OctoML today announced the close of a $3.9 million funding round to grow its engineering team and prepare for the release of a SaaS service for machine learning engineers. The service aims to automate machine learning optimization with the open source Apache TVM software stack. TVM is used today by companies like Facebook, Amazon, and Microsoft to run models efficiently in the cloud or on edge hardware.
Madrona Venture Group led the funding round, with participation from Amplify Partners. OctoML cofounder and CEO Luis Ceze is also a part-time general partner at Madrona Venture Group and heads its technical advisory board.
TVM was created four years ago as part of a collaboration between University of Washington colleagues like Ceze and Carlos Guestrin. It was developed to address the problem of mapping machine learning models to cloud and edge hardware like CPUs, GPUs, or FPGA accelerators. TVM was made part of the Apache Software Foundation in March and counts more than 250 contributors.
“The key magic in TVM is to use machine learning to optimize machine learning code,” Ceze told VentureBeat in a phone interview. “Once you have a model that you’re ready to run … you want to bridge the gap between what the model wants to do and what the hardware can do and then make that as efficient as possible.”
TVM uses machine learning to automate the optimization of machine learning models after being trained with a framework like TensorFlow or PyTorch.
“There’s quite a bit of work that needs to be done once you have a model to make sure that our model runs well on a specific hardware target,” Ceze said.
Data collected during optimization experiments to discover efficient mappings between models and hardware is used to train TVM’s machine learning and improve model optimization.
OctoML was founded in July 2019 and is based in Seattle, with 10 employees.
Other AI startups that received funding for machine learning optimization include Ople, and Determined AI, which raised an $11 million funding round led by GV in March.