model_name: AIDE-Chip-Surrogates
license: cc-by-nc-sa-4.0
library_name: xgboost
pipeline_tag: tabular-regression
tags:
- computer-architecture
- gem5
- cache
- surrogate-model
- explainable-ai
- shap
- monotonic-constraints
- systems-ml
datasets:
- uralstech/AIDE-Chip-15K-gem5-Sims
AIDE Chip Surrogates
This is a collection of physics-aware, monotonicity-constrained XGBoost models that replace expensive gem5 cache simulations during design-space exploration.
Each model predicts either IPC or L2 miss rate for a specific workload, using only cache configuration parameters as input. The models are interpretable via SHAP and enforce microarchitectural monotonicity where physically justified.
This model release accompanies the paper:
Udayshankar Ravikumar . Fast, Explainable Surrogate Models for gem5 Cache Design Space Exploration. Authorea. January 14, 2026. https://doi.org/10.22541/au.176843174.46109183/v1
Model Architecture
- Algorithm: XGBoost Regressor
- Targets:
- IPC
- L2 miss rate
- Features:
- Logâ‚‚ cache sizes & associativities
- Set-count proxies
- Cache hierarchy ratios
- Constraints:
- Monotonic constraints encoding cache physics
- Selective relaxation for latency-sensitive workloads
Available Models
| Workload | IPC Model | L2 Miss Model |
|---|---|---|
| crc32 | ✔ | ✔ |
| dijkstra | ✔ | ✔ |
| fft | ✔ | ✔ |
| matrix_mul | ✔ | ✔ |
| qsort | ✔ | ✔ |
| sha | ✔ | ✔ |
Total models: 12
Performance
- Test set accuracy: R² ≈ 0.999
- OOD validation:
- 26 unseen cache configurations
- ~817× critical-path speedup
- Low absolute error even when R² is unstable
Explainability
Each model is uploaded with its SHAP summary plot. They confirm:
- Cache sizes dominate IPC & miss behavior
- Associativity effects are workload-dependent
- Learned relationships align with microarchitectural intuition
Intended Use
- Architecture research
- Design-space exploration
- Educational use
- Explainable systems ML
Not for commercial deployment without separate licensing.
Limitations
- Single-core, single-thread models
- Cache hierarchy only (no pipeline, prefetcher, or multicore effects)
- Accuracy depends on training coverage; extreme OOD configs are flagged
Patent Notice
The models uploaded here implement techniques described in an accompanying research paper. The author has filed a pending patent application that may cover broader design-space exploration workflows beyond these specific model implementations.
The open-source license (CC BY-NC-SA 4.0) governs use of these models. This notice is informational only.
Pickle Model Inference
Environment confirmed to be able to load the pickled models:
> python --version
Python 3.13.7
> pip list
Package Version
----------------- -----------
cloudpickle 3.1.2
colorama 0.4.6
contourpy 1.3.3
cycler 0.12.1
fonttools 4.61.1
joblib 1.5.3
kiwisolver 1.4.9
llvmlite 0.46.0
matplotlib 3.10.8
numba 0.63.1
numpy 2.3.5
packaging 25.0
pandas 2.3.3
pillow 12.1.0
pip 25.2
pyparsing 3.3.1
python-dateutil 2.9.0.post0
pytz 2025.2
scikit-learn 1.8.0
scipy 1.16.3
shap 0.50.0
six 1.17.0
slicer 0.0.8
threadpoolctl 3.6.0
tqdm 4.67.1
typing_extensions 4.15.0
tzdata 2025.3
xgboost 3.2.0