Powered by Indico
v3.2.8
Silicon photonics holds significant promise in achieving ultra-low latency processing and data movement.
Salience Labs is developing one such solution. They have designed a Photonic Tensor Processing Unit that achieves orders of magnitude performance improvement in low precision matrix maths. This can be leveraged for AI inference, signal processing and other matrix-maths heavy compute workloads. Salience’s electronic/optical approach delivers a 60x reduction in latency for full AI inference workloads. This is made possible by Salience’s unique amplitude-based approach to photonic computing that allows for a chip that clocks at 10’s of GHz. It has been developed to integrate with commonly used software stacks. The technology is based on research from University of Oxford and Münster University. The Salience team has built a prototype that runs neural networks and is currently designing a fully integrated hybrid electronic/optical chip for which they have partnered with production level foundries. Additional information can be found in this nature paper: https://www.nature.com/articles/s41586-020-03070-1. This talk will provide a technical overview of the technology and plans for development of customer technology demonstrators. Salience is very interested in understanding the unique needs of the DOE Labs, especially in applications that can take advantage of several orders of magnitude of reduced latency.
Please see talk recording here: