MinIO ups object storage ante for AI
MinIO is making a case for object to be the storage protocol for AI. But one analyst said the market has yet to decide.
Software-defined object storage vendor MinIO is bolstering and rebranding its enterprise offering with a focus on AI and simplifying the use of object when it comes to model and data storage.
MinIO AIStor is a revamp of the vendor's exascale, high performance object storage offering Enterprise Object Store, including new features and higher performance. Added capabilities include promptObject, an S3 API that moves past basic object storage commands to natural language text prompts. AIStor also includes AIHub, a repository where customers can store and manage AI models and is API compatible with Hugging Face, an open source AI platform. MinIO also supports S3 over Remote Direct Memory Access (RDMA) for higher performance.
"The most valuable asset of an enterprise is their data. And within that data, the most valuable data is your trained models with specific knowledge about your business," said A.B. Periasamy, co-founder and CEO of MinIO.
Customers are not looking to train their data and move it to public repositories, Periasamy said. They want to keep AI models and datasets in the same place as their raw data, and that's what MinIO provides with AIStor.
Objectifying AI
With promptObject as an extension of the S3 API, users can query unstructured data using language text prompts to describe or compare objects, according to the vendor. Users don't need expertise with other AI technologies including vector databases and retrieval-augmented generation (RAG).
The promptObject API isn't a unique feature, according to Sanjeev Mohan, principal analyst at SanjMo. Snowflake's Document AI enables customers to ask a question of a PDF and get an answer. MinIO is doing that without the third party.
"You could have MinIO on bare metal Kubernetes with an object of a PDF in the object store and literally ask questions of that PDF," Mohan said. This simplifies the whole process, he said, adding, "Why do RAG when you can directly ask questions to the PDF?"
AIHub enables customers to store their model repositories on a private cloud, reducing the risks when using Hugging Face, the vendor said.
While Hugging Face has more than a million models for customers to experiment with, to reach the large language model, users have to go outside of the security perimeter, Mohan said.
"AIHub is an API-compatible copy of Hugging Face inside your company's security," he said.
MinIO's AIStor also works on-premises, which means customers can use AIHub to access large language models without moving their dataset, Mohan said.
Faster network, higher performance
MinIO's support for RDMA allows for direct data access without accessing the CPU and supports emerging networking speeds, including 400GbE and 800GbE.
While the adoption of these networks is still early, MinIO is addressing systems that will be built for AI and the data access those systems will need, according to Camberley Bates, an analyst at The Futurum Group.
"This is for customers who are trying to maximize the performance of their GPUs," she said.
Mohan said the adoption of RDMA is MinIO's way of delivering higher-performance object stores, which they couldn't do alone.
"You are skipping the operating system and the CPU, doing directly memory-to-memory between objects for faster network bandwidth," he said.
Is AI's future object storage?
The cloud is built on object storage, and so are the major AI players including Anthropic and OpenAI, according to Periasamy. With new AI applications, new infrastructure will be needed, including storage. Given this and its ability to scale to exabytes, Periasamy sees object as the storage of choice for AI.
Bates said that while object is the preferred method of storage by Kubernetes and cloud providers, the primary protocol for AI hasn't been set yet. Object storage generally lacks the performance needed by AI.
"MinIO has a very strong base and following," Bates said. "The question is how will the market move and standardize."
Adam Armstrong is a TechTarget Editorial news writer covering file and block storage hardware and private clouds. He previously worked at StorageReview.
