Machine learning acceleration startup OctoML Inc. has secured yet more funding with a new $28 million round that brings its total amount raised to $47 million.

Today’s Series B round was led by Addition, with participation from existing investors Madrona Venture Group and Amplify Partners. It follows a $15 million Series A round in March 2020 that came just six months after its initial seed funding round.

OctoML was founded in 2019 by a group of computer scientists from the University of Washington. Led by Chief Executive Officer Luis Ceze, the company has created an open-source framework called Apache TVM that can be used to optimize artificial intelligence models for the specific processors they’re designed to run on.

The Apache TVM framework is said to provide quite a boost, improving the performance of neural networks by a factor of 30 or more in some cases. It works by fine-tuning AI models to run on specific chip architectures, and supports all manner of standard chips, graphics processors and even some specialized AI accelerators. The framework is popular too, with Facebook Inc. using it to help improve the performance of its speech recognition models, for example. Other big-name customers include Microsoft Corp. and Amazon.com Inc.

Although Apache TVM is completely open-source and therefore free for anyone to use, OctoML does offer a paid version of that tool called Octomizer (below). According to the company, Octomizer is a more user-friendly version of the framework that developers can use to optimize their models just by uploading their code and specifying which chips they want to optimize it to run on.

OctoML says that AI model optimization is a cumbersome task when done manually, and that because of this, about 91% of AI models that developers build never make it into production.

“This is because improving model performance without sacrificing accuracy requires endless manual optimizations and fine tuning, especially given the growing stack of ML software and hardware backends,” said Ceze.

So OctoML’s pitch is that Octomizer can help to make AI development much more efficient. The platform supports dozens of popular ML frameworks including PyTorch, TensorFlow, and ONNX serialized models, as well as hardware backends like nVidia/CUDA, x86, AMD, ARM, Intel, MIPS. Using Octomizer, companies will be able to “extract full value and efficiency from their CPU and GPU investments,” Ceze said.

OctoML said its Octomizer tool is available now via its early access program.

Image: OctoML