Intel Gaudi 3, Optimum-Habana (sw 1.24): Llama-3.3-70B-Instruct FP8 8HPU bs600 2048in/2048out aggregate tok/s
10747 output tok/s AGGREGATE across 8x Gaudi 3 (batch 600; ~1343 tok/s per HPU)
Fields
id
benchmarks/gaudi3-8hpu-llama33-70b-fp8-2048-2048Metric
tps_aggregate
Value
10,747
Unit
output tok/s AGGREGATE across 8x Gaudi 3 (batch 600; ~1343 tok/s per HPU)
Model
Llama-3.3-70B-Instruct
Methodology
VENDOR-MEASURED (Intel model-performance table, Gaudi software 1.24, Optimum-Habana + Hugging Face model, maxabs FP8). Llama-3.3-70B-Instruct FP8 on 8 HPU, batch_size=600, input_len=2048, output_len=2048. This is MAX THROUGHPUT under a very large batch — NOT a single-user decode figure. ~1343 tok/s per HPU.
Measured by
vendor
Reproducible
yes
notes
Optimum-Habana run_generation.py path, use_hpu_graphs + flash_attention_recompute. Reproducible via the vendor's Docker recipe; numbers are peak-serving, not latency-oriented.
Notes
Optimum-Habana run_generation.py path, use_hpu_graphs + flash_attention_recompute. Reproducible via the vendor's Docker recipe; numbers are peak-serving, not latency-oriented.