Skip to content

Dear customers! Please send orders and inquiries to [email protected] or to Telegram @servix_uz instead of calling by phone: due to the high workload, the operator cannot answer all incoming calls.

Servix

HPE ProLiant DL320 Gen11: review of the single-socket 1U server

Servix technical team · Updated

In brief. The DL320 Gen11 is a 1U HPE rack server with one 4th or 5th Gen Intel Xeon Scalable processor: up to 60 cores and 270 W per socket, 16 DDR5 slots (up to 4 TB), PCIe 5.0 bus, iLO 6 management. The point of the model is that one modern Xeon handles a workload that five years ago required two processors — and virtualization licenses are counted by physical sockets and cores. In the Servix catalog the configuration costs 41,000,000 UZS, the dual-socket DL360 Gen11 — 40,000,000 UZS, the tower ML30 Gen11 — 31,000,000 UZS.

Quick navigation: what the machine is · one socket instead of two · memory and channels · drives and controllers · networking: OCP3 and PCIe · power and cooling in 1U · what to compare with · catalog prices · what to check before ordering · questions and answers

HPE ProLiant DL320 Gen11: review of the single-socket 1U server

HPE ProLiant DL320 Gen11 — single-socket 1U rack server

What this machine is

The DL320 Gen11 took a place in the lineup that did not exist before: a full-fledged 1U rack server, but with a single processor socket. Before Gen11, the 1U choice looked like this: either the dual-socket DL360 or nothing. Single-socket machines lived in the tower form factor or in the DL20 family, which is noticeably weaker in memory and drives.

  • Processor. One 4th Gen (Sapphire Rapids) or 5th Gen (Emerald Rapids) Intel Xeon Scalable. The upper limit is 60 cores and a 270 W thermal design power; the fifth generation on the same socket platform offers more cache and higher memory frequencies.
  • Memory. 16 registered DDR5 slots, up to 4 TB with 256 GB modules. Eight channels per socket, two modules per channel.
  • Bus. PCIe 5.0 — twice the bandwidth per lane compared to Gen10 Plus. The number of slots depends on the riser, plus a separate OCP 3.0 slot for a network adapter.
  • Drives. Depending on the drive cage, 4 bays of 3.5″ or 8 bays of 2.5″; there are versions with 10 SFF bays and for EDSFF E3.S drives.
  • Management. iLO 6 with a dedicated network port and a hardware root of trust — firmware signature verification at every power-on.

The practical conclusion: this is a machine for an edge site, a branch office with a rack, a mid-sized virtualization node, and for tasks where density matters but a second processor is not needed.

One socket instead of two: when it pays off

The main argument for a single-socket platform in 2026 is not the hardware price but licenses and memory topology.

Licenses. Both Microsoft and VMware charge per core, but with minimums. VMware's minimum is 16 cores per physical processor: two eight-core CPUs in a dual-socket machine are billed as 32 cores, while one sixteen-core CPU is billed as 16. Windows Server is licensed from 16 cores per server and at least 8 per processor, so “spreading” a workload across two weak CPUs is almost always more expensive than consolidating it on one strong one.

NUMA. A single-socket system has one memory node. A virtual machine with a large amount of memory cannot end up “split” between two processors — and this is exactly what most often explains the fluctuating performance of 1C and databases on dual-socket servers where nobody kept track of virtual machine sizes.

When two sockets are unavoidable. You need more than 60 cores in one system; you need more than 4 TB of memory; you need many PCIe lanes for several adapters at once — for example, two FC HBAs and two 25G cards; or the corporate standard explicitly requires processor-level redundancy. In these cases the right choice is the DL360 Gen11 in the same 1U or the DL380 Gen11 in 2U if you need drives and slots.

Memory: eight channels and the identical-modules rule

Sixteen slots means two modules for each of the eight channels. Two practical rules follow from this.

