HPE ProLiant DL380 Gen9 vs Gen10: What Actually Changed

The DL380 is HPE's most widely deployed 2U server, and Gen9 and Gen10 units are both plentiful secondhand. From the front they look nearly identical, same 2U chassis, same drive cage layout, same SmartDrive carriers. The differences are underneath, and they are large enough that the two should not be treated as interchangeable.

Gen9 uses two Socket 2011-3 processors from the Xeon E5-2600 v3 (Haswell-EP) or v4 (Broadwell-EP) families, with four memory channels per socket. Gen10 uses two Socket P (LGA 3647) processors from the 1st- or 2nd-generation Xeon Scalable families, with six memory channels per socket. Both platforms give you 24 DDR4 DIMM slots, so the channel count is the change that matters, not the slot count.

What changed at the silicon level

Item	DL380 Gen9	DL380 Gen10
Socket	2 x Socket 2011-3	2 x Socket P (LGA 3647)
Processors	Xeon E5-2600 v3 / v4	Xeon Scalable 1st and 2nd gen
Max cores per socket	22 (E5-2699 v4)	28 (Platinum 8280)
Interconnect	QPI, up to 9.6 GT/s	UPI, up to 10.4 GT/s
Vector extensions	AVX2	AVX-512
Memory channels per CPU	4	6
DIMM slots	24	24
Max memory speed	2133 MT/s (v3), 2400 MT/s (v4)	2666 MT/s (1st gen), 2933 MT/s (2nd gen)
Memory capacity	1.5 TB with 24 x 64 GB LRDIMMs; up to 3 TB with 128 GB LRDIMMs on E5-2600 v4	3 TB with 24 x 128 GB LRDIMMs, plus persistent memory with 2nd gen CPUs
PCIe	Gen 3, up to 6 slots	Gen 3, up to 8 slots
Storage controller	Smart Array P440ar, P840	Smart Array E208i-a, P408i-a, P816i-a
Network	FlexibleLOM slot	FlexibleLOM slot
Management	iLO 4	iLO 5 with silicon root of trust

The move from four to six memory channels per socket is the headline. At the same DIMM count and the same speed grade, Gen10 has half again the theoretical memory bandwidth per socket, and it reaches higher rated speeds on top of that. It also changes how you populate: DL380 Gen9 balances in groups of four DIMMs per CPU, Gen10 in groups of six. Reusing a Gen9 memory plan on Gen10 will leave channels empty and give up bandwidth the platform is otherwise capable of.

One caution on Xeon Scalable: not every SKU delivers the platform's headline numbers. Silver 4100-series parts cap memory at 2400 MT/s and carry two UPI links rather than three, and Bronze parts are slower still. A well-specified E5-2600 v4 pair can outperform a poorly chosen Scalable pair. Look at the specific processor model, not the generation label. If you plan to fit 2nd-generation Scalable processors, check the system ROM level as well, since that support arrived in a later firmware release.

Storage and I/O

Both generations offer the familiar drive box arrangement: 8, 16 or 24 front 2.5-inch SFF bays, or 12 front 3.5-inch LFF bays, with optional rear cages. HPE kept the SmartDrive carrier design across this range, so 2.5-inch trays are generally interchangeable between Gen8, Gen9 and Gen10, which is a genuine convenience if you keep spares. That is not true of every vendor.

NVMe is where they part company. On Gen9, NVMe arrived through an Express Bay enablement kit, a limited number of bays wired through dedicated PCIe paths. On DL380 Gen10, NVMe backplane options are part of the standard configuration menu across multiple drive boxes, and the additional PCIe slots give you room for the adapters those layouts need.

The controller change is also more than a part number. Gen10 Smart Array controllers support mixed mode, presenting some drives as RAID volumes and others as pass-through HBA devices from the same card at the same time. That is directly useful for software-defined storage, where the OS wants raw disks but you still want a hardware-mirrored boot pair. On Gen9 the usual approach is to flash or add a separate HBA.

Networking stayed on HPE's FlexibleLOM slot for both generations, so the same class of 10 GbE and 10/25 GbE FlexibleLOM adapters is available. The shift to OCP 3.0 came later, with Gen10 Plus, not with Gen10.

iLO 4 versus iLO 5, and the firmware question

iLO 4 works. It gives you remote console, virtual media, health monitoring, power control and a REST interface. iLO 5 is faster, has a cleaner HTML5 console, and adds a hardware root of trust that validates firmware against silicon-embedded fingerprints at boot and can recover a corrupted image. If firmware integrity is part of your compliance posture, that is a concrete reason to prefer Gen10.

On both generations, the remote console and virtual media features that make a server manageable from a desk require an iLO Advanced license. Verify the license status of any unit before you plan around it.

Firmware access deserves planning. HPE gates Service Pack for ProLiant downloads and most component firmware behind an active warranty or support agreement. Decide up front how you will obtain system ROM, iLO, Smart Array and NIC updates for the life of the deployment, and factor that into the choice. Also check your hypervisor's compatibility list against the exact CPU model, since newer releases have steadily dropped older processor generations.

When Gen9 is still the right call

Gen9 remains a sensible choice for workloads that are not bandwidth-bound: file and print, backup targets, lab and staging clusters, dedicated application hosts, branch-office virtualization with a modest VM count. It is also the pragmatic answer if you already run a Gen9 fleet, because carriers, rails, FlexibleLOM cards and controllers are already on your shelf and your runbooks already work.

Gen10 earns the move when memory bandwidth per host is the limit, or when you need large capacity at speeds Gen9 cannot reach, when you want NVMe as a first-class option rather than an add-on kit, when you need firmware integrity assurance, or when the host has to stay on a supported hypervisor build for several more years.

Where to go from here

Have a look at the available HPE ProLiant DL380 Gen9 systems and DL380 Gen10 systems to see which processor pairs, drive cages and controllers are in stock. If the configuration you need is not on the shelf, request a quote and specify CPUs, memory, drive layout, controller and FlexibleLOM adapter for the build you want. Include the workload and rack details with your request and we will point out anything in the specification that will not do what you expect.