Start Date
1-5-2026 12:00 PM
End Date
1-5-2026 1:00 PM
Description
This project explores how to improve the performance of machine learning systems at the kernel level. As models grow in size, efficient execution depends on reducing data movement and using limited on-chip memory effectively.
Techniques such as tiling and optimized execution order are used to improve data locality and reduce memory overhead. These approaches enable better use of fast on-chip memory and minimize costly data transfers.
Using benchmark workloads, improved scheduling strategies are shown to reduce latency by avoiding suboptimal local decisions and enabling better global execution patterns.
Overall, this work demonstrates that kernel-level optimization can significantly improve performance and efficiency in modern machine learning workloads.
Optimizing Machine Learning Workloads Performance in ML accelerators
This project explores how to improve the performance of machine learning systems at the kernel level. As models grow in size, efficient execution depends on reducing data movement and using limited on-chip memory effectively.
Techniques such as tiling and optimized execution order are used to improve data locality and reduce memory overhead. These approaches enable better use of fast on-chip memory and minimize costly data transfers.
Using benchmark workloads, improved scheduling strategies are shown to reduce latency by avoiding suboptimal local decisions and enabling better global execution patterns.
Overall, this work demonstrates that kernel-level optimization can significantly improve performance and efficiency in modern machine learning workloads.