Research & Papers

Terastal framework cuts DNN deadline misses by 40% with layer variants

New scheduling algorithm reduces missed deadlines by 40.58% while keeping accuracy loss under 2.24%

Deep Dive

Terastal is a soft real-time framework for scheduling multiple DNNs on heterogeneous accelerators. It uses "

Key Points
  • Terastal reduces deadline miss rate by 40.58% vs FCFS, 30.53% vs EDF, and 36.27% vs DREAM
  • Uses 'layer variants' — custom layer implementations that trade up to 2.24% accuracy for lower latency on non-preferred accelerators
  • Combines offline virtual budget assignment with online scheduling for joint accelerator-mapping and variant selection

Why It Matters

Enables real-time multi-DNN workloads on edge devices with tighter deadlines and minimal accuracy loss.

📬 Get the top 10 AI stories daily