Technical data sheet · ProLiant Compute Gen12 · 2U NVIDIA GH200 NVL2

HPE ProLiant Compute DL384 Gen12Specifications, configurations & comparisons

The HPE ProLiant Compute DL384 Gen12 is a 2U server built around NVIDIA’s GH200 Grace Hopper Superchip — one of them, or two joined as GH200 NVL2. It has no DIMM slots: CPU and GPU memory sit on each superchip, and HPE sets the slot, OCP and drive mix at the factory for good. This page explains what those fixed choices mean before you configure one.

2U Rack1 or 2 socketsNVIDIA GraceHPE iLO 6

144Arm Neoverse V2 cores72 per GH200 · two in NVL2

960GBLPDDR5X on the superchips480 GB per GH200 · no DIMM slots

288GBHBM3e GPU memory144 GB per Hopper GPU

900GB/sNVLink-C2C per superchipCPU↔GPU · 1800 GB/s for NVL2

8E3.S NVMe drivesdual GH200 · not hot-plug

4Flex Slot supplies1800–2200 W · 3+1 with NVL2

Configure a DL384 Gen12 →Get a UK quote

Every figure checked against the HPE ProLiant Compute DL384 Gen12 QuickSpecs (Version 16 - 03-Aug-2026)

Image: Hewlett Packard Enterprise

Form factor: 2U Rack · 2 sockets
Processors: NVIDIA Grace
Drive bays: Up to 8 · 7 drive-cage kits
Expansion: 4 × PCIe · 2 × OCP 3.0
GPUs: Up to 2 · 2 listed
Boot: M.2 · 4 boot options
Management: HPE iLO 6

01 — Overview

The DL384 Gen12 at a glance

Our read of HPE’s documentation: where the DL384 Gen12 is strong, and the limits worth knowing before you spec one.

Strengths

Memory is part of the superchip: each GH200 carries 480 GB of LPDDR5X for its 72 Grace cores and 144 GB of HBM3e for its Hopper GPU, linked by NVLink-C2C at 900 GB/s.
GH200 NVL2 puts two superchips, joined by NVLink, in one 2U box — 144 cores, 960 GB of LPDDR5X and 288 GB of HBM3e.
HPE optimises it for large-language-model inference that needs a lot of memory, and also names non-AI work — simulation at scale, EDA and weather forecasting.
Air-cooled in a standard 19-inch 2U chassis, and it keeps ProLiant iLO management, silicon Root of Trust and a three-year onsite warranty.
NVIDIA fabric cards run through the option list: CX-7 class InfiniBand/Ethernet, CX8 XDR adapters and BlueField B3220 and B3140H DPUs.

Watch-outs

The core build is frozen at the factory — superchip count and the PCIe, OCP and drive maxima cannot change later, and HPE sells no field upgrade kits.
No hardware RAID and no hot-plug drives, so any drive redundancy is a software matter.
Grace is an Arm Neoverse V2 CPU; management is iLO 6, with some iLO features unsupported on this Arm server, and the QuickSpecs name three Linux releases as initial certification plans.
HPE rates each Flex Slot 1800–2200 W supply for 200–240 V AC input, and a dual-superchip server needs four of them.
200 and 400 Gb adapters carry inlet-temperature limits, and several two-card combinations are not supported in the single-GH200 one-slot settings.

Processors2 SKUs

Cores / CPU72

Drive-cage kits7

Max bays8

GPUsup to 2

PSU options1 · to 2,200 W

02 — Processors

2 supported processors, mapped

There is no processor list to choose from. The CPU is the 72-core NVIDIA Grace — Arm Neoverse V2 — on each GH200 superchip, at 3.1 GHz base and 3.0 GHz all-core SIMD with 114 MB of L3 cache. Adding the second superchip (GH200 NVL2) takes the server to 144 cores and 228 MB of L3.

