Intel Gaudi 3, Optimum-Habana (sw 1.24): Llama-3.1-70B-Instruct FP8 2HPU bs223 2048in/128out aggregate tok/s
650 output tok/s AGGREGATE across 2x Gaudi 3 (batch 223; ~325 tok/s per HPU)
Fields
id
benchmarks/gaudi3-2hpu-llama31-70b-fp8-2048-128Metric
tps_aggregate
Value
650
Unit
output tok/s AGGREGATE across 2x Gaudi 3 (batch 223; ~325 tok/s per HPU)
Model
Llama-3.1-70B-Instruct
Methodology
VENDOR-MEASURED (Intel model-performance table, Gaudi software 1.24, Optimum-Habana + Hugging Face model, maxabs FP8). Llama-3.1-70B-Instruct FP8 on 2 HPU, batch_size=223, input_len=2048, output_len=128. MAX THROUGHPUT under a large batch, NOT a single-user decode figure. ~325 tok/s per HPU.
Measured by
vendor
Reproducible
yes
notes
Optimum-Habana run_generation.py path, use_hpu_graphs + flash_attention_recompute. Sibling rows from the same table were already ingested; these are the previously unrecorded input/output/HPU combinations.
Notes
Optimum-Habana run_generation.py path, use_hpu_graphs + flash_attention_recompute. Sibling rows from the same table were already ingested; these are the previously unrecorded input/output/HPU combinations.