UK’s trusted IT infrastructure partner since 2003
Servnet
FinanceToolsConfiguratorGet in Touch
Technical data sheet · PowerEdge XE rack-scale · 1U GB300 NVL72 compute sled

Dell PowerEdge XE9712Specifications, configurations & comparisons

The XE9712 is one compute sled of an NVIDIA GB300 NVL72 rack, built into Dell’s 48 RU IR9048 Open Rack v3 cabinet. Each 1U sled holds two GB300 boards — one Grace CPU and two Blackwell Ultra GPUs apiece — and relies on the rack for power, coolant and GPU-to-GPU NVLink. This data sheet shows what one sled contributes and what the rack must provide around it.

Rack-scale, liquid-cooled 1U compute sled2 socketsNVIDIA GraceOpenBMC on Aspeed AST2600
4GPUs per sledBlackwell Ultra · 288 GB HBM3e each
2GB300 boards per sled1 Grace CPU + 2 GPUs on each
2,900Wtotal TDP per GB300 boardDell’s figure; its part TDPs sum higher
8E1.S NVMe bayseach slot filled or blanked
4800 Gb/s OSFP portstwo CX8 mezzanines
72GPUs per rack, maximumNVLink, 1.8 TB/s GPU-to-GPU (spec sheet)

Every figure checked against the Dell PowerEdge XE9712 Technical Guide (Rev. A00, November 2025)

Form factor
Rack-scale, liquid-cooled 1U compute sled · 2 sockets
Processors
NVIDIA Grace
Drive bays
Up to 8 · 1 drive-bay layout
Expansion
2 × PCIe
GPUs
Up to 4 · 1 listed
Boot
M.2 boot riser
Management
OpenBMC on Aspeed AST2600
01 — Overview

The XE9712 at a glance

Our read of Dell’s documentation: where the XE9712 is strong, and the limits worth knowing before you spec one.

Strengths

  • Four 288 GB Blackwell Ultra GPUs and two Grace CPUs per 1U sled, with rear NVLink connectors into the rack’s NVLink switches; Dell’s spec sheet puts up to 72 GPUs in one NVLink rack with 1.8 TB/s of GPU-to-GPU bandwidth.
  • Two CX8 mezzanines give four 800 Gb/s OSFP ports per sled, which the spec sheet calls 1:1 NIC-to-GPU, and the guide lists InfiniBand NDR and XDR mezzanine support with GPU direct over Gen6 x16.
  • Power comes from the rack: the sled carries no power supplies and takes up to 54 VDC from a busbar fed by 33 kW shelves, and a Delta capacitance shelf can charge or discharge 15 kW for five seconds.
  • Leak handling is designed in: the cold plates and inner manifolds carry leak sensors, and Dell’s Integrated Rack Controller can isolate fluid and cut power through OpenManage Enterprise.
  • Eight hot-swap E1.S NVMe bays per sled — four times the two E3.S bays Dell gives the XE8712 tray.

Watch-outs

  • A lone sled has no use outside the NVL72 rack: GPU-to-GPU NVLink runs through the rear connectors to the rack’s NVLink switches, and power and coolant come from the busbar and manifold. Neither Dell document prints a sled count; the spec sheet’s maximum of 72 GPUs at four per sled works out to 18 by our arithmetic.
  • Software choice is narrow: the guide lists only GraceOS, an NVIDIA-qualified OS, and the spec sheet lists Ubuntu. The standard sled runs OpenBMC; iDRAC 10 appears only on the “XE9712 with iDRAC” version in the spec sheet.
  • Little PCIe freedom: Dell’s thermal table allows one BlueField-3 B3240 in slot 2 or two CX8 C8180 cards in slots 1 and 2, each with specified NVIDIA transceivers, and pairs the two options with different E1.S drive counts.
  • Dell’s documents disagree on several numbers — Grace TDP (500 W in Table 1, 300 W per board in the TDP table), the board total (2900 W, though the listed CPU and GPU TDPs add to 3,100 W), NVLink-C2C count (up to 10 or up to four), slot count (one or two), sled depth (807 or 836.8 mm), busbar voltage and the storage controller — so confirm the build with Dell.
  • The coolant is prescribed: Recochem PG25 is the only authorised fluid, which sets a −5 °C non-operating floor for liquid-filled parts.
Processors1 SKU
Cores / CPU72
Drive layouts1
Max bays8
GPUsup to 4
Rack power parts1 · to 5,500 W
02 — Processors