  • Populate channels, not slots. Eight 32 GB modules give 256 GB with all channels at full speed. Four 64 GB modules give the same capacity but leave half the channels empty — bandwidth drops by roughly half, and this shows up immediately on databases and virtualization.
  • A second module per channel costs frequency. When all sixteen slots are populated, the controller lowers the DDR5 frequency. If memory expansion is planned for the future, it is cheaper to take eight higher-capacity modules right away than to add a second row a year later and lose frequency.

There is no need to mix modules of different capacity, rank and manufacturer within one server: the system will start but will run at the level of the worst module, and diagnosing memory errors will turn into trial and error.

Drives and controllers

The drive cage in the DL320 Gen11 is chosen at the ordering stage and can later be changed only together with the backplane and cables. Three typical options:

  • 4 × 3.5″ (LFF). Maximum capacity per rack unit: four 22 TB drives such as the WD Ultrastar DC HC570 (8,625,882 UZS per drive) give 66 TB of usable capacity in RAID 5 at a height of 1U. Random access speed is low, though — this is an option for archives and backups.
  • 8 × 2.5″ (SFF). A universal cage for SAS SSDs and SATA SSDs. For virtualization this is the main configuration.
  • NVMe. The bays connect directly to the processor via PCIe 5.0. What matters here is not the number of bays but the number of lanes: full x4 per drive is not available in every backplane variant.

A separate note on system boot. Using two of the eight drive bays for the hypervisor is wasteful; for this there is the HPE NS204i-p (10,063,529 UZS) — a board with two NVMe M.2 drives in a mirror, which the operating system sees as one boot volume. Dell implements the same idea in BOSS cards (7,907,058 UZS).

RAID controllers: for simple volumes and software-defined storage, the E208i-a (5,750,588 UZS) without cache is enough; disk arrays with write workloads need the MR416i-a (8,625,882 UZS) with 4 GB cache and non-volatile protection. A breakdown of the lineups is in the guide to PERC and Smart Array controllers.

Networking: OCP 3.0 slot and PCIe cards

As standard, the server provides gigabit ports — enough for management and light office traffic, but not for backup or for connecting to a storage system. The platform is designed for the network adapter to be installed in the OCP 3.0 slot: the card does not occupy a regular PCIe slot and can be replaced without disassembling the server.

  • 10 GbE. HPE 562SFP+ — 5,031,764 UZS. A working minimum for connecting to the network core and to storage over iSCSI.
  • 25 GbE. HPE 640SFP28 — 5,750,588 UZS. Works at both 10 and 25 gigabits, so it is chosen “for growth” with a future switch in mind.
  • 4 × 1 GbE. Intel I350-T4 OCP3 — price on request; chosen when many ports are needed and speed is not.

In 1U the number of slots is limited, so the order is: first OCP 3.0 for networking, then PCIe for a RAID controller or FC HBA. If the project needs two network adapters, an external RAID and an HBA, that is already an argument for 2U. What goes into what is covered in the article on network adapters and transceivers.

Power and cooling in 1U

A 270 W processor in a chassis 44 mm high is a separate engineering challenge, and it has consequences for the order.

  • Fan kit. Top-end processors and dense drive cages require a high-performance cooling kit. This is a configurator item, and if it is skipped, the server will throttle under load or will not accept the selected CPU at all.
  • Power supplies. Two hot-swap power supplies are the norm, but the wattage is chosen for the configuration, not “at the maximum”: a power supply running at 20% load has worse efficiency than a correctly sized one.
  • Room temperature. A 1U with a high TDP has a lower upper limit of operating temperature than a 2U. A server room with a household air conditioner that is turned off at night is not a place where such a machine will last long.
  • Noise. A 1U with turbine fans does not go into an office room. If there is no rack and the room is occupied by people, the right choice is the tower ML30 Gen11.

