Rating Standard

NPUBench Standard v1.0 · Section 7 · Performance rating thresholds based on real hardware measurements

🚧 Results Coming Soon

Leaderboard Under Development

The benchmark tool is currently in development. Official community leaderboard will be published following the public release. The rating thresholds below are defined in NPUBench Standard v1.0.

Performance Rating Tables

All thresholds derived from real hardware measurements — average of 5 runs on Llama 3.2 3B Instruct (int4), translation task

TTFT Lower is Better
Rating Threshold Reference
★★★ < 0.1 s Qualcomm Snapdragon X NPU: 0.029 s
★★ 0.1 ~ 2.0 s AMD Ryzen AI 300 NPU: 1.29 s
★ 2.0 ~ 5.0 s Intel Core Ultra Series 2 NPU: 2.91 s
✗ > 5.0 s CPU baseline
TPS Higher is Better
Rating Threshold Reference
★★★ > 40 t/s Far exceeds reading speed
★★ 20 ~ 40 t/s Exceeds reading speed
★ 10 ~ 20 t/s Comparable to reading speed
✗ < 10 t/s Below reading speed (~4–5 t/s ref)
NPU Acceleration Ratio TTFT Higher is Better
Rating Threshold Reference
★★★ > 50 × Qualcomm Snapdragon X: 314 ×
★★ 10 ~ 50 × Intel Core Ultra Series 2: 26 ×
★ 5 ~ 10 × AMD Ryzen AI 300: 7 ×
✗ < 5 × NPU acceleration ineffective
Power Efficiency NPU Power Higher TPS/W is Better
Rating Threshold Reference
★★★ NPU < 6 W Qualcomm / Intel NPU: 3–5 W
★★ 6 ~ 15 W AMD Ryzen AI 300 NPU: 5–10 W
★ 15 ~ 30 W CPU lower bound
✗ > 30 W CPU upper bound
★★★ = Excellent · ★★ = Good · ★ = Acceptable · ✗ = Poor | Source: NPUBench Standard v1.0 Section 7

About the Rating System

The NPUBench rating system evaluates four independent dimensions: TTFT (Prefill phase latency), TPS (Decode phase generation rate), NPU Acceleration Ratio (NPU vs CPU speedup), and Power Efficiency (TPS per watt). Each dimension is rated independently. A device may excel in one metric while being acceptable in another. Thresholds are fixed in NPUBench Standard v1.0 and will not change between versions to ensure cross-version comparability.

Legal Notice
NPUBench Standard v1.0 · Apache 2.0 License · Not affiliated with MLCommons · MLPerf® is a registered trademark of MLCommons Association