1 supported processor, mapped

The guide calls the sled two-socket, but the two NVIDIA Grace processors (“Grace Ultra” in the chapter-one summary) are not separate options: each sits on its own GB300 board with two GPUs, and the pair is joined by six cabled C-Links for fixed, coherent CPU-to-CPU traffic. Dell gives 72 cores per processor and speeds up to 3.6 GHz; thread count, base clock and cache are not published.

Processor landscape — cores × TDP

1 SKUs · 1 family
Dell PowerEdge XE9712: every supported processor by cores and TDP081624324048566472500 WCores per processorNVIDIA Grace (Grace Ultra, 72 cores) — 72 cores · 500 WNVIDIA Grace (Grace Ultra, 72 cores)72 cores · 6.9 W/core — top SKU and most efficient
NVIDIA Grace 1Bubble size = L3 cache · hover a bubble for the full SKU

Processor explorer

filter · sort · compare
ProcessorCores / threadsBaseL3 cacheTDPW / coreMemory
NVIDIA Grace (Grace Ultra, 72 cores)72500 W6.9
  • Table 1 lists a maximum TDP of 500 W. The TDP table budgets 300 W per CPU and notes that the CPU can reach that 300 W maximum by borrowing power from its GPUs.
  • Each processor reaches its GPUs over NVLink-C2C with 900 GB/s of coherent memory; Table 2 says up to 10 NVLink-C2C links, while the GPU appendix says up to four.
  • The same table lists 2900 W as the total per board, while its CPU (300 W) and GPU (2 × 1400 W) entries add up to 3,100 W; Dell does not reconcile the two.
  • The TDP table gives 40 °C as the supported temperature for the GB300 board.
03 — Memory

Memory built onto the processor board — no DIMM slots

All memory is soldered and there are no DIMM slots. Each GPU has 288 GB of HBM3e, which multiplies to 1,152 GB per sled. For the Grace LPDDR5X, Dell’s feature tables and spec sheet give 480 GB per CPU — 960 GB per sled if multiplied, a total neither document prints — while the opening summary states 480 GB with no per-CPU qualifier, which can be read as 480 GB for the whole sled.

Integrated memory

fixed at the factory

480 GB LPDDR5X (ECC) per CPU, solder-down on each GB300 HPM (2 HPMs per sled) + 288 GB HBM3e per GPU

  • Wording by section: the chapter-one summary gives 480 GB of LPDDR5 with ECC and no qualifier; Tables 1 and 2 and the spec sheet say 480 GB per CPU; the memory chapter speaks of 480 GB of solder-down memory on the GB300 HPM. Confirm the per-sled total with Dell.
  • No LPDDR5X speed is given; Table 2 mentions “DIMM Speed” without a figure.
04 — Storage

1 drive-bay layout, up to 8 bays

Eight hot-swap EDSFF E1.S NVMe bays (S0–S7) run across the front of the sled, and the operating system lives apart from them on one M.2 2280 NVMe drive on a riser. Dell’s guide names a single BlueField-3 B3240 as the storage controller for the eight bays; its spec sheet instead describes direct NVMe through the CX8 mezzanine’s PCIe.

Headline layouts

front panel view

Front panel

8 × EDSFF NVMe at the front — 8 bays in total.

