Advanced Micro Devices Inc. said today it’s collaborating with a startup called Rapt AI Inc. to improve artificial intelligence training and inference performance on its Instinct brand of graphics processing units.

Rapt AI is the creator of an intelligent platform that uses AI smarts to automate workload management on high-end GPUs, helping to maximize performance and scale, simplify application deployment and reduce the cost overhead of AI applications.

According to the companies, many enterprises are struggling to get a handle on their AI applications. The challenge stems from the fact that customers must rely on huge clusters of GPUs to support their most complex workloads, but many struggle to manage these resources effectively. As such, there’s an urgent need for more efficient resource allocation to avoid performance bottlenecks for GPU workloads.

“As more organizations move to production AI, maximizing infrastructure efficiency and cost effectiveness becomes paramount,” said Rapt AI Chief Technology Officer Anil Ravindranath.

Rapt AI’s software is designed to work with AMD Instinct accelerators such as the MI300X, MI325X and the upcoming MI350 GPUs, which are alternatives to Nvidia Corp.’s better known H100, H200 and new “Blackwell” AI accelerators.

By using Rapt AI’s automation software to intelligently manage fleets of AMD GPUs, companies can expect to squeeze the maximum performance out of their silicon for any kind of AI workload, ensuring they fully utilize those resources to lower the total cost of ownership.

The software also helps to simplify the deployment of AI applications in both on-premises and cloud environments. According to Rapt AI, it allows organizations to save hours of time experimenting with different infrastructure configurations by automatically setting up the most optimal workload balance, even in diverse compute clusters made up of multiple kinds of GPUs.

The result will be improved inference and training performance and increased scalability for production AI deployments, with Rapt AI’s unique auto-scaling software optimizing resource allocation based on application demand.

AMD’s collaboration with Rapt AI means that the software will work perfectly, out-of-the-box, with all AMD Instinct GPUs, helping customers to realize immediate performance benefits with simple deployment. Moreover, the companies plan to collaborate in future to enable further optimizations in areas such as GPU scheduling, memory utilization and more, continuously boosting performance to ensure customers have access to the most optimal and cost-effective AI infrastructure.

Rapt AI Chief Executive Charlie Leeming said that by working more closely with AMD, it can develop more intricate performance optimizations, increasing the benefits for joint customers.

“This joint solution is set to transform AI infrastructure management, driving better performance, cost efficiency and faster time-to-value for our mutual customers,” he promised.

Image: AMD