Inference energy modeling for BERT models
Bibliographic record
Abstract
BERT achieved great success in Natural Language Processing (NLP).BERT models are computationally expensive due to their model size with hundreds of millions of parameters making them impractical to deploy to resource-constrained hardware platforms such as mobile devices.One way to optimize these complex models is to utilize hardware-aware neural architecture search (NAS).However, hardware-aware NAS needs to incorporate hardware performance metrics, such as energy consumption.Although on-device measurements provide accurate feedback, the overhead is huge.To address this problem, we propose an energy modelling framework on the Hikey 970 ARM big.LITTLE CPU and Mali GPU to predict the inference energy consumption of BERT models.We train energy predictor models on BERT and MobileBERT design space.We evaluate our BERT energy predictor's ability to generalize to the DistilBERT and DynaBERT design space Combined with the real accuracy values on the QNLI task, we evaluate our energy model's ability to predict Pareto-optimal front in the 2D accuracy-energy design space of DynaBERT models.Our model correctly predicts 14 out of the 17 true Pareto-optimal models.List of Tables xi 5.4 DistilBERT on CPU big cluster: The performance of predictor in terms of MAPE, percentage of models in 5%, 10%, 20% error band . . . . . . . . . .54 5.5 DistilBERT on CPU LITTLE cluster: The performance of predictor in terms of MAPE, percentage of models in 5%, 10%, 20% error band . . . . . . . . .55 5.6 MobileBERT on CPU big cluster: The performance of predictor in terms of MAPE, percentage of models in 5%, 10%, 20% error band . . . . . . . . . .56 5.7 MobileBERT on CPU LITTLE cluster: The performance of predictor in terms of MAPE, percentage of models in 5%, 10%, 20% error band . . . . . . . . .56 6.1 DynaBERT design space . . . . . . . . .
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.014 |
| Meta-epidemiology (narrow) | 0.001 | 0.002 |
| Meta-epidemiology (broad) | 0.002 | 0.002 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.010 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".