At FMS 2026, we saw a really neat system. The Supermicro ASG-4116S-NU160R is a single-socket AMD EPYC server that sports 160x U.2 NVMe drive bays. Taking only 4U of space, and combined with even “just” 122-128TB class NVMe SSDs, like the Solidigm D5-P5336 122.88TB NVMe SSD we reviewed, this is roughly 20PB of storage in a single system.
Supermicro ASG-4116S-NU160R Overview
Looking at the system, you can tell that Supermicro is using a design where there are slide-out trays for servicing, with the top being for SSDs and the bottom being for the CPU, memory, and networking.
Slide-out trays are important in top-loading storage systems because they allow servicing without pulling the entire system out of the rack.
Pulling out the drive shelf, we can see just how neat this system is. There are rows of 10 drives on either side of the center column. Since this was on a table, not secured on a rack, we could only slide this out far enough to show you 60 of the 16 drives, but you get the idea.
In the center section, there are PCIe switches. As you might imagine, 160x NVMe devices all running with four lanes would require 640 PCIe lanes, and that is without NICs to move data on and off of the system.
Supermicro’s solution is to use a setup similar to what we would expect from a top-loading SAS storage server, but instead of SAS expanders, it is using PCIe switches.
Underneath the SSD tray, you can see the CPU tray starting with a VGA port and a power button on the left-hand side.
When we say this is a 160-SSD design, that is technically selling it short. There are four E1.S NVMe SSD bays on the front panel next to the IPMI port, dual USB ports, and then two 10Gbase-T ports (Broadcom BCM57416). Given there are full-height I/O expansion slots and four E1.S bays between the VGA port and the USB ports, it must be interesting when you have a KVM cart in the cold aisle and hook up to one of these.
On the right side, there are more expansion slots.
The CPU tray pulls out as well, and you can see that those expansion slots are not just empty risers. There are three available PCIe Gen5 x16 slots for your 400GbE NICs and DPUs in this system.
One of the other interesting bits is that the system uses a single AMD EPYC 9005 CPU up to 360W. If you had 24-bay 2U servers, you would need 7 or so of them (168 bays total) to have a similar capacity. So this design saves you roughly 6 CPUs and 72 DIMMs over using standard 2U servers.
Perhaps that is really the point of all of these servers.
Final Words
Taking a step back, this is likely an $8-10M machine with 122TB-class SSDs putting it into the 19-20PB in 4U range. While the PCIe switch architecture means that each drive will not achieve its maximum sequential throughput, there is another benefit. By limiting the lanes to each drive, the drives can run at lower power. This system, including the drives, NICs, and CPU, is driven through redundant 2.6kW power supplies. Call that only 130-137W per PB of storage all-in at the outlet, which seems quite amazing. That power envelope is below just the power required to drive the 500 hard drives you would need for a similar capacity.
At FMS, this was the densest storage platform we could find. It also would house something like 5TB of memory just in the NVMe SSDs. Given memory prices, we actually used this server as an example of how Sandisk’s innovative prototype drive in a platform like this server (with E3.S SSDs instead) could deliver massive DRAM savings in the future. Still, it was neat to see on the FMS show floor.
 that can deliver crazy storage density){kind=link}