Menno Lindwer (Grai Matter Labs)

Less is more: do 20 times more machine learning with 95 percent less compute


We can and should avoid (>>) 95 percent of the operations currently being executed on machine learning. Physics dictates that there’s no magic wand by which we can suddenly become 20-100 times more efficient. Hence, the trick is in actually not doing those operations in the first place. This talk will explain how to avoid at least 95 percent of the compute, memory accesses, data transfers, and so on. In doing so, we can handle complex ML tasks with very low power consumption.

Menno Lindwer is leading GML’s engineering department, which includes developing neuromorphic and neural network processors, programming and learning tools, and AI applications. Prior to joining GML, he was a processor technology architect, an engineering manager and a program manager at Intel. He received his MSc from University of Twente and his PDEng from Eindhoven University of Technology in 1993. Since then, he’s studied asynchronous devices, database matching technology, 3D graphics renderers, media processor design and AI processors. In 1995, he joined Philips Research, working on the design of Java and media processors. In 2002, he was part of the team that started up Silicon Hive. He was the director of product management for Silicon Hive’s Hivelogic parallel processing platform. Silicon Hive was acquired by Intel in February 2011. At Intel, he headed teams that worked on processor construction tools and optimizing auto-targeting compilers. He also led Intel’s activities in several EU projects.

Menno Lindwer

Grai Matter Labs