What to compare with

Model

Form factor

Sockets

When to choose

DL320 Gen11

1U rack

1

Density, modern Xeon, license savings, edge site with a rack

DL360 Gen11

1U rack

2

More than 60 cores, more than 4 TB of memory or processor-level redundancy needed

DL380 Gen11

2U rack

2

Many drives and expansion slots: FC HBA, several network adapters, GPU accelerators

ML30 Gen11

tower

1

No rack, the server sits in a room with people, light workload

MicroServer Gen11

compact

1

Branch office for 10–⁠20 people, four drives, no hot swap needed

A common question is “why the DL320 if the DL360 costs about the same?” The answer is that you need to compare not chassis but the total cost of ownership: two processors require twice as many memory modules to populate all channels, give twice as many licensable cores under virtualization minimums and consume noticeably more electricity. On a workload that fits in one socket, the difference over three years adds up to a substantial amount.

Catalog prices as of November 2026

Prices in the listings are for specific configurations; the final amount depends on the processor, memory capacity, drive cage and controller. An exact calculation for your task is easiest to start with the configurator.

What to check before ordering

  • How many cores you really need and how they are licensed: VMware's minimum of 16 cores per processor often settles the debate between one and two sockets.
  • Which drive cage: 4 LFF for capacity, 8 SFF for general-purpose workloads, NVMe for databases. After assembly, the cage cannot be changed without replacing the backplane.
  • Whether a separate boot module is included so that the hypervisor does not occupy drive bays.
  • Eight memory modules instead of four — whether all channels are populated.
  • A fan kit for the selected processor: for top-end CPUs it is mandatory.
  • Free 10 or 25 Gbps ports on the switch and modules for them.
  • iLO license: the remote graphical console and image mounting are available only in the advanced version.
  • Rack depth and whether there is a rail kit with a cable management arm.

The DL320 Gen11 is the right choice where you need a modern rack server, but a second processor would be a tribute to habit. We will build a configuration for your task, calculate the licenses and tell you honestly if the DL360 or ML30is more cost-effective in your case. Write to us; the entire Gen11 lineup with prices is here.

Questions and answers

How does the DL320 Gen11 differ from the DL360 Gen11?

In the number of processor sockets: the DL320 has one, the DL360 has two. Both have a 1U chassis. Everything else follows from the single socket: 16 memory slots instead of 32, fewer PCIe lanes and, as a result, fewer expansion slots. For a workload that fits within 60 cores and 4 TB of memory, the DL320 is more cost-effective in terms of licenses and electricity.

How many drives fit in the DL320 Gen11?

Depending on the selected drive cage, 4 drives of 3.5″ or 8 drives of 2.5″; there are versions with 10 SFF bays and for EDSFF E3.S drives. The cage is chosen when ordering: changing it later means replacing the backplane and cables, and sometimes the chassis itself.

Which processor goes into the DL320 Gen11?

A 4th Gen (Sapphire Rapids) or 5th Gen (Emerald Rapids) Intel Xeon Scalable, up to 60 cores and up to 270 W. AMD EPYC does not fit this platform — HPE's single-socket EPYC machine is called the DL325.

Is a separate RAID controller needed?

If you need hardware RAID 5 or 6 with write cache — yes, and with non-volatile cache protection. For a mirror of two drives or for software-defined storage such as Storage Spaces and ZFS, a controller without cache is enough, and sometimes it is better to pass the drives through directly in HBA mode.

How much does the HPE ProLiant DL320 Gen11 cost in Tashkent?

The configuration in the Servix catalog is 41,000,000 UZS. The final price depends on the processor, memory capacity, drive cage, controller and network adapters; a calculation for a specific task is made on request, with delivery across Tashkent and the regions.

Is the DL320 Gen11 suitable for 1C?

Yes, and this is one of the tasks it suits best: a single NUMA node, high frequencies in the lower Xeon models and fast NVMe drives give good responsiveness on a typical database. Configuration sizing by number of users is on the page server for 1C.

Related reading