Zesty adds automated CUDA-native GPU fractioning for Kubernetes Run:ai

Zesty has added “Automated Fractional GPUs” to its platform, aiming to automate how NVIDIA CUDA-native GPU fractioning gets allocated and adjusted inside Kubernetes clusters. The feature is designed to continuously adapt GPU memory limits, compute-sharing mode, and workload placement as demand changes, with the goal of fitting more workloads per GPU while staying within team-defined performance and availability policies.

The capability builds on NVIDIA’s open-source release of Run:ai, which brings CUDA-native GPU fractioning to Kubernetes via the GPU Fractioning Operator. That operator enforces per-workload GPU memory limits and integrates with Kubernetes schedulers, but it doesn’t decide how to divide a GPU among competing workloads over time. Zesty’s layer is intended to make those decisions automatically: which workloads should share a given GPU, how many can run together, how much GPU memory each workload receives, which sharing mode to use, and when allocations should increase or decrease.

Mechanically, Zesty says it bases those decisions on observed GPU memory usage and utilization, workload behavior over time, and workload latency requirements and resource limits set by teams in advance. It also says changes are constrained by the performance and availability policies operators define per workload. For platform teams trying to keep inference and training clusters busy without constant hand-tuning, the practical value is straightforward: fractioning defines what’s allowed, but automation determines whether it actually gets used efficiently.

Deployment-wise, Zesty describes Automated Fractional GPUs as a lightweight agent that runs in a Kubernetes cluster alongside the NVIDIA Run:ai GPU Fractioning Operator and the Kubernetes scheduler. It supports time-slicing and NVIDIA Multi-Process Service (MPS). Zesty also says it requires no application code or model changes, doesn’t restart running workloads during installation, and manages the Kubernetes configuration required for optimization on a supported GPU stack without manual manifest edits.

“GPUs are among the most expensive and constrained resources infrastructure teams manage today,” said Pini Ben Nahum, VP R&D at Zesty. “Autonomous GPU Resource Management adjusts GPU memory and compute sharing while guiding workload placement as each workload’s needs rise and fall.”

Automated Fractional GPUs is available today as part of the Zesty platform, with more details at zesty.co/platform/automated-fractional-gpus.