Processor landscape — cores × TDP

2 SKUs · 1 family

NVIDIA Grace 2Bubble size = L3 cache · hover a bubble for the full SKU

Processor explorer

filter · sort · compare

Sort

Processor	Cores / threads	Base	L3 cache	TDP	W / core	Memory
NVIDIA GH200 Grace Hopper Superchip	72	3.1 GHz	114 MB	W	0.0	—
NVIDIA GH200 NVL2 (2 x GH200 Grace Hopper Superchip, NVLink)	72	3.1 GHz	114 MB	W	0.0	—

Besides the base server (P71411-B21), HPE lists two superchip kits — the single GH200 bundle (P71406-B21) and the dual GH200 NVL2 bundle (P71407-B21) — each with the enablement parts for any supported build.
L1 cache is 64 KB instruction plus 64 KB data, and L2 is 1 MB per core.
HPE prints no CPU TDP or thread count for the superchip.

03 — Storage

7 drive-cage kits, up to 8 bays

The maximum drive count is fixed by the factory configuration you order: two or four E3.S NVMe drives with one GH200, four or eight with two. Every setting gives each superchip two EDSFF drives and one FHHL slot, then one extra choice per superchip — a second slot, an OCP 3.0 slot or two more drives.

Headline layouts

front panel view

Front panel

Dual GH200, 2 PCIe / 8xEDSFF setting (P72466-B21): the maximum of eight Gen5 E3.S NVMe drives, two FHHL x16 slots and no OCP slot.

SAS/SATANVMeSAS/SATA + NVMe

Controllers & boot

HPE 960GB NVMe Gen4 Mainstream Performance Read Intensive M.2 2280 Self-encrypting PE9010 SSD (P80327-B21) — boot
HPE 480GB NVMe Gen4 Mainstream Performance Read Intensive M.2 2280 PE9010 SSD (P80318-B21) — boot
HPE 960GB NVMe Gen4 Mainstream Performance Read Intensive M.2 2280 PE9010 SSD (P80321-B21) — boot
HPE 1.92TB NVMe Gen4 Mainstream Performance Read Intensive M.2 2280 PE9010 SSD (P80324-B21) — boot

No drives ship as standard; drives can be added later up to the factory-set maximum, but that maximum cannot be raised.
With two superchips, drives either fill the first GH200 before spilling to the second (unbalanced, P72469-B21) or split evenly across both (balanced, P72472-B21).
Up to two M.2 NVMe boot devices sit inside. HPE calls them PCIe Gen3 and recommends them for boot rather than cache, although the M.2 SSDs listed are named NVMe Gen4.
Boot targets — M.2, E3.S or PXE — are set in iLO, since there is no RAID controller.
The maximum-storage table prints 122.8 TB from 8 × 15.36 TB hot-plug SFF U.3 drives, which matches neither the E3.S-only drive list (it includes 30.72 TB E3.S models) nor the no-hot-plug rule, so we quote drive counts rather than that total.
The three single-GH200 setting kits carry “Gen11” in their HPE names (for example “HPE DL384 Gen11 1P 2EDSFF 2PCIe FIO”), although this is a Gen12 document.

Every documented drive-cage kit

6 with drive bays

2.5-inch3.5-inchEDSFF (E1.S / E3.S)

04 — Accelerators & I/O

GPUs, PCIe and networking

The GPU is the Hopper half of each GH200, not a PCIe card, and no other accelerator is listed. Per superchip HPE quotes 34 teraFLOPS FP64, 67 teraFLOPS FP64 Tensor Core and 3958 teraFLOPS FP8 Tensor Core with sparsity (1979 without); for the NVL2 pair the FP8 figure is 7916 teraFLOPS with sparsity. The document gives no GPU TDP. 1 or 2 Hopper GPUs (one per GH200) · 144 GB HBM3e each · no PCIe GPUs.

NVIDIA GH200 Grace Hopper Superchip - Hopper GPU 144GB HBM3e

144GB HBM3e · 1

NVIDIA GH200 NVL2 - 2 x Hopper GPU (288 GB HBM3e total)

288 GB HBM3e · 2

I/O topology

2 sockets

Each superchip owns its I/O. Depending on the factory setting, a GH200 NVL2 server has four FHHL x16 Gen5 slots, or two slots plus two OCP 3.0 x16 slots, or two slots and eight drives.

