Views
No views yet
Benchmark validation card — This repository documents the inference performance of Qwen/Qwen2.5-0.5B when optimized with the RunuX-AI runtime on Google TPU v5e.
| Metric | PyTorch (torch_xla) | RunuX-AI | Improvement |
|---|---|---|---|
| Throughput | 328.4 tok/s | 1024.3 tok/s | 3.12× faster |
| Energy | 0.61 J/tok | 0.2 J/tok | 3.12× lower |
| MXU Util | ~32% | 88% | 2.75× higher |
do_sample=False)1# Install baseline framework
2pip install torch torch_xla[tpu] transformers accelerate
3
4# Run baseline benchmark
5python benchmark_baselines.py --models qwen2.5-0.5b1@article{callens2026runux,
2 title={RunuX-AI: Memory-Efficient, Energy-Aware Inference Runtime for Edge and Cloud Accelerators},
3 author={Callens, Xavier},
4 year={2026},
5 note={Socrate AI Lab}
6}