Smart IOPS and H3 Platform Target One Billion IOPS for AI Infrastructure

The collaboration centers on the Unobtanium T50, a PCIe Gen6 x4 SSD from Smart IOPS that utilizes a specialized TruRandom controller to hit 50 million random-read IOPS per device. By integrating 20 of these units into a single air-cooled H3 Platform appliance, the companies intend to provide the massive, granular I/O performance required by modern AI workloads, such as graph neural networks and recommendation systems. The appliance is designed to support NVIDIA Storage-Next and the Scaled Accelerated Data Access (SCADA) programming model, which allow GPUs to bypass traditional CPU bottlenecks when accessing large datasets.

This architecture addresses the growing disparity between high-speed GPU compute and conventional storage interfaces. By using TLC NAND configured in pseudo-SLC mode, the T50 balances extreme performance with the manufacturing scalability of standard flash components. The system configuration includes four NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs and four NVIDIA ConnectX-8 network cards to ensure balanced throughput. Evaluation units are scheduled for release in the first quarter of 2027, with full production expected later that year.