Network adapters HPE lists

11 options

Adapter	Ports	Form
Mellanox MCX631102AS-ADAT Ethernet 10/25Gb 2-port SFP28 Adapter for HPE	2 × 10/25GbE	PCIe
Mellanox MCX623106AS-CDAT Ethernet 100Gb 2-port QSFP56 Adapter for HPE	2 × 100GbE	PCIe
Mellanox MCX631432AS-ADAI Ethernet 10/25Gb 2-port SFP28 OCP3 Adapter for HPE	2 × 10/25GbE	OCP 3.0
HPE InfiniBand NDR200/Ethernet 200Gb 1-port OSFP PCIe5 x16 MCX75310AAS-HEAT Generic Adapter	1 × InfiniBand NDR200 / 200GbE	PCIe
HPE InfiniBand NDR/Ethernet 400Gb 1-port OSFP PCIe5 x16 MCX75310AAS-NEAT Generic Adapter	1 × InfiniBand NDR / 400GbE	PCIe
HPE InfiniBand XDR/Ethernet 2x400GbE 1-port OSFP PCIe6 x16 HHHL CX8 Crypto Adapter	1 × InfiniBand XDR / 2 x 400GbE	PCIe
HPE InfiniBand XDR400/Ethernet 400GbE 2-port QSFP112 PCIe6 x16 HHHL CX8 Crypto Adapter	2 × InfiniBand XDR400 / 400GbE	PCIe
HPE InfiniBand XDR PCIe Gen6 x16 Multi Host/Socket Direct Auxiliary Card with 250mm MCIO Cable Kit	×	PCIe
HPE Data Processing Unit InfiniBand NDR200/Ethernet 200Gb 2-port QSFP112 FHHL B3220 Adapter	2 × InfiniBand NDR200 / 200GbE	PCIe
HPE Data Processing Unit InfiniBand NDR/Ethernet 400Gb 1-port QSFP112 HHHL B3140H Adapter	1 × InfiniBand NDR / 400GbE	PCIe
MCX755106AC-HEAT 200Gb 2-port CX-7 class adapter	2 × 200Gb	PCIe

HPE supports one north–south BlueField B3220 (200 Gb, 2-port) and up to four east–west B3140H (400 Gb) per server; other NIC and fabric adapters are supported up to four per server.
10/25 Gb and 100 Gb Ethernet cards may go in any free slot; 200/400 Gb cards carry ambient limits — two B3140H in the single-GH200 two-slot setting are rated to 18 °C with optical cables and 23 °C with copper.
The CX8 XDR adapters are listed as PCIe6 x16 cards, while the server’s slots are specified as x16 Gen5.
The only OCP adapter listed is the Mellanox MCX631432AS-ADAI 2 × 10/25GbE; for RJ-45 HPE points to an SFP-to-RJ-45 adaptor.

05 — Power & cooling

1 power supply, 2,200 W

One supply is listed: the HPE 1800W–2200W Flex Slot Titanium hot-plug kit (P44712-B21), up to 96% efficient, on 200–240 V AC with C13 cords. A single-GH200 server runs on two (N+0), three (2+1) or four (2+2); a GH200 NVL2 server needs four, giving 3+1.

HPE gives the maximum heat dissipation of a fully populated GH200 NVL2 system as 3048 W (10.4k BTU/hr) and refers to its Power Advisor for input power.
Airflow at 25 °C ranges from 52 CFM (entry, idle, no DPU) to 240 CFM at maximum, and HPE’s full-load acoustic test declares 80 dBA.

