Technology

DDN Infinia

DDN's software-defined S3-compatible object storage for AI, sold alongside the AI400X3 appliance line — holder of the strongest MLPerf Storage v2.0 *audited* throughput numbers in the 2026 field (120.68 GB/s on 3D U-Net from 2RU).

8 connections2 resources1 post

Summary

What it is

Where it fits

The reference point for what a fabric-attached object-storage tier delivers when GPUDirect and S3-over-RDMA are wired end-to-end. In the VAST/WEKA/DDN appliance tier, DDN's differentiator this cycle is provenance: it submitted to MLPerf's audit while WEKA's 10.2 TB/s rack claim went unaudited.

Misconceptions / Traps

The headline 4 TB/s Eos number belongs to EXAScaler (Lustre/NVMe-oF), not the S3-compatible Infinia path — don't conflate the two product lines.
"MLPerf-audited" vs "vendor-published" is the load-bearing distinction when comparing this tier; insist on the former.

Key Connections

implements S3 API + RDMA (RoCE v2 / InfiniBand) — the S3-over-RDMA data path
solves Data Loading Bottleneck / GPU Starvation — audited proof the storage tier can feed GPUs
competes_with VAST Data / WEKA — the AI storage appliance tier

Definition

What it is

DDN's software-defined, S3-compatible object storage platform for AI workloads, sold alongside the AI400X3 appliance line (EXAScaler-based). Together they represent the current high-water mark of the GPUDirect Storage + S3-over-RDMA integration this index tracks: storage targets that let GPUs read data directly into HBM over the network fabric, bypassing the host CPU entirely. DDN's EXAScaler lineage includes powering NVIDIA's own Eos supercomputer at 4 TB/s aggregate throughput.

Why it exists

Historically, object storage rode standard TCP/IP, capping single-stream throughput far below what a GPU cluster consumes. DDN's bet is that when object storage pairs with XDR InfiniBand or Ultra Ethernet fabrics via GPUDirect and RDMA, it stops being a passive data lake and acts as a high-throughput disaggregated device capable of sustaining AI-training ingestion rates and real-time KV-cache retrieval.

Primary use cases

AI-training data delivery at fabric line rate, checkpoint read/write for large-model training, S3-over-RDMA data paths in GPU clusters, high-density appliance deployments where rack throughput-per-RU is the constraint.

Recent developments

Latest signals

Audited MLPerf Storage v2.0 results: 120.68 GB/s on 3D U-Net from a 2RU appliance. The DDN AI400X3 sustained 120.68 GB/s on 3D U-Net training workloads and processed Llama3-8b model checkpoints at 30.6 GB/s read and 15.3 GB/s write — from a compact 2RU enclosure (independent-methodology benchmark, MLPerf-audited). These are the strongest audited storage-throughput numbers in the 2026 field. Per StorageReview — best storage arrays 2026: AI leaders and audited results.
The audited-vs-claimed contrast is the provenance story. VAST Data and WEKA are similarly positioned in the parallel-filesystem-and-object space, but WEKA did not submit audited figures for the MLPerf Storage v2.0 round — leaving its 10.2 TB/s rack-scale claim unverified by independent auditors. When comparing vendors in this tier, "MLPerf-audited" versus "vendor-published" is the load-bearing distinction. Per StorageReview — audited MLPerf Storage v2.0 results.
EXAScaler at NVIDIA Eos: 4 TB/s aggregate. DDN's parallel-filesystem line powers NVIDIA's own Eos supercomputer over Lustre/NVMe-oF — the non-S3 ceiling that frames what the S3-compatible Infinia path is chasing. Per Introl — AI-optimized storage 2025.