Hardware Acceleration of Sparse Cognitive Algorithms
Abstract
Hardware accelerators were designed for both the Sparsey and Numenta HTM cortical algorithms. Two versions were designed -- a programmable 65 nm SIMD version with Processor in Memory (PiM) extensions and a 65 nm ASIC version. They were compared against a 28 nm GPU baseline using the KTH video action recognition benchmark. Performance/power improvement over the GPU were (Sparsey) SIMD with PiM: 1490; ASIC: 1300; and (HTM) SIMD with PiM: 537; ASIC: 47,100.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 2016
- Accession Number
- ADA631951
Entities
People
- Joshua Schabel
- Lee Baker
- Paul D Franzon
- Sumon Dey
- Weifu Li
Organizations
- North Carolina State University