TITLE: Industrial-grade benchmarking and optimization of Computer Systems for Machine Learning

ABSTRACT: The field of Machine Learning is teeming with innovation. Computer Systems are unsung heroes of this innovation: from huge datacentres packed with accelerators for training Large Language Models like GPT-3 - to tiny microcontrollers detecting anomalies in the behaviour of industrial equipment. The growing importance of Machine Learning workloads mandates designing Computer Systems that are fit-for-purpose in terms of quality, performance, efficiency, size and cost. But how do we ensure a good fit and sensible use of resources?

In this talk, we will discuss MLPerf, an industry-leading initiative for fair and useful benchmarking of Computer Systems, and its special interest working groups focused on Training, Inference, Mobile, Tiny, Power and other topics. We will also discuss how on-going benchmarking efforts are paving the way to synthesizing complete neuralware/middleware/hardware stacks and optimizing them for given requirements.

BIO: Dr Anton Lokhmotov has been working on designing and optimizing computer systems for over 20 years, as a researcher, engineer and entrepreneur. He is Founder of KRAI, a Cambridge-based computer engineering startup, working with leading ML chip companies ranging from global corporations such as Qualcomm to stealth-mode startups. The team at KRAI has been contributing to the industry-leading MLPerf benchmark since its inception in 2018, having submitted more results to MLPerf Inference than all the other 40+ submitters combined. For example, KRAI have implemented and optimized MLPerf Inference benchmarks on Datacenter servers and Edge appliances equipped with Qualcomm Cloud AI 100 accelerators, demonstrating some of the fastest and most power efficient results in the history of MLPerf.