Argonne MI250 x4 llama.cpp Qwen/Qwen2-72B FP16 b1 128io
4.58 tok/s aggregate (batch*(in+out)/latency)
Fields
id
benchmarks/argonne-mi250-llamacpp-qwen2-72b-fp16-4x-b1-128io-tpsMetric
tps_aggregate
Value
4.58
Unit
tok/s aggregate (batch*(in+out)/latency)
Model
Qwen/Qwen2-72B
Methodology
THIRD-PARTY MEASURED (Argonne LLM-Inference-Bench, arXiv:2411.00136, All_results.csv). llama.cpp on AMD MI250 x4 (Polaris node; MI250X silicon, filed as MI250), FP16, 128 in / 128 out, batch 1. Throughput = batch*(in+out)/latency = AGGREGATE across batch, NOT per-request decode. Latency 55.84s.
Measured by
third_party
Reproducible
yes
Thermal state
unknown
notes
Polaris node, 4x MI250. llama.cpp llama-bench-style throughput; batch-1 value approximates single-request decode.
Notes
Polaris node, 4x MI250. llama.cpp llama-bench-style throughput; batch-1 value approximates single-request decode.