What is it about?
Florets is a space-filling curve inspired network design which enables accelerating datacenter-scale workloads with up to 2X performance improvement than state-of-the-art counterparts. Florets employ a dataflow aware scheme to connect multiple chiplets (hundreds of smaller cores) on the interposer and scales to higher number with lower fabrication costs and better performance.
Featured Image
Photo by Hal Gatewood on Unsplash
Why is it important?
Floret thematically draws similarity to the brain to be able to serve concurrent CNN tasks with an optimized neural layer to chiplet mapping.
Perspectives
It bridges the gap of scalable chiplet based architectures. As in case of chiplet-based systems the biggest bottleneck is the communication due to relatively farther distance between chiplets. This limits the performance gain while being energy efficient. Floret as an NoI architecture solves that problem with an efficient design inspired by space-filling curves, which are naturally occurring all around us.
Harsh Sharma
Washington State University
Read the Original
This page is a summary of: Florets for Chiplets: Data Flow-aware High-Performance and Energy-efficient Network-on-Interposer for CNN Inference Tasks, ACM Transactions on Embedded Computing Systems, September 2023, ACM (Association for Computing Machinery),
DOI: 10.1145/3608098.
You can read the full text:
Resources
Contributors
The following have contributed to this page
