NPUWattch

NPUWattch is an ML-based power, area, and timing (PAT) modeling tool. Given an accelerator description — and, optionally, activity recorded by a simulator — it produces per-component estimates of energy, area, and timing, collected into a self-contained HTML report. The models are trained on measurements collected through our holistic modeling approach.

Install

pip install -e .

Tutorial

tutorial/ holds two ready-to-run examples with real simulator data — AlexNet on an Eyeriss-like array (Timeloop input) and a 1024³ matmul on a TPUv3-like NPU (PyTorchSim input). Each is one command:

cd tutorial/timeloop   && ./run
cd tutorial/pytorchsim && ./run

Run

There are three ways to run NPUWattch, depending on what you have.

1. From a PyTorchSim run (recommended). NPUWattch reads the simulator's own output folders, figures out what hardware was modeled, and charges each part with the activity the simulator recorded:

npuwattch --harness pytorchsim \
          --togsim-dir togsim_results/ \
          --gem5-dir   gem5_outputs/ \
          --node 7nm --report report_dir/

When both folders live under one run root, ./run.sh <run_root> [flags...] locates them for you.

2. From an Accelergy/Timeloop architecture YAML. Add the mapping's stats file (or a directory of per-layer stats files) to charge the real access counts Timeloop reports; without --stats the result is an activity-free estimate labeled VECTORLESS:

npuwattch --harness timeloop --arch-yaml architecture.yaml \
          --stats timeloop-model.stats.txt --node 7nm

Stats level names bind to the description's components by (leaf) name; the optional --stats-map map.yaml (levels: renames, ignore: deliberate drops) covers the rest. A directory of per-layer stats becomes one report window per layer (--stats-mode aggregate sums them instead).

3. From NPUWattch's own files. A native description, plus an optional activity CSV:

npuwattch -d description.yaml -l activity.csv

Useful flags:

• --report DIR writes report.html + report.json; --tree shows how the run was interpreted.
• Technology and operating point (harness modes only — a native description carries its own): --node, --transistor, --corner, --voltage-offset, --temperature, --clock-mhz.
• The node axis is continuous: characterized nodes (5/7/10/16/20nm) answer directly, anything between is log-interpolated, up to ±50% beyond the range (2.5–30nm) is extrapolated with a WARNING, and anything outside that envelope is clamped to it with a WARNING on the CLI and in the report.
• PyTorchSim extras: --config-yml, --booksim-dir (anynet NoC runs), --energy-table (DRAM cost table, e.g. hbm2.yml).
• Any run without activity data falls back to a vectorless estimate (25% of random switching, tunable with --vectorless-activity) and is labeled accordingly.

Citation

NPUWattch :

@inproceedings{kim_hpca2026,
    title       = {NPUWattch: ML-based Power, Area, and Timing Modeling for Neural Accelerators},
    author      = {Kim, Sehyeon and Kim, Minkwan and Park, Chanho and Park, Hanmok and Kim, Seonghoon and Song, Taigon and Song, William J.},
    booktitle   = {IEEE International Symposium on High-Performance Computer Architecture},
    month       = {Jan.},
    year        = {2026},
    pages       = {1-14},
}

Related technology libraries :

@inproceedings{shin_iscas2024,
    title       = {FS2K: A Forksheet FET Technology Library and a Study of VLSI Prediction for 2nm and Beyond}, 
    author      = {Shin, Yunjeong and Park, Daehyeok and Koh, Dohun and Heo, Dongryul and Park, Jieun and Lee, Hyundong and Kim, Jongbeom and Lee, Hyunsoo and Song, Taigon},
    booktitle   = {IEEE International Symposium on Circuits and Systems}, 
    month       = {May},
    year        = {2024},
    pages       = {1-5},
}

@article{kim_tvlsi2023,
    title       = {NS3K: A 3nm Nanosheet FET Standard Cell Library Development and Its Impact},
    author      = {Kim, Taehak and Jeong, Jaehoon and Woo, Seungmin and Yang, Jeonggyu and Kim, Hyunwoo and Nam, Ahyeon and Lee, Changdong and Seo, Jinmin and Kim, Minji and Ryu, Siwon and Oh, Yoonju and Song, Taigon},
    journal     = {IEEE Transactions on Very Large Scale Integration Systems},
    volume      = {31},
    number      = {2},
    month       = {Feb.},
    year        = {2023},
    pages       = {163-176},
}

Related Links

Technology Libraries from 20nm to 2nm

License

NPUWattch is released under the MIT license. See LICENSE for additional details. Thanks to the I3D VLSI Laboratory.

Questions

Leave github issues or please contact ikamusume@yonsei.ac.kr