ClearML today released a set of software tools designed to help companies use their artificial intelligence infrastructure more efficiently.

San Francisco-based ClearML, officially Allegro Artificial Intelligence Ltd, is backed by $11 million in funding. It provides a software platform that developers can use to train AI models, deploy them in production and manage the data they process. ClearML says its platform is used by more than 1,600 organizations, including Intel Corp. and other major tech firms.

The first new addition to the company’s product portfolio is an open-source tool dubbed Fractional GPU. According to ClearML, the software allows multiple AI workloads to run side-by-side on a single Nvidia Corp. graphics processing unit.

Many AI models don’t use the full compute capacity of the GPU on which they run. As a result, a part of the chip’s processing power is left unused, which can prevent companies from making the most out of their hardware investments. ClearML says that its new Fractional GPU tool makes unused GPU resources available for applications and thereby allows companies to more efficiently operate their AI infrastructure.

Under the hood, the software is powered by two features that Nvidia ships with its chips: Multi-Instance GPU and time slicing.

Multi-Instance GPU makes it possible to split a graphics card into multiple sections that each function as a separate processor. Each such virtual processor is assigned its own memory pool and on-chip cache. If the software running on one of the GPU sections experiences a technical issue, the others can continue processing data without interruption. 

The second Nvidia feature ClearML’s Fractional GPU tool uses is called time slicing. It’s designed for the same task as Multi-Instance GPU, but can split a graphics card into a larger number of sections. The tradeoff is that those sections are isolated to a lesser degree, which means a software malfunction in one of them can potentially interrupt processing across the others.

According to ClearML, using an Nvidia chip’s built-in software features to run multiple workloads side-by-side typically requires customers to deploy Kubernetes. Its Fractional GPU tool removes that requirement. Additionally, the tool makes it possible to create multiple virtual processors in consumer-grade Nvidia chips that don’t include the chipmaker’s Multi-Instance GPU feature.

“ClearML is democratizing access to compute as part of our commitment to help our community build better AI at any scale, faster,” said ClearML co-founder and Chief Executive Officer Moses Guttmann. “We hope that organizations that might have a mixture of infrastructure are able to use ClearML and get more out of the compute and resources they already have.”

The company debuted Fractional GPU today alongside an update to its flagship AI development platform. ClearML has added a tool that administrators can use to manage developers’ access to their company’s GPU environment. Additionally, a new observability dashboard will make it easier to monitor AI models deployed in production for technical issues.

Image: Unsplash