Real-Time Machine Learning (RTML) Compiling Hardware Neural-Net Accelerators (CHANNA)
Abstract
The CHANNA project developed an open-source machine-learning accelerator generator (named Gemmini) that supports integration into a full system-on-a-chip design. Additionally, the team completed evaluations of generated accelerators running various machine learning (ML) workloads; demonstrated that the generator can produce competitive accelerators for a wide range of ML tasks; and developed a novel accelerator virtualization mechanism, AuRORA, to enable virtualized and disaggregated accelerator integration for many-accelerator-many-application systems.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2024
- Accession Number
- AD1229977
Entities
People
- Borivoje Nikolic
- Jonathan Ragan-Kelley
- Krste Asanovic
- Sophia Shao
Organizations
- University of California, Berkeley