Zesty launches automated fractional GPUs, enabling native NVIDIA CUDA GPU partitioning automation.
With NVIDIA open-sourcing the Run:ai platform, Zesty’s new autonomous capabilities can continuously adjust GPU memory allocation, compute sharing, and workload placement in response to changing demands, enabling each GPU to run more workloads.
San Mateo, California, October 15 /PRNewswire/ — Zesty, a Kubernetes optimization company, today announced support for NVIDIA’s CUDA-native GPU partitioning. Following NVIDIA’s open-sourcing of Run:ai as part of the NVIDIA DSX AI Factory Platform, Zesty’s Automated Fractional GPUs now add a layer of autonomous decision-making that continuously determines GPU memory allocation and compute sharing patterns, and adjusts workload placement in response to fluctuations in AI workload demands. Zesty says that with these autonomous capabilities, teams can increase GPU utilization, run more workloads on existing infrastructure, reduce the need for additional GPU capacity, and lower infrastructure costs while protecting application performance.
Zesty stated that 66% of organizations hosting generative AI models run some or all of their inference workloads on Kubernetes, and the GPU capacity required for these workloads is expensive and often limited. In the past, teams typically allocated entire GPUs to single workloads or shared them via time-slicing, NVIDIA Multi-Process Service (MPS), or NVIDIA Multi-Instance GPU (MIG), which divides each GPU into fixed hardware partitions. Zesty noted that these allocation methods cannot keep pace with dynamic AI workloads and real-time demand changes, resulting in idle GPU memory and compute power—while teams continue purchasing additional GPUs despite existing unused capacity.
Zesty noted that NVIDIA’s recently open-sourced Run:ai introduces native CUDA GPU partitioning into Kubernetes via the GPU Fractioning Operator, which enforces GPU memory limits per workload and works in coordination with the Kubernetes scheduler. Zesty pointed out that while the partitioning mechanism enables flexibility by defining how a single GPU is divided among workloads, it does not determine how these capacities are allocated. Deciding which workloads share each GPU, how much memory each workload receives, and how these allocations adapt as demand rises or falls requires an additional layer of decision-making.
Zesty provides exactly this layer. Its Automated Fractional GPU feature continuously determines which workloads should share the same GPU, how many workloads can safely run simultaneously, how much GPU memory each workload should receive, which sharing mode is best suited for each workload, and when these allocations should be increased or decreased. These decisions are based on actual GPU memory usage and utilization, observed workload behavior over time, and the team’s predefined latency requirements and resource constraints. All changes remain within the performance and availability policies set by the team for each workload.
Pini Ben Nahum, Vice President of Engineering at Zesty, said: “GPUs are among the most expensive and tightly constrained resources managed by today’s infrastructure teams. They are difficult to secure, costly to scale, and shared among AI workloads with constantly shifting demands. Now that NVIDIA has opened up GPU partitioning to the ecosystem, Zesty is making this allocation intelligent and autonomous.”
He said: "Autonomous GPU Resource Management adjusts GPU memory and compute sharing in response to changes in workload demand and guides workload placement. Teams can fit more workloads onto existing hardware without compromising performance quality, reducing the need for new GPU purchases and lowering the cost of running AI in production environments."
This feature is deployed as a lightweight proxy within a Kubernetes cluster and works in conjunction with the NVIDIA Run:ai GPU Fractioning Operator and the Kubernetes scheduler to support time-slicing and MPS. It requires no modifications to application code or models and does not restart running workloads during installation. On supported GPU stacks, Zesty manages the necessary Kubernetes configurations without requiring manual changes to manifest files.
Automated Fractional GPUs are now available as part of the Zesty platform. For more information, visit http://zesty.co/platform/automated-fractional-gpus.
About Zesty
Zesty helps businesses reduce infrastructure costs by autonomously optimizing resources. The platform continuously aligns resources with real-time workload demands, automatically eliminating overprovisioning and improving infrastructure efficiency while safeguarding application performance. By replacing manual, ongoing resource adjustments and performance tuning with policy-driven automation, Zesty enables teams to run leaner Kubernetes environments without increasing operational risk.
With Automated Fractional GPUs, Zesty is extending this approach to GPU infrastructure, helping teams extract more value from existing GPU capacity while managing the rising costs of AI workloads.
Zesty was founded in 2019, focusing on making infrastructure more efficient, automated, and cost-effective.
For more information, visit zesty.co.
Media contact
Kate Schoenstadt
[email protected]
IL +972 (0) 54-777-6684