I am awarded the NWO Open Competition KLEIN-1 grant for € 350,000 to explore the research direction for building efficient storage support for large-model ML workloads. Here is the public summary:
Title: Machine Learning from Storage: Democratizing Machine Learning for All
Public Summary: In machine learning, large models are associated with more accuracy and intelligence. Building and training of large models processes large amounts of data in CPU-attached and accelerator on-board DRAM memories, which are not scalable, energy-inefficient, and extremely expensive, thus putting large model training out of reach for many users. We propose to leverage Non-Volatile Memory (NVM) storage technology to design a Machine-learning-from-Storage model training paradigm, where data is dynamically moved between NVM storage and DRAM memories on-demand, thus unleashing a new class of previously-unavailable cost- and energy-efficiency for all.
The project is expected to start in the second half of 2021.
Open PhD position: I am looking for a motivated PhD student to start work on this. See application details https://workingat.vu.nl/ad/phd-position-in-storage-support-for-machine-learning/w3s7ih.