FADiff: Optimizing DNN Scheduling on Tensor Accelerators with Fusion-Aware Differentiable Optimization

This research explores differentiable optimization techniques for DNN scheduling, specifically targeting tensor accelerators. The paper's contribution lies in the fusion-aware aspect, likely improving performance by optimizing operator fusion.

• •Addresses the challenge of efficient DNN scheduling on specialized hardware.
• •Employs differentiable optimization to achieve improved performance.
• •Incorporates fusion awareness for potentially more optimized execution plans.