Argonne LLM-Inference-Bench: Gaudi2 x8, DeepSpeed, Llama-2-70B BF16, 128in/128out, batch 64
4764.8 tok/s aggregate across 64 batched requests on 8x Gaudi2
Fields
id
benchmarks/argonne-gaudi2-deepspeed-llama2-70b-bf16-8x-b64-128io-tpsMetric
tps_aggregate
Value
4,764.8
Unit
tok/s aggregate across 64 batched requests on 8x Gaudi2
Model
meta-llama/Llama-2-70b-hf
Methodology
THIRD-PARTY MEASURED (Argonne LLM-Inference-Bench, arXiv:2411.00136, All_results.csv). FP16/BF16 weights, 128 in / 128 out. Throughput = batch*(in+out)/end-to-end latency = AGGREGATE over the batch, NOT per-request decode tok/s. Batch 64, 8x Gaudi2 TP (RoCE fabric between HPUs). Latency 3.439s.
Measured by
third_party
Reproducible
no
notes
Complements the existing MLPerf v4.0 llama2-70b Gaudi2 row (closed-division, Habana's own submission) with a third-party DeepSpeed datapoint.
Notes
Complements the existing MLPerf v4.0 llama2-70b Gaudi2 row (closed-division, Habana's own submission) with a third-party DeepSpeed datapoint.