06 — Management, security & OS

Run it, secure it, choose the OS

Management

HPE iLO 6 ASIC
Most-used iLO features enabled on Arm; exceptions listed in the iLO 6 User Guide
iLO RESTful API (Redfish conformant)
Active Health System
RESTful Interface Tool and Scripting Tools
Optional HPE iLO Advanced licence

Security

HPE silicon Root of Trust
UEFI Secure Boot and Secure Start
UEFI boot mode only — no legacy mode

Operating systems

HPE refers to its Servers Support & Certification Matrices for completed certifications
Initial certification plans named in the QuickSpecs: RHEL 9.4, Ubuntu 24.04 and SLES 15 SP5 QU1
NVIDIA AI Enterprise not included with the GH200 — sold separately through HPE

07 — Compare

How the DL384 Gen12 compares with its family

The DL384 is the only Grace Hopper system among the HPE rack servers Servnet documents; it has no x86 socket and no DIMMs. The matrix compares it with Xeon 6 models from the same Gen12 line, each drawn from its own verified spec file.

This modelSiblings and successorEach value from that model’s own verified spec file · bars scale per column

DL380a Gen12 data sheet DL380 Gen12 data sheet DL340 Gen12 data sheet Dell PowerEdge XE8712 (Grace Blackwell) data sheet Lenovo ThinkSystem SC777 V4 (Grace Blackwell) data sheet HPE DL380a Gen12 data sheet

08 — Use cases

Three builds that suit the DL384 Gen12

Three builds covering both superchip counts. The configurator offers only options it accepts for the DL384; the slot, OCP and drive maxima come from the factory setting you pick.

LLM inference

GH200 NVL2 inference server

HPE optimises the DL384 for inference on large language models that need a lot of memory. Two superchips give 288 GB of HBM3e and 960 GB of LPDDR5X, and the recommended 4 PCIe setting leaves four slots for east–west fabric.

2 × NVIDIA GH200 (NVL2) — 2 × 72-core Grace + 2 × Hopper GPU
960 GB LPDDR5X + 288 GB HBM3e, on the superchips
Dual GH200 4 PCIe factory setting (P72460-B21)
4 × 7.68 TB E3.S NVMe
BlueField B3140H 400Gb DPU
4 × 1800W–2200W Flex Slot Titanium (3+1)

The link opens with two supplies; GH200 NVL2 requires four, so add two. HPE rates one B3140H in this setting to 26 °C inlet with optical cables and 30 °C with copper.

Open this build in the configurator →

RAG & fine-tuning

Storage-heavy NVL2 for retrieval

HPE names fine-tuning and retrieval-augmented generation among GH200 NVL2 workloads. The 8xEDSFF setting keeps a document corpus or vector store on eight local Gen5 NVMe drives beside the superchips.

2 × NVIDIA GH200 (NVL2) with Grace CPUs
8 × 30.72 TB E3.S NVMe — 245.76 TB raw
Dual GH200 2 PCIe / 8xEDSFF factory setting (P72466-B21)
HPE NDR200 / 200GbE MCX75310AAS-HEAT adapter
4 × 1800W–2200W Flex Slot Titanium (3+1)
OS supplied by you (e.g. Ubuntu 24.04)

The link opens with two supplies — add two for the four NVL2 needs. No hardware RAID, so protect the eight drives in the OS or application; this setting has two PCIe slots and no OCP.

Open this build in the configurator →

HPC & EDA

Single-GH200 simulation node

Large-scale simulation, EDA and weather forecasting are HPE’s named non-AI targets. One GH200 with 480 GB of LPDDR5X and 144 GB of HBM3e suits scale-out nodes, and a 100 GbE card is a low-speed adapter HPE allows in any free slot.

1 × NVIDIA GH200 — 72-core Grace + Hopper GPU
480 GB LPDDR5X + 144 GB HBM3e, on the superchip
1 GH200 2 PCIe factory setting (P72451-B21)
2 × 7.68 TB E3.S NVMe
Mellanox MCX623106AS 2 × 100GbE
2 × 1800W–2200W Flex Slot Titanium (N+0)

Two supplies are enough for one GH200 but give no redundancy — add a third for 2+1. A single-GH200 server cannot be converted to NVL2 later.

