NVIDIA L40S 48GB llama.cpp cuBLAS: LLaMA-3 8B F16 pp1024 prefill tok/s
2491.65 tok/s (pp1024 avg, per-request batch-1)
Fields
id
benchmarks/xd-nvidia-l40s-8b-f16-pp1024Metric
prefill_tok_s
Value
2,491.65
Unit
tok/s (pp1024 avg, per-request batch-1)
Model
LLaMA-3 8B
Methodology
THIRD-PARTY MEASURED (XiongjieDai GPU-Benchmarks-on-LLM-Inference, RunPod rental). llama.cpp built with cuBLAS (make LLAMA_CUBLAS=1), May-2024 snapshot. NVIDIA L40S 48GB full offload (-ngl 10000); llama-bench -p 512/1024/4096/8192 -n same. Ubuntu 22.04, CUDA 12.1.1. Value is the 1024-token column. Per-request batch-1, NOT aggregate.
Measured by
third_party
Reproducible
yes