ZettaLane posts MLPerf Storage v3.0 results with object-backed MayaNAS

ZettaLane Systems said Tuesday it submitted first-time MLPerf Storage v3.0 results on standard Google Cloud virtual machines and unveiled MayaNAS, a storage engine that runs Lustre directly on cloud object storage. The company says the setup is meant to give AI teams HPC-class performance with NAS simplicity while keeping data inside a customer’s own cloud.

Why it matters: - ZettaLane is targeting a long-running split in enterprise storage: teams often use one system for everyday NAS and another for high-performance AI and HPC workloads. - MayaNAS is designed to combine those workloads on the same data path, with object-storage economics instead of premium block-storage pricing. - The company is also pitching sovereignty benefits, with storage, buckets and encryption keys staying inside the customer’s own cloud account.

What happened: - ZettaLane Systems published its first MLPerf® Storage v3.0 submission and introduced MayaNAS. - The submission ran entirely on standard Google Cloud virtual machines. - ZettaLane submitted CLOSED-division results across the AI data pipeline with two cloud-native engines: MayaNAS and MayaScale. - MayaNAS is an object-backed parallel file system. - MayaScale is a companion NVMe-over-TCP block engine.

The details: - MayaNAS sustained 32.42 GB/s checkpoint write and 21.20 GB/s read on Llama 3 70B with two clients. - The MayaNAS data path ran entirely on cloud object storage, with no local scratch tier. - One MayaScale client on 200 Gbps networking fully fed 72 B200-class accelerators on RetinaNet at 86.68% utilization. - 3D U-Net training held about 92% accelerator utilization on MayaNAS and 93.31% on MayaScale. - The same platform served a Llama 3.1 8B inference cache at 524.65 tokens per second. - The platform checkpointed Llama 3 8B at 14.43 GB/s write and 10.40 GB/s read from a single client. - For MayaNAS, each OST is an OpenZFS dataset derived from regional Standard-class Google Cloud Storage buckets. - A small NVMe device holds metadata only. - ZettaLane said the configuration keeps bulk data read and written directly to object storage. - MLPerf® Storage v3.0 added support for an S3 object-storage access layer alongside its POSIX layer. - ZettaLane said MayaNAS fuses Lustre with object storage so existing applications do not need to be rewritten. - Both engines run inside the customer’s own cloud account. - ZettaLane said the file system, buckets and encryption keys never leave that account. - MayaNAS and MayaScale run across Google Cloud, Microsoft Azure and Amazon Web Services. - The stack deploys as code through the open-lustre-cloud project on GitHub. - ZettaLane positions the storage family alongside its other products, including MayaScale.

Between the lines: - ZettaLane is trying to show that object storage can support demanding AI training and inference workloads without forcing customers into separate HPC storage stacks. - The MLPerf submission gives the company a standardized benchmark to back up that claim. - The software-only, cloud-native approach also fits customers that want to keep storage in their own cloud environments for compliance or data-residency reasons. - MLCommons said new MLPerf Storage participants add data points that help build a fuller picture of storage performance under real-world training workloads.

What's next: - ZettaLane is likely to use the MLPerf results to push MayaNAS and MayaScale as a full-stack option for AI storage on public cloud infrastructure. - The company is also signaling broader deployment plans through support for multiple clouds and infrastructure-as-code delivery. - Further benchmark submissions or customer deployments could determine whether the approach gains traction beyond this initial MLPerf showing.

The bottom line: - ZettaLane is betting that AI teams want one cloud-native storage stack that can handle NAS, parallel file systems and block workloads without sacrificing performance or sovereignty.