KubeCon + CloudNativeCon North America 2024

Kubernetes is the ideal platform for running AI and ML workloads, such as LLMs. GPU nodes are often used for their parallel processing capabilities and higher performance benefits; however, they are known to be costly. Many factors impact the cost of running AI/ML workloads such as GPU utilization, GPU VM size, idle time, etc. These costs are often ignored and considered inherent in running GPU workloads. But if running workloads at scale and left unoptimized, costs will quickly spin out of control. In this talk, we leverage NVIDIA DCGM exporter with Prometheus for GPU metrics monitoring alongside OpenCost to measure the Kubernetes spend of our GPU workloads. We will provide an overview of OpenCost, highlighting its role in bridging the gap between the developer and platform teams through visibility and accountability of spend. We will demonstrate how to use the NVIDIA GPU Operator and how techniques such as partitioning can lead to significant cost savings.