Research map / Architectures and efficiency / RNNs and normalization
Faster Activation Functions at the Edge for Post-Training Speedups
Anton Lydike, Jun Bi, Jackson Woodruff
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace exact SiLU with FFCC FastSiLU post-training | trained models · 8 models (DenseNet, MobileNetV2, ResNet-18, ShuffleNetV2, SqueezeNet, MLPerf Tiny) | accuracy | No clear effect |
| replace exact SiLU with FastSiLU post-training | MLPerf Tiny Anomaly Detection autoencoder · MLPerf Tiny Anomaly Detection | task metric: 91.94 -> nan | Hurt |
| train with Hardswish instead of SiLU | MobileNetV2, DenseNet-121 · image classification | accuracy: 74.0 -> 72.1 (DenseNet-121) | Hurt |
| replace exact GELU with FastGELU in pretrained ConvNeXt | ConvNeXt-Tiny · ImageNet | Top-1: 82.1 -> 82.0 | No clear effect |
| replace exact GELU with FastGELU in pretrained ConvNeXt | ConvNeXt-Femto · ImageNet | Top-1: 77.4 -> 77.2 | Hurt |
| tune constants holistically over the whole expression instead of exponentiation alone (ablation) | FFCC FastGELU · ConvNeXt ImageNet | Top-1 accuracy: +0.1 (Tiny) to +0.3 pp (Femto) | Improved |
| approximate SiLU with FFCC instead of exact SiLU | Arm Cortex-M7 · activation microbenchmark | cycles per activation: 138.2 -> 21.1 | Improved |
| approximate activations with FFCC | Snapdragon and Intel chips · activation microbenchmark | wall time: 15-22x speedup | Improved |
| replace exact GELU with FFCC approximation in ConvNeXt-Femto | ConvNeXt-Femto · end-to-end per-stage latency | speedup: 1.2x to 1.37x | Improved |
| replace exact GELU with FFCC approximation in ConvNeXt-Tiny | ConvNeXt-Tiny · end-to-end per-stage latency | speedup: 1.19x to 1.37x in Stage 1 | Mixed |
| subtract small offset epsilon from exponent bias (ablation) | bit-cast exp2 approximation · approximation error | error | Improved |
Closest papers by meaning
- FALQON: Accelerating LoRA Fine-tuning with Low-Bit Floating-Point Arithmetic (NeurIPS 2025 · similarity 0.84)
- Stepping Forward on the Last Mile (NeurIPS 2024 · similarity 0.84)
- Analog Foundation Models (NeurIPS 2025 · similarity 0.84)
- Agile-Quant: Activation-Guided Quantization for Faster Inference of LLMs on the Edge (AAAI 2024 · similarity 0.84)
- Scaling FP8 training to trillion-token LLMs (ICLR 2025 · similarity 0.84)
- Pushing the Limits of BFP on Narrow Precision LLM Inference (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