SAS/SATANVMeSAS/SATA + NVMe

Controllers & boot

  • M.2 Riser Card (boot) — Boot. 1 x M.2 2280 NVMe Gen4 (16 GB/s), RAID not supported, OS on M.2
  • NVIDIA BlueField-3 B3240 DPU (Storage Controller per Table 2) — HBA. Table 2 lists 'Storage Controllers: BF3-B3240 x 1'; BF3-B3240 x 1 supports eight E1.S drives on slots S0-S7. Spec sheet instead says 'Direct NVMe using CX8 mezz PCIe'. No PERC/HBA offered.
  • M.2 boot riser (1 x M.2 2280 NVMe Gen4 SSD, no RAID; thermal config lists 960 GB M.2 NVMe x 1) — boot
  • Every E1.S slot must hold a drive or a blank. Dell’s thermal table lists two configurations: two C8180 cards with two Gen5 E1.S drives in S0 and S1, or one B3240 in slot 2 with eight Solidigm PS1010 E1.S drives — and the guide says the B3240 supports all eight slots.
  • The M.2 riser takes one 2280 NVMe Gen4 drive (16 GB/s) with no RAID; the thermal table lists a 960 GB unit.
  • No PERC or HBA is offered, and Dell’s guide names no E1.S capacities.
05Accelerators & I/O

GPUs, PCIe and networking

Each sled carries four NVIDIA Blackwell Ultra GPUs, two on each GB300 board, with 288 GB of HBM3e and a 1400 W TDP apiece; Dell gives 2900 W as the total TDP for each board. GPU-to-GPU traffic leaves the sled through four rear NVLink5 connectors to what Dell calls the Non-Scalable NVS array, so the NVLink switching sits in the rack rather than in the sled. 4 × Blackwell Ultra per sled, 288 GB HBM3e and 1400 W each · up to 72 per rack (spec sheet).

NVIDIA Blackwell Ultra (GB300) 288GB HBM3e

288GB · 4 per sled; up to 72 per IR9048 rack (NVL72, spec sheet)

I/O topology

2 sockets
CPU 1PCIe 5.0 / 6.00 DIMMCPU 2PCIe 5.0 / 6.00 DIMM6 cabled C-Links (CPU–CPU)Front riser: 1 or 2 FHHLB3240 DPU, or 2 × C8180 CX82 × CX8 mezzanineGen5/Gen6 · 2 × 800G OSFP each1 GbE LOM · system LANon the front PCBOpenBMC · AST2600 BMCRJ45 BMC · HMC for GB300 boards

Apart from the rear NVLink, busbar and coolant connections, all I/O is at the front: two CX8 mezzanines with four OSFP cages, a front riser bay for PCIe cards, a 1 GbE LOM and the DC-SCM management ports. Dell describes the PCIe as Gen5/Gen6 at 32 GT/s, with the mezzanines on Gen5/Gen6 links.

Network adapters Dell lists

5 options
AdapterPortsForm
NVIDIA ConnectX-8 dual mezzanine (2 x 800 Gb/s OSFP each; 2 mezz = 4 x OSFP; IB (InfiniBand) NDR and XDR with GPU direct Gen6 x16, Super-NIC Mezzanine)2 × 800Gb/sMezzanine (Gen5/Gen6 x16)
NVIDIA BlueField-3 B3240 SuperNIC/DPU FHHL (up to 400 Gb/s QSFP112, up to 150 W, slot 2) × 400Gb/sPCIe
NVIDIA ConnectX-8 PCIe card C8180 (OSFP; up to 2, slots 1 and 2) × 800Gb/sPCIe
1 GbE LOM (system LAN, front PCB)1 × 1GbELOM
BMC management LAN RJ451 × 1GbEembedded
  • Table 2 lists one FHHL Gen5 slot on the front riser and Table 9 a single RF2 cable riser (CPU 1, Gen5 x16, 150 W); the PCIe overview and the spec sheet say up to two FHHL x16 slots. The thermal table covers both: one B3240 in slot 2, or C8180 cards in slots 1 and 2.
  • One DPU per sled at most, up to 150 W, with an auxiliary power cable.
  • Transceivers are specified: an NVIDIA dual QSFP112 (MMA1Z00-NS400) for the BlueField-3, and one of four listed dual-OSFP modules for the CX8 C8180 card.
  • The BlueField-3 B3240 supports up to 400 Gb/s QSFP and brings its own RJ45 management port to the front panel.
  • Only Dell-validated GPUs, NICs and PCIe devices are supported; unvalidated hardware voids the warranty.
06 — Power & cooling

