Task-Conditional Compute Skipping for Multi-Task Inference Accelerators
A new hardware-software co-design approach enables task-conditional compute skipping in multi-task inference accelerators. The method uses a lightweight gating network, trained jointly with the backbone, to predict per-tile binary execution masks based on the task command. This allows masked tiles to be skipped with zero overhead, reducing energy and cycles by activating only the subset of the network required for each task. The approach does not change the model architecture or inference pipeline. The paper is available on arXiv with ID 2607.22038.
Key facts
- Multi-task inference models execute identical computation regardless of the active task.
- Task command provides a free signal to skip unnecessary computation at hardware level.
- Lightweight gating network predicts per-tile binary execution masks conditioned on task input.
- Each tile corresponds to a fixed group of output channels (native scheduling granularity).
- Masked tiles are skipped with zero overhead.
- Task-dependent reduction in compute without changes to model architecture or inference pipeline.
- Full system co-design is described in the paper.
- arXiv ID: 2607.22038
Entities
Institutions
- arXiv