Argonne: A100 40GB SXM x1, llama.cpp CUDA, Llama-2-7B FP16, 128io, b1
58.9 tok/s aggregate (batch*(in+out)/latency)
Fields
id
benchmarks/argonne-a100sxm40-llamacpp-llama2-7b-fp16-b1-128io-tpsMetric
tps_aggregate
Value
58.9
Unit
tok/s aggregate (batch*(in+out)/latency)
Model
meta-llama/Llama-2-7b-hf
Methodology
THIRD-PARTY MEASURED (Argonne LLM-Inference-Bench, arXiv:2411.00136, All_results.csv). llama.cpp (ggml CUDA backend), FP16/BF16, 128 in / 128 out. Throughput = batch*(in+out)/end-to-end latency (paper Eq.2) = AGGREGATE over the batch, NOT per-request decode. Batch 1, 1x A100-40GB-SXM4. Latency 4.345s.
Measured by
third_party
Reproducible
yes
notes
llama.cpp FP16 ~2.9x slower than vLLM at identical bs1. CSV column 'Nvidia A100 GPU' = ALCF ThetaGPU node (4x A100 40GB SXM4 + NVSwitch, paper Table II). Filed under nvidia-a100-40gb-sxm4 ONLY; the PCIe 40GB SKU shares the 1555 GB/s HBM2 but was not the measured part, so nvidia-a100-40gb-pcie stays uncovered.
Notes
llama.cpp FP16 ~2.9x slower than vLLM at identical bs1. CSV column 'Nvidia A100 GPU' = ALCF ThetaGPU node (4x A100 40GB SXM4 + NVSwitch, paper Table II). Filed under nvidia-a100-40gb-sxm4 ONLY; the PCIe 40GB SKU shares the 1555 GB/s HBM2 but was not the measured part, so nvidia-a100-40gb-pcie stays uncovered.