No supplies of its own — powered by the rack

No power supplies sit inside the XE9712. IR9048 power shelves (PS33, 33 kW, six 5500 W AC supplies each) drive the rack busbar, and the sled draws up to 54 VDC from it through a rear connector. Budget from Dell’s board total: two GB300 boards at 2900 W each is 5,800 W per sled by our multiplication, before fans, drives and network cards.

050010001500200025003000350040004500500055005500 W 97.5% peak / 96.5% at full load (no 80 PLUS class stated) — Rack-scale power: sled has NO internal PSUs; it draws 54 V DC from the IR9048 rack busbar. IR9048 Power Shelf PS33 = 33 kW output (49-50 V, 674 A) from 6 x 5500 W AC PSUs, 7-pin 60 A AC input, hot-pluggable Power Management Controller; upstream 60 A/phase breaker recommended. A Delta Power Capacitance Shelf (52 V, 15 kW DCH/CHG @ 5 s, 1RU, BK500 500 A connector) buffers the busbar.5500DCWatts · dashed line = output at 100–120 V low line
AC
  • Each shelf delivers 33 kW at 674 A (49 V no-load, 50 V full-load, as printed) from supplies rated at 97.5% peak and 96.5% full-load efficiency, with a hot-pluggable Power Management Controller.
  • Voltage wording varies: the shelf overview says 54 V DC output, Table 4 prints 49 V / 50 V, the output clip mates with 48 V busbars, and the capacitance shelf is rated 52 V.
  • Each shelf takes AC through a 7-pin 60 A connector and has no fuses of its own, so Dell advises a 60 A per-phase breaker ahead of it.
  • A Delta power capacitance shelf (1RU, 52 V, BK500 500 A connector) charges from and discharges into the busbar at 15 kW for five seconds and is hot-swappable.
  • The XE9712 guide does not publish a shelf-count table for the IR9048.
  • Cooling is liquid-to-liquid: in-rack CDUs (Vertiv XDU100B or 121, Delta LCHD124A4QA) or in-row CDUs (Motivair MCDU-50, Vertiv XDU 1350) circulate coolant to the NVLink switches and GPU sleds, and eight fans handle the air-cooled parts.
  • The environmental appendix lists operating category A2 (10–35 °C, derated above 900 m), while the TDP table gives 40 °C for the GB300 board.
07 — Management, security & OS

Run it, secure it, choose the OS

Management

  • OpenBMC on an Aspeed AST2600 BMC (DC-SCI compatible); front I/O on a DC-SCM card
  • BMC-to-NVIDIA HMC link for managing the GB300 boards
  • Front ports: dedicated RJ45 BMC, BIOS serial out, Mini-DisplayPort, USB 3.0 Type-A
  • Dell Integrated Rack Controller: leak detection, fluid isolation and power shut-down via OpenManage Enterprise; eRDHx cooling control
  • Power shelves: hot-pluggable Power Management Controller
  • Spec sheet: an “XE9712 with iDRAC” version with iDRAC 10, a USB 2.0 Type-C port and a dedicated iDRAC RJ45
  • Matrox G200 video with 16 MB, up to 1920 × 1200 at 60 Hz

Security

  • TPM module on the board
  • NVIDIA BMC and reference boot firmware
  • Silicon-based root of trust, multi-factor authentication and role-based access (spec sheet)
  • Only Dell-validated GPUs, NICs and PCIe devices supported
  • Dell data sanitisation to NIST SP 800-88 r1, or physical destruction, at retirement

Operating systems

  • GraceOS — NVIDIA-qualified OS, the only OS in the guide
  • Ubuntu (spec sheet)
  • No Windows Server, VMware ESXi or RHEL listed
08 — Compare

How the XE9712 compares with its family

The XE9712 is the NVL72-class member of Dell’s liquid-cooled XE line. Its nearest relative, the XE8712, is a GB200 NVL4 tray for the 21-inch IR7000 rack, with four GPUs, more PCIe slots and only two E3.S bays; the 3 OU XE9780L and XE9785L nodes, also for IR7000, are built around eight-GPU boards; the XE9680L and XE9780 are self-contained HGX servers, 4U liquid-cooled and 10U air-cooled. Bars come from each spec file.

