# GPU usage — Amazon 100K persona extraction (reference run)
Charged hours from the batch scheduler's accounting records. Allocation fully spent.
| GPU | Precision / parallelism | Charged GPU-hours | |---|---|---:| | NVIDIA H100 80GB (HBM3) | FP8, TP=1 (1 model / GPU) | 645.6 | | NVIDIA A100 40GB (SXM4) | BF16, TP=4 (1 model / 4-GPU node) | 1,037.7 |
The A100 figure is 257.1 node-hours × 4 GPUs/node (GPU nodes are charged whole).
The two lines are separate allocations on different hardware; they are reported separately and not summed.
Model Qwen/Qwen3.6-35B-A3B on vLLM ≥ 0.24.0, greedy (temp=0).
Output: 38,219 unique users (38.2% of the 100K target).