decode plus prefill share — treat as end-to-end request throughput.
Notes
MI250 is the dual-GCD 128GB part on Polaris. At batch 1 the 128-out decode phase is bandwidth-bound; 139.9 tok/s aggregate for 1 request approximates ~22 tok/s/user effective decode plus prefill share — treat as end-to-end request throughput.