Advanced Micro Devices Inc. today announced that it has acquired Nod.ai, a startup that develops open-source software for speeding up artificial intelligence models, for undisclosed financial terms.

Nod.ai, officially Nod Inc., is the developer of an open-source AI tool called SHARK. It’s a so-called runtime system designed to increase the inference speed of neural networks. According to Nod.ai, SHARK can significantly outperform other open-source tools designed for the same task.

Programs such as AI models can’t be deployed on their own, but rather must be bundled with various auxiliary software components. Many of a program’s auxiliary components are packaged into an application known as a runtime system. Nod.ai’s SHARK is a runtime system optimized specifically for the requirements of neural networks.

In a blog post last year, the company said SHARK enables AI models to run more than three times faster than PyTorch and Torchscript. PyTorch is a popular framework for building neural networks. Torchscript is a complementary technology used to optimize neural networks’ inference speed. 

Nod.ai says that SHARK can also provide better performance than XLA. The latter technology is used to improve the performance of AI models created with TensorFlow, a popular PyTorch alternative. 

Nod.ai not only maintains SHARK but also contributes to a number of other open-source projects. Those projects are likewise designed to make AI applications’ code faster. 

Developers turn their source code files into a functioning program using a tool known as a compiler. The compiler takes the code and turns it into a so-called intermediate representation, which is a kind of abstract description of how a program works. It then turns this abstract description into instructions that a computer’s processor can run.

Torch-MLIR, one of the open-source projects to which Nod.ai contributes, is an intermediate representation specifically geared toward AI use cases. The company also contributes to IREE, which is a compiler that shares many technical components with Torch-MLIR. Developers can use IREE to turn the source code of a neural network into a functioning program capable of running on graphics cards and other AI-optimized chips.

AMD is actively working to increase its share of the AI chip market. The company’s Epyc central processing units ship with built-in machine learning optimizations. AMD also sells specialized AI processors such as the recently introduced MI300X, which is designed to compete with Nvidia Corp.’s market-leading graphics cards.

Enterprises developing AI software require not only fast chips but also a simple way to deploy their neural networks on those chips. The acquisition of Nod.ai could enable AMD to address that requirement more effectively. Optimizing neural networks’ performance, the task that Nod.ai’s SHARK tool eases, is one of the most complicated tasks involved in deploying AI software.

“The addition of the talented Nod.ai team accelerates our ability to advance open-source compiler technology and enable portable, high-performance AI solutions across the AMD product portfolio,” said Vamsi Boppana, senior vice president of AMD’s Artificial Intelligence Group. “Nod.ai’s technologies are already widely deployed in the cloud, at the edge and across a broad range of end point devices today.”

It’s possible AMD will seek to enhance Nod.ai’s SHARK tool following the acquisition. The chipmaker could, for example, add optimizations that improve the performance of SHARK-powered AI models deployed on its processors.

AMD’s rivals are also involved in the open-source ecosystem. Intel Corp., for example, is one of the largest corporate code contributors to Linux and Chromium, the open-source project on which Google Chrome is based.

At the go-to-market level, the Nod.ai acquisition could help AMD make its AI accelerators more competitive with Nvidia’s graphics cards. Alongside its chips, the latter company provides a software platform called Nvidia Enterprise AI that makes it easier for customers to optimize their neural networks’ performance. The platform also includes prepackaged AI models, scientific computing software and a range of other tools. 

Image: AMD