HPE ProLiant ML350: The Tower That Thinks It Is a Rack Server

Most tower servers are compromises: fewer memory slots, one power supply, fixed fans, and a management controller that is a step down from whatever the rack line gets. The ProLiant ML350 is not that. It carries 24 DIMM slots, redundant hot-plug power supplies and fans, up to two dozen hot-plug drive bays, and the same iLO management stack that ships on the DL rack servers. It is a 5U tower with a conversion kit that turns it into a rack chassis when the room finally gets a rack.

Gen9 and Gen10 are different machines

They share a model number and very little else. The ML350 Gen9 is a Socket R3 (LGA 2011-3) platform running Xeon E5-2600 v3 or v4 processors, with four memory channels per socket. Twenty-four DIMM slots across two sockets means twelve slots per processor and three DIMMs per channel. Management is iLO 4, storage is Smart Array of the P440ar/P840 era, and the I/O is PCIe 3.0.

The ML350 Gen10 moves to Socket P (LGA 3647) and Xeon Scalable processors, with six memory channels per socket. The same 24 slots now mean two DIMMs per channel instead of three, which is a better arrangement. Management is iLO 5 with HPE's Silicon Root of Trust firmware validation, storage controllers are the P408i-a / P816i-a generation, and the chassis is rated for processors up to 205 W — a ceiling most towers never approach.

Nothing crosses over. Different socket, different heatsinks, different controllers, and memory qualified for one generation should not be assumed to work in the other even though both are DDR4. If you are consolidating spares across a mixed fleet, plan for two parts pools.

Memory: the part that gets misconfigured

On Gen10, six channels per socket means the balanced populations are six DIMMs per socket (one per channel) or twelve (two per channel). Eight DIMMs per socket — a number people reach for because it looks tidy — leaves two channels doubled and four single, and memory-bound workloads notice. If you are buying twelve modules, buy identical ones.

The clock ceiling comes from the processor, not the DIMM. On first-generation Xeon Scalable, Bronze parts run memory at 2133 MT/s, Silver 41xx and Gold 51xx at 2400, and Gold 61xx and Platinum 81xx at 2666. Second-generation Gold 62xx and Platinum 82xx parts move that to 2933. Buying 2933 MT/s modules for a Silver-class CPU buys you nothing but the ability to reuse them later.

Capacity has a second gate that surprises people: per-socket memory limits are set by the processor SKU. Standard parts cap well below the physical slot capacity, and reaching the headline maximum across all 24 slots requires the M- or L-suffix processors that were specifically built for large memory footprints. Decide the memory target first, then pick a CPU that can address it.

On Gen9 the arithmetic is different. Four channels per socket and twelve slots per socket means three slots per channel, so four, eight, and twelve DIMMs per socket are all balanced populations — the penalty for the third DIMM on a channel is clock, not balance. Populating the third slot on every channel drops the bus to 1600 MT/s with RDIMMs, where a two-per-channel population holds the processor's rated speed. Sixteen DIMMs at 2400 MT/s therefore move more bandwidth than twenty-four at 1600, which is worth knowing before you fill every slot for the sake of filling it.

Storage and expansion

Drive capacity in the ML350 comes from cages rather than a fixed backplane. Small-form-factor builds stack eight-bay hot-plug cages up to 24 bays; large-form-factor cages trade bay count for per-spindle capacity, up to 12 LFF. That flexibility is one of the reasons the ML350 ends up as a virtualization host and a backup target in the same server room.

Controller choice is the usual fork. A Smart Array with flash-backed write cache (the P408i-a and its larger sibling) is what you want in front of spinning disks doing random writes. If a software-defined storage layer owns the disks, run the controller in HBA mode or fit a plain host bus adapter and let the software see the drives directly.

With both processors installed the Gen10 exposes eight PCIe 3.0 slots. That is enough for a storage controller, 10 or 25 GbE networking, a SAS HBA for tape, and a double-width accelerator card with the right auxiliary power cabling — something few towers can physically or thermally accommodate. Typical configurations ship with an embedded quad-port gigabit adapter, with faster networking added by card.

Rack-class serviceability

The features that separate the ML350 from an office pedestal server are the boring ones. Power supplies and fans are hot-plug and redundant, so a fan failure is a scheduled part swap rather than an outage. Drives are hot-plug. iLO 5 gives you remote console, virtual media, hardware health telemetry, and the Active Health System log that HPE support will ask for. Firmware is maintained through the Service Pack for ProLiant rather than a scavenger hunt through individual downloads.

The tower-to-rack conversion kit is worth remembering at purchase time. A machine bought as a tower for a closet today can be racked later without replacing it, which is not true of most pedestal servers.

How it compares

	ML350 Gen9	ML350 Gen10	DL380 Gen10
Form factor	5U tower, rack convertible	5U tower, rack convertible	2U rack
Sockets	2	2	2
Processors	Xeon E5-2600 v3/v4 (LGA 2011-3)	Xeon Scalable (LGA 3647)	Xeon Scalable (LGA 3647)
Memory channels per socket	4	6	6
DIMM slots	24 (3 per channel)	24 (2 per channel)	24 (2 per channel)
Hot-plug drive bays	Up to 24 SFF or 12 LFF	Up to 24 SFF or 12 LFF	24 SFF front, rear cages optional
Management	iLO 4	iLO 5, Silicon Root of Trust	iLO 5, Silicon Root of Trust

The DL380 column is there to make the point: on paper the ML350 Gen10 and the DL380 Gen10 are the same server. The difference is 5U of floor space and acoustics you can live with versus 2U of rack space and fans you cannot sit next to.

Where to go from here

Start by deciding the generation, because that decides sockets, memory rules, and spares. From there, browse the current ML350 Gen9 configurations or the ML350 Gen10 configurations and check the drive cage layout against your storage plan. If the machine is going into a rack anyway, compare them with the DL380 Gen10 before you decide. When you want a specific processor pair, a balanced memory population, and a particular controller, send the configuration through on request a quote — include the workload and we will flag anything in the specification that the chassis or the CPU SKU will not actually support.