Hardware Accelerated Similarity Search

Hardware Accelerated Similarity Search

Standard databases are composed of structured tables containing symbolic information.

The latest deep learning tools generate non-symbolic, high-dimensional feature vectors (x2vec) for images, video, text, music, speech, molecules, and many other objects. These representations are a more powerful method for achieving recognition accuracy, and more flexible than fixed symbolic representations.

However, searching for similar, non-symbolic, objects within billions of records – each with a high-dimensional representation – presents a computational challenge.   

GSI technology has developed a breakthrough, patented technology: An in-memory associative processing chip based on in-place computing. It is capable of performing similarity search on billions of  high-dimensional records, with very low latency (< 2 msec) and a high throughput of over 100,000 queries per second. The technology works with standard software frameworks such as TensorFlow.

The presentation will cover the technology, the product, and its performance.

Schedule
Room
Room C

Tracks:

Architecture
Speakers
Profile picture for user george.williams
George
Williams
Director of Data Science
at
George is Director of Computing and Data Science at GSI Technology, an embedded hardware and artificial intelligence company.  He's held senior leadership roles in software, data science, and research, including tenures at Apple's New Product Architecture group and at New York University's Courant Institute.  He can talk on a broad range of topics at the intersection of data science, e-commerce, machine learning, software development, and cybersecurity.  He is an author on several patents and research papers in computer vision and deep learning, including publications at NIPS, CVPR, ICASSP, and SIGGRAPH.

Slides & Recordings

   Download Slides