NVIDIA launched Minerva on Oct. 22 to identify problems across AI data centers, with capacity to analyze billions of data points an hour, the chip company said.
The company described the platform in its Oct. 22 blog post. The post was titled “NVIDIA Unveils Minerva to Bring End-to-End Observability for AI Factories.” NVIDIA said Minerva combines information that traditional monitoring tools often separate into views for networking, computing and machine learning teams.
NVIDIA said the platform connects readings from graphics processing units, or GPUs, with data from networks, storage, systems that assign computing jobs and the workloads themselves. The company said that combined view helps operators trace a slow training job to the component affecting it.
The announcement described product capabilities and deployment plans, but supplied no measured improvements in productivity, reliability or time spent resolving faults. NVIDIA provided no study period, test dataset or comparison of customer performance before and after adopting Minerva.
NVIDIA said Minerva targets AI factory yield, a measure of useful output from infrastructure used to train models and run AI applications. The company listed GPU hours, useful processing output, time that computing clusters remain operational and production of tokens, the units AI models process, among its measures.
NVIDIA said a single component performing poorly can slow an entire computing job in these facilities. Warning signs can include a GPU running hotter than nearby units or a storage drive losing performance under load, the company said.
The company said Minerva creates a changing map of equipment, connections and computing jobs, allowing operators to follow problems across facilities with hundreds of thousands of GPUs. NVIDIA said its machine learning models learn normal behavior, detect unfamiliar problems and predict developing failures.
NVIDIA said the system also groups related alerts into an incident with a likely cause, reducing the number of individual symptoms operators must investigate. The company said this can help teams address problems earlier and keep GPUs running.
NVIDIA said an integrated AI assistant lets operators ask questions about computing cluster health and receive suggested fixes based on live operating data. Teams can implement fixes manually or connect Minerva to external systems that automate changes, the company said.
IREN is already using Minerva, NVIDIA said. Denis Skrinnikoff, IREN’s chief technology officer, said the combined view of networks, GPUs and jobs helps its operations team spend more time on preventive maintenance.
NVIDIA said Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla plan to deploy Minerva. Dell Technologies, Lenovo and Supermicro will offer the platform within their AI factory operations portfolios, the company said.