Open this build in the configurator →

09 — Full specifications

HPE ProLiant Compute DL384 Gen12 specifications

Chassis & cooling

Form factor	2U, standard 19-inch rack · air cooled
Dimensions (H × W × D)	87.5 × 448 × 789 mm
Weight	Up to 26 kg with one superchip · up to 31 kg with two
Airflow at 25 °C	52 CFM (entry, idle) to 240 CFM (operating maximum)
Heat dissipation	3048 W (10.4k BTU/hr), fully populated GH200 NVL2
Rails	Rail kit included as standard

Superchip

Options	GH200 (P71406-B21) or GH200 NVL2 (P71407-B21)
CPU	72 Arm Neoverse V2 cores per GH200 · 3.1 GHz base / 3.0 GHz all-core SIMD
Cache	L3 114 MB per GH200 (228 MB NVL2) · L2 1 MB per core
GPU	Hopper with 144 GB HBM3e, up to 4.9 TB/s, per GH200
CPU–GPU link	NVLink-C2C 900 GB/s per GH200 (1800 GB/s NVL2)

Memory

Type	LPDDR5X on the superchip — no DIMM slots
Capacity	480 GB (one GH200) · 960 GB (two)
Bandwidth	512 GB/s · 1024 GB/s

Storage

E3.S NVMe Gen5	2 or 4 (one GH200) · 4 or 8 (two) · not hot-plug
M.2	Up to 2 NVMe · PCIe Gen3 per HPE (SSDs named Gen4) · boot use
RAID	No hardware RAID controllers

Expansion & networking

PCIe	Up to 4 × FHHL x16 Gen5 (factory setting)
OCP 3.0	Up to 2 × x16 (factory setting)
Adapter limits	1 × B3220 DPU · up to 4 × B3140H · other NICs up to 4
Built-in network	1Gb dedicated iLO port only

Power

Supplies	HPE 1800W–2200W Flex Slot Titanium (P44712-B21)
Count	2–4 with one GH200 · 4 with two
Input	200–240 VAC per supply · C13 cords

Ports, management & support

Rear	UID, 2 × USB 3.0, 1Gb iLO, serial, VGA
Front	LED panel, power button, iLO service port, 1 × USB 3.2
Management	HPE iLO 6
Acoustics	LWA,m 9.5 B · LpAm 80 dBA (full-load test)
Warranty	3 years parts, labour and onsite, next business day

10 — FAQ

DL384 Gen12 questions, answered from the documentation

Does the HPE DL384 Gen12 have DIMM slots?

No. Memory is integrated on the NVIDIA GH200 superchips: 480 GB of LPDDR5X for the Grace CPU and 144 GB of HBM3e for the Hopper GPU on each one. A GH200 NVL2 server therefore has 960 GB of LPDDR5X and 288 GB of HBM3e, and none of it can be expanded.

What processor does the DL384 Gen12 use?

The 72-core NVIDIA Grace CPU (Arm Neoverse V2) on each GH200 superchip, at 3.1 GHz base with 114 MB of L3 cache. The QuickSpecs list no other processor option. With two superchips the server has 144 cores.

What is the difference between one GH200 and GH200 NVL2?

GH200 NVL2 fits two superchips linked by NVLink, doubling cores, LPDDR5X and HBM3e. It raises the drive maximum from four to eight and the FHHL slots from two to four, and it needs four power supplies where one GH200 runs on two to four.

Can I upgrade a single-GH200 DL384 to two superchips later?

No. HPE sells the DL384 only as a factory-assembled system and offers no field upgrade kits; the number of superchips and the drive, PCIe and OCP maxima are set when it is built. Drives can be added later up to the factory maximum.

How many drives does the DL384 hold, and does it support RAID?

Up to eight E3.S NVMe Gen5 drives with GH200 NVL2, or four with one GH200, plus up to two M.2 boot devices. There is no hardware RAID controller and the drives are not hot-pluggable, so redundancy must come from software.

What networking can the DL384 take?

