Supermicro 160-Bay NVMe Server Reaches 20PB In 4U

A 160-bay NVMe storage server shown at FMS 2026 puts roughly 19PB to 20PB of flash capacity into 4U of rack space, ServeTheHome reported, turning storage density and serviceability into the central design tradeoff rather than raw drive speed alone.

The FMS unit is the Supermicro ASG-4116S-NU160R, a single-socket AMD EPYC 9005 storage server with 160 U.2 NVMe drive bays.

With 122TB- to 128TB-class SSDs, the system reaches about 20PB in one chassis, a level of density that would otherwise require several conventional storage servers.

160 U.2 Bays In A Top-Loading Chassis

The hardware is arranged around slide-out trays.

The upper tray carries the SSDs, while a lower tray holds the CPU, memory and networking hardware.

That layout matters for operators because top-loading storage systems must be serviced without pulling a fully loaded chassis out of the rack.

The drive shelf is built around rows of 10 drives on each side of a centre column.

The FMS floor unit could be opened far enough to show 60 of the 160 drives, and the centre section carries PCIe switches rather than SAS expanders.

A full four-lane connection for 160 NVMe devices would require 640 PCIe lanes before network adapters are added.

The system instead uses PCIe switching to fan the available lanes across the drives, accepting a throughput compromise in exchange for density, cable control and lower per-drive power.

One EPYC Socket Replaces Several Storage Nodes

The front panel adds four E1.S NVMe bays next to the IPMI port, dual USB ports and two 10Gbase-T ports using Broadcom BCM57416 silicon.

The platform also includes three available PCIe Gen5 x16 slots for 400GbE network cards or DPUs.

The single AMD EPYC 9005 CPU is rated up to 360W.

A similar bay count built from 24-bay 2U servers would need about seven systems, or 168 bays, which would add roughly six more CPUs and 72 DIMMs for comparable capacity.

That consolidation changes the operating calculation.

Fewer host platforms reduce the number of processors, memory channels and management endpoints tied to a given amount of flash storage.

The chassis still leaves room for front-panel management and expansion.

Four E1.S bays sit beside the IPMI connection, while the available Gen5 x16 slots give operators a path for higher-speed network cards or acceleration devices without turning the storage shelf into a separate external enclosure.

For large object stores, backup targets and analytics repositories, rack count and platform overhead become part of the storage-cost equation.

2.6kW Power Supplies Define The Tradeoff

The estimated system value with 122TB-class SSDs sits around $8 million to $10 million.

Redundant 2.6kW power supplies cover the drives, CPU and networking, producing an all-in outlet estimate of about 130W to 137W per PB of storage.

That power profile sits below the power required to run roughly 500 hard drives for similar capacity.

The PCIe switch architecture means each SSD will not reach maximum sequential throughput, but the design is aimed at dense capacity and lower power per stored petabyte rather than every drive operating at its individual peak.

The FMS example also points to a separate memory-cost angle.

Supermicro's chassis could house about 5TB of memory inside the NVMe SSDs themselves, and the source connected that platform idea to Sandisk's prototype work on reducing DRAM needs in future storage servers.

The disclosed design therefore combines about 20PB in 4U with fewer host platforms and an explicit PCIe-switching limit on per-drive throughput.