XE9712XE8712XE9780L/XE9780LAPXE9785LXE9680LXE9780Cores / CPUXE9712: 7272XE8712: 7272XE9780L/XE9780LAP: 128128XE9785L: 192192XE9680L: 6464XE9780: 8686DIMM slotsXE9712: 00XE8712: 00XE9780L/XE9780LAP: 3232XE9785L: 2424XE9680L: 3232XE9780: 3232Max memoryXE9712: 0.469 TB0.469XE8712: 0.938 TB0.938XE9780L/XE9780LAP: 4 TB4XE9785L: 3 TB3XE9680L: 4 TB4XE9780: 4 TB4Max drive baysXE9712: 88XE8712: 22XE9780L/XE9780LAP: 1616XE9785L: 1616XE9680L: 88XE9780: 1616Max GPUsXE9712: 44XE8712: 44XE9780L/XE9780LAP: 88XE9785L: 88XE9680L: 88XE9780: 88Largest PSUXE9712: 5,500 W5,500XE8712: 5,500 W5,500XE9780L/XE9780LAP: 33,000 W33,000XE9785L: 33,000 W33,000XE9680L: 3,000 W3,000XE9780: 3,200 W3,200
This modelSiblings and successorEach value from that model’s own verified spec file · bars scale per column
09 — Use cases

Three builds that suit the XE9712

Three sled builds. The configurator carries only this sled’s fixed complement — Grace pair, LPDDR5X, four GB300 GPUs, the CX8 mezzanines and the E1.S backplane — so each build notes what the IR9048 rack, the NVLink switch trays and the front-riser card add at quote stage.

Large-scale training

Multi-rack training sled

LLM training and large-scale AI training both appear on Dell’s workload list, and its spec sheet pitches the platform at multi-rack GPU clusters with high-speed east-west traffic. Inside a rack the GPUs talk over NVLink; between racks, the two CX8 mezzanines put four 800 Gb/s OSFP ports on InfiniBand NDR or XDR, and two CX8 C8180 cards in the front riser add more. Dell validates that card pair with two E1.S drives, so this build fits two.

  • 2 × NVIDIA Grace, 72 cores each, one per GB300 board
  • 2 × 480 GB LPDDR5X, soldered
  • 4 × NVIDIA Blackwell Ultra (GB300), 288 GB HBM3e each
  • CX8 dual mezzanine — 4 × 800 Gb/s OSFP
  • 2 × 7.68 TB EDSFF E1.S NVMe Gen5 in S0–S1, other slots blanked
  • Canonical Ubuntu Server 24.04 LTS
  • ProSupport Plus, 5 years

The two C8180 cards and their NVIDIA dual-OSFP transceivers, the rack and the NVLink switch trays are quoted separately. Ubuntu comes from the spec sheet, which gives no release; the guide lists only GraceOS.

Open this build in the configurator →
LLM inference

Large-model inference sled

LLM inference is the second workload Dell lists. Each sled holds 1,152 GB of HBM3e across its four GPUs, and Dell’s tables give each Grace 480 GB of LPDDR5X over coherent NVLink-C2C. A BlueField-3 B3240 in slot 2 adds up to 400 Gb/s of QSFP networking beside the four mezzanine ports, and Dell’s thermal table pairs that card with a full set of eight E1.S drives, which hold local model copies.

