Smart IOPS and H3 Platform detailed an AI compute storage design that pairs Smart IOPS’ Unobtanium T50 SSD with a GPU-centric, air-cooled H3 Platform appliance, targeting up to one billion 512-byte random IOPS in a single system. The configuration is intended for NVIDIA’s Storage-Next initiative and the NVIDIA SCADA (Scaled Accelerated Data Access) programming model, which focus on GPU-initiated access to NVMe storage for working sets that don’t fit in local GPU memory.
The Smart IOPS Unobtanium T50 is described as a PCIe Gen6 x4, NVMe 2.0 device in an E3.S 2T form factor. Smart IOPS lists preliminary performance targets of up to 50 million random-read IOPS and up to 10 million random-write IOPS at a 512-byte block size, along with up to 28 GB/s sequential reads and 24 GB/s sequential writes. Capacity options listed are 9 TB, 18 TB, and 36 TB.
Four T50 devices are presented as an aggregate, device-level target of up to 200 million random-read IOPS, positioned to align with the I/O capability of next-generation PCIe Gen6 x16 GPUs such as NVIDIA Rubin, with the companies noting that actual system performance depends on workload, software, topology, configuration, and other system-level factors.
One design choice Smart IOPS calls out is using “broadly available” TLC NAND operated in pseudo-SLC mode to reach the high IOPS targets. The company also says the devices can support specialized low-latency NANDs when workload requirements justify the cost.
On the system side, H3 Platform’s planned appliance is built around four NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, four NVIDIA ConnectX-8 NICs, and 20 E3.S 2T storage slots rated for up to 60 W per slot. With each slot designed to support up to 50 million IOPS, the 20-slot configuration is where the one-billion-IOPS figure comes from. It’s an unusually storage-dense design for an air-cooled GPU appliance, and the 60 W-per-slot envelope is a practical detail operators will care about when they start mapping the platform to rack-level power and thermal budgets.
Ashutosh Das, CEO of Smart IOPS, said, “With the Unobtanium T50 ACS device, our goal is to deliver 50 million IOPS in a compact PCIe Gen6 x4 E3.S form factor using standard TLC NAND.” Brian Pan, CEO of H3 Platform, said the appliance is being developed to combine NVIDIA accelerated computing and networking with “20 high-power E3.S slots in an efficient, air-cooled platform.”
Smart IOPS and H3 Platform expect evaluation samples and evaluation systems in Q1 2027, with production planned for Q2 2027.
Source: Smart IOPS
