Argonne: PVC x4 llama.cpp Llama-2-70B FP16 128io b1
0.29 tok/s aggregate (batch*(in+out)/latency)
Fields
id
benchmarks/argonne-pvc-llamacpp-llama2-70b-fp16-4x-b1-128io-tpsMetric
tps_aggregate
Value
0.29
Unit
tok/s aggregate (batch*(in+out)/latency)
Model
meta-llama/Llama-2-70b-hf
Methodology
THIRD-PARTY MEASURED (Argonne LLM-Inference-Bench arXiv:2411.00136). llama.cpp (ggml SYCL backend), FP16, 128in/128out, AGGREGATE over batch. 4 PVC tiles, bs1, lat 884.54s.
Measured by
third_party
Reproducible
yes
notes
CSV column 'Intel PVC GPU' = ALCF Aurora/Sunspot Intel Data Center GPU Max (Ponte Vecchio, filed under xpu-intel-dc-gpu-max-1550; the PVC product line maps to the Max series). vLLM xpu backend, 1/2/4-tile configurations. 70B FP16 on 4 PVC tiles is barely-viable; documents the floor.
Notes
CSV column 'Intel PVC GPU' = ALCF Aurora/Sunspot Intel Data Center GPU Max (Ponte Vecchio, filed under xpu-intel-dc-gpu-max-1550; the PVC product line maps to the Max series). vLLM xpu backend, 1/2/4-tile configurations. 70B FP16 on 4 PVC tiles is barely-viable; documents the floor.