Supermicro ASG-4116S-NU160R：4U 伺服器容納 160 個 NVMe 插槽

At FMS 2026 (the Future of Memory and Storage conference, formerly Flash Memory Summit), Supermicro pulled back the curtain on the ASG-4116S-NU160R, a 4U server crammed with 160 NVMe SSD bays. That number is not a typo. Getting that many drives into a standard 4U chassis is an engineering challenge worth understanding, whether you are designing storage infrastructure or just keeping an eye on where the industry is heading.

💡 Why This Matters

Most high-density NVMe servers top out somewhere in the 24 to 48 drive range for a 2U or 4U footprint. Supermicro pushing to 160 bays in 4U changes the math on petabyte-scale flash deployments. Fewer servers mean fewer network ports, fewer power feeds, and less management overhead. For hyperscalers and large storage clusters, that consolidation is a direct cost reduction.

It also shows that density does not require a new drive form factor. Rather than moving to E3.S or other EDSFF drives, Supermicro sticks with 2.5-inch U.2 SSDs and borrows the layout of a top-loading SAS storage server, using PCIe switches where that design would use SAS expanders. The ASG-4116S-NU160R is a preview of what dense all-flash arrays can look like at the server level rather than the appliance level.

🏗️ The Architecture Behind 160 Bays

Fitting 160 NVMe drives into 4U requires a few design choices happening in parallel. Supermicro confirmed at FMS 2026 that the drives are 2.5-inch U.2 NVMe SSDs, not a smaller EDSFF form factor. It reaches 160 bays in 4U with a two-level slide-out tray design: the top tray carries the hot-swap SSDs, and the bottom tray holds the CPU, memory, and networking.

The connectivity side is the harder problem. NVMe drives connect over PCIe, and PCIe lanes are a finite resource. Running all 160 drives at four lanes each would take 640 PCIe lanes, far more than a single CPU provides. Supermicro solves this with PCIe switches in the center of the chassis, fanning out from the host CPU to the drive bays the way SAS expanders do in a top-loading SAS server. Supermicro lists the 160 U.2 bays as PCIe 4.0 x1, so each drive gets a single Gen4 lane rather than the usual x4. The switch chips have not been named, so check the official product page and the ServeTheHome coverage at the source link below for updates.

Power and thermals are the other constraint. 160 NVMe drives all active under load generate substantial heat. Supermicro's spec sheet lists hot-swappable rear fans and redundant (1+1) 2600W Titanium-level power supplies. ServeTheHome estimates roughly 130 to 137W per petabyte of storage at the outlet when the system is filled with 122TB-class drives.

📐 Form Factor and Drive Standards

The small form factor NVMe ecosystem is worth knowing if you are evaluating platforms like this:

E3.S: An EDSFF standard designed specifically for data center density. About 76mm by 112.75mm (7.5mm thick for 1T drives), with a native PCIe connector, NVMe, and hot-swap support. It uses a different connector and shape, so it does not fit 2.5-inch U.2 bays.
E1.S: Even smaller than E3.S, used in ultra-dense JBOF designs. Lower capacity per drive but maximum bays per rack unit.
M.2 (22110): Sometimes used in density builds but limited in hot-swap support and thermal headroom at high drive counts.
U.2 / U.3: The workhorse for enterprise NVMe. The ASG-4116S-NU160R uses 2.5-inch U.2 drives. Supermicro fits 160 of them in 4U with a two-level slide-out tray design rather than moving to a smaller form factor.

Supermicro's product page confirms the design: 160 top-loading hot-swap 2.5-inch NVMe bays running at PCIe 4.0, four front hot-swap E1.S PCIe 5.0 x4 bays, a single AMD EPYC 9005 series processor, up to 6TB of DDR5 across 12 DIMM slots, and three PCIe 5.0 x16 slots for networking or DPUs. See the official ASG-4116S-NU160R product page and the ServeTheHome coverage for the full datasheet.

🖧 Where This Fits in a Storage Stack

A 160-bay NVMe server is not a general-purpose compute node. It is purpose-built for specific workloads:

Distributed storage clusters running Ceph, MinIO, or similar software-defined storage layers
AI training dataset storage where random read throughput matters more than latency
Database tiering where the entire working set needs to live on flash
Video ingest and media asset storage at scale
NVMe-oF (NVMe over Fabrics) targets serving compute nodes over Ethernet or InfiniBand

For homelab use, this is firmly in the "interesting to know about" category rather than the "go buy one" category. But the architectural patterns it uses, PCIe switching in place of SAS expanders and slide-out service trays, show up at smaller scales too.

🔧 What to Watch For in the Spec Sheet

Supermicro's product page is already live, and these are the numbers worth paying attention to:

Total PCIe bandwidth: Supermicro lists the U.2 bays as PCIe 4.0 x1, and PCIe switches share the CPU's lanes across all 160 drives, so per-drive throughput is well below a typical x4 U.2 link. That trade-off favors capacity over peak per-drive speed.
Drive form factor: confirmed as 2.5-inch U.2, with four E1.S bays on the front panel.
Host interface: Onboard networking is just two 10GbE RJ45 ports (Broadcom BCM57416), so plan on adding NICs or DPUs in the three PCIe 5.0 x16 slots to move data at a speed that matches 160 NVMe drives.
Supported CPUs: a single AMD EPYC 9005 series processor (up to 360W TDP) drives the platform. The E1.S bays and expansion slots are PCIe 5.0, while the 160 U.2 bays run at PCIe 4.0 behind PCIe switches. Lane count from the CPU constrains everything downstream.
Redundancy: Drives and rear fans are hot-swap, and the two 2600W Titanium power supplies are 1+1 redundant. Whether the drive backplane and PCIe switch layer have any redundancy is worth asking Supermicro directly.

🛡️ Security Considerations at Scale

When you consolidate 160 NVMe drives into one chassis, the blast radius of a single point of failure goes up. A few things to think through before deploying dense storage like this:

Enable self-encrypting drive (SED) support if the drives offer it, so physical drive removal does not equal data exposure.
Plan your erasure coding or replication factor in software-defined storage to tolerate the loss of an entire chassis, not just a single drive.
Secure BMC/IPMI access on the management interface. A chassis this dense is a high-value target for ransomware via out-of-band management.
Keep firmware update cadence tight. NVMe drive firmware vulnerabilities are rare but real, and patching 160 drives manually is painful without automation.

✅ Wrap Up

The Supermicro ASG-4116S-NU160R is a signal, not just a product. It shows that all-flash density at the server level is catching up to what purpose-built appliances have offered, and that PCIe switching plus a top-loading 2.5-inch U.2 layout is the architecture getting it there. Whether you are planning a petabyte-scale Ceph cluster or just tracking where storage hardware is heading, this is worth bookmarking.

For confirmed specs and drive form factor, go straight to the Supermicro product page and the full ServeTheHome coverage at servethehome.com.

Related from the lab

Originally published at codewithromi.com.