PCIe adapters including CX-7 class InfiniBand/Ethernet, CX8 XDR cards and BlueField B3220 and B3140H DPUs, plus one listed OCP 3.0 card. HPE supports one B3220, up to four B3140H and up to four other adapters per server. The only built-in port is the 1Gb iLO link.

What power does the DL384 Gen12 need?

HPE 1800W–2200W Flex Slot Titanium supplies on 200–240 V AC. One GH200 runs on two, three or four supplies (N+0, 2+1 or 2+2); GH200 NVL2 needs four for 3+1. A fully populated NVL2 system dissipates up to 3048 W.

Which operating systems run on the DL384 Gen12?

HPE points to its Servers Support & Certification Matrices for completed certifications. The QuickSpecs name RHEL 9.4, Ubuntu 24.04 and SLES 15 SP5 QU1 as the initial certification plans, the server boots in UEFI mode only, and NVIDIA AI Enterprise is not bundled with the GH200.

How big and heavy is the DL384 Gen12?

It is a 2U, 19-inch rack server measuring 87.5 mm high, 448 mm wide and 789 mm deep. It weighs up to 26 kg with one superchip and up to 31 kg with two, and a rail kit is included.

Is the HPE ProLiant DL384 Gen12 end of life?

HPE shows no withdrawal: its QuickSpecs first appeared in July 2024 and were updated to Version 16 in August 2026. They carry no end-of-service date for the model, and the Servnet end-of-life checker adds dates as HPE releases them.

Sources & verification

Where these figures come from

HPE ProLiant Compute DL384 Gen12 QuickSpecs (Version 16 - 03-Aug-2026) — Overview, Standard Features (superchip table, interfaces, management, OS), Configuration Information (factory PCIe/OCP/EDSFF settings, networking ambient tables, power supplies, drive balancing), Core Options and Technical Specifications.
Hewlett Packard Enterprise product imagery (HPE DAM) — Front render (badge “ProLiant Compute DL384 Gen12”) and rear render s00013155, the image used in the QuickSpecs “DL384 Gen12 – Rear View” figure.

Specifications are extracted from HPE’s published documents and independently re-checked line by line. Configurations are quote-based; availability and exact options are confirmed at quotation. Last reviewed 2026-09-11.

Get a quote for the HPE ProLiant Compute DL384 Gen12

Tell us the workload and we’ll spec and price it — or build it in the configurator first.

Business enquiries only — please use your company email. Gmail, Hotmail and other personal addresses are not accepted.

Delivery country (optional)

We only ship hardware to the countries in the dropdown. Software and maintenance services are global.

Enquiring about: HPE ProLiant Compute DL384 Gen12

Full Name *

Email *

Company

Quantity Required

Additional Notes

Response within 4 business hours · No obligation

Quote for

HPE ProLiant Compute DL384 Gen12

0800 987 4111

The DL384 Gen12 at a glance

Strengths

Watch-outs

2 supported processors, mapped

Processor landscape — cores × TDP

Processor explorer

7 drive-cage kits, up to 8 bays

Headline layouts

Controllers & boot

Every documented drive-cage kit

GPUs, PCIe and networking

NVIDIA GH200 Grace Hopper Superchip - Hopper GPU 144GB HBM3e

NVIDIA GH200 NVL2 - 2 x Hopper GPU (288 GB HBM3e total)

I/O topology

Network adapters HPE lists

1 power supply, 2,200 W

Run it, secure it, choose the OS

Management

Security

Operating systems

How the DL384 Gen12 compares with its family

Three builds that suit the DL384 Gen12

GH200 NVL2 inference server

Storage-heavy NVL2 for retrieval

Single-GH200 simulation node

HPE ProLiant Compute DL384 Gen12 specifications

Chassis & cooling

Superchip

Memory

Storage

Expansion & networking

Power

Ports, management & support

DL384 Gen12 questions, answered from the documentation

Where these figures come from

Buy & configure

AI & Arm guides

Plan the deployment

Support & lifecycle

Get a quote for the HPE ProLiant Compute DL384 Gen12

Talk to a UK specialist