Unigen has launched two PCIe 5.0 enterprise SSD families designed to feed increasingly storage-hungry AI infrastructure without turning DRAM into the storage budget.
The Cheetah High Capacity range uses PCIe Gen5 x4 and NVMe 2.0, with 16 NAND channels and a DDR5 interface. Unigen’s announcement says capacities reach 122TB in U.2 and E1.L formats.
Unigen’s product page lists 16TB, 32TB, 64TB and 128TB versions, while elsewhere on the same page it still claims 122TB. Someone in marketing and someone in engineering may need a quiet chat.
The HC drives offer sequential reads up to 14.17GB/s and writes up to 2.355GB/s, with up to 836,000 random read IOPS. Endurance reaches 0.3 drive writes per day for random workloads and 1.8 DWPD for sequential workloads.
The more exotic Cheetah Storage Class Memory range sacrifices enormous capacity for speed and endurance. It comes in 1.6TB and 3.2TB versions across U.2, E1.L, E1.S and E3.S form factors.
Cheetah SCM manages sequential reads of 14.7GB/s and writes of 11.1GB/s, with 3.37 million random read IOPS and 1.01 million random write IOPS. Random read latency can fall to 7.08 microseconds.
Unigen rates SCM for 60 DWPD on random workloads and 80 DWPD sequentially, putting it between conventional NAND storage and expensive DRAM for persistently hammered data.
Unigen wants customers to pair the two families. SCM handles active datasets, caching, write acceleration and scratch space, while HC stores foundation models, training data, embedding databases and longer-term archives.
That makes more sense as AI systems become storage problems as much as GPU problems. Huge models and vector databases need to be fed quickly enough that costly accelerators don’t spend their lives waiting for data to arrive.
Unigen targets AI training clusters, GPU inference farms, hyperscale clouds, HPC systems, vector databases, and GPU Direct Storage deployments. Both Cheetah SCM and Cheetah HC are already in production.
