Zesty Automates GPU Resource Sharing for AI Workloads
Zesty has launched a new automated feature that dynamically optimizes NVIDIA GPU resource allocation for AI workloads, enhancing utilization and reducing costs.
SAN MATEO, Calif. – Zesty, a company specializing in Kubernetes optimization, has released a new automated capability for dynamically fractioning NVIDIA CUDA-native GPUs. This feature, building on NVIDIA's recent open-source release of the Run:ai platform, aims to improve the performance of AI workloads and optimize the utilization of GPU resources.
The new automation system continuously adapts GPU memory allocation, compute sharing, and workload placement based on changing demands. Its goal is to enable more workloads to run on existing GPUs, reduce the need for additional GPU capacity, and lower infrastructure costs without compromising application performance.
Traditionally, GPUs have either been allocated entirely to individual workloads or shared into fixed partitions, failing to keep pace with the dynamic requirements of AI tasks. This has led to resource wastage and the need to procure more GPUs, even when existing capacity remains underutilized.
NVIDIA's open-source release of Run:ai facilitates GPU sharing within Kubernetes environments. Zesty's automation complements this by providing a decision-making layer that optimizes resource usage based on real-time data and predefined policies.
Zesty's "Automated Fractional GPUs" is now available as part of the company's platform. The solution deploys as a lightweight agent within the Kubernetes cluster and requires no modifications to existing applications or models.