  • 2 × NVIDIA Grace, 72 cores each
  • 2 × 480 GB LPDDR5X, 480 GB per CPU in Dell’s tables
  • 4 × NVIDIA Blackwell Ultra (GB300) — 1,152 GB HBM3e in total
  • CX8 dual mezzanine — 4 × 800 Gb/s OSFP
  • 8 × 7.68 TB EDSFF E1.S NVMe Gen5, all slots filled
  • Canonical Ubuntu Server 22.04 LTS
  • ProSupport Plus, 3 years

The BlueField-3 B3240 and its NVIDIA MMA1Z00-NS400 QSFP112 transceiver are quoted separately; Dell’s thermal table puts the DPU in slot 2, a sled takes only one, and the validated eight-drive set names Solidigm PS1010 units, which our configurator doesn’t list.

Open this build in the configurator →
Rack-scale training

Single-rack NVL72 training sled

Dell also lists rack-scale architecture and the NVL72 design itself among the sled’s workloads. When one rack is the whole job, GPU-to-GPU exchanges stay on NVLink — 1.8 TB/s in Dell’s spec sheet — and every sled adds eight E1.S bays of local NVMe; this build fills all eight rather than blanking any.

  • 2 × NVIDIA Grace, 72 cores each
  • 2 × 480 GB LPDDR5X, soldered
  • 4 × NVIDIA Blackwell Ultra (GB300)
  • CX8 dual mezzanine — 4 × 800 Gb/s OSFP
  • 8 × 3.84 TB EDSFF E1.S NVMe Gen5
  • No operating system from us
  • ProSupport, 5 years

No operating system is included: Dell’s guide lists GraceOS, an NVIDIA-qualified OS, and the spec sheet Ubuntu. The guide ties the full eight-drive set to a BlueField-3 B3240 in slot 2, which is quoted separately. Dell recommends ProSupport Plus for PowerEdge XE systems, so weigh that against the plain ProSupport shown.

Open this build in the configurator →
10 — Full specifications

Dell PowerEdge XE9712 specifications

Sled & cooling

Form factor1U (1 RU) two-socket compute sled for the Dell IR9048 48 RU ORv3 rack
Dimensions43.6 × 438 × 807 mm (guide) · 481.54 mm width at Xa (appendix) · spec sheet depth 836.8 mm
Weight30 kg with 8 × E1.S
CoolingDirect liquid cooling for CPUs and GPUs (combined cold plates with leak sensors) plus air · liquid-to-liquid CDU
FansEight · Delta, AVC or Foxconn · up to 45.64 CFM each
RailsStatic rails for the ORv3 IR9048

Processor & memory

Processors2 × NVIDIA Grace (Grace Ultra), 72 cores each, up to 3.6 GHz
TDP500 W maximum (Table 1) · 300 W per CPU in the GB300 TDP table
CPU–CPU6 cabled C-Links, fixed and coherent
CPU memoryLPDDR5X with ECC, soldered · 480 GB per CPU in Tables 1–2 and the spec sheet; 480 GB with no per-CPU qualifier in the summary and memory chapter

Accelerators

GPUs4 × NVIDIA Blackwell Ultra per sled · 2 per GB300 board
GPU memory288 GB HBM3e per GPU
GPU TDP1400 W per GPU · 2900 W total per GB300 board (Dell’s figure; its CPU and GPU entries add to 3,100 W)
CPU–GPUNVLink-C2C, 900 GB/s coherent memory · up to 10 links (Table 2) or up to four (appendix)
GPU–GPU4 rear NVLink5 connectors to the NVS array · up to 72 GPUs per rack, 1.8 TB/s GPU-to-GPU (spec sheet)

Storage

Front bays8 × EDSFF E1.S NVMe, hot-swap (S0–S7) · each filled or blanked
ControllerBlueField-3 B3240 (guide) · direct NVMe via the CX8 mezzanine (spec sheet)
BootM.2 riser: 1 × M.2 2280 NVMe Gen4, no RAID

Expansion & networking

PCIeGen5/Gen6 at 32 GT/s · 1 or 2 FHHL x16 front slots (documents differ)
Card options1 × BlueField-3 B3240 in slot 2 (with 8 × E1.S), or 2 × CX8 C8180 in slots 1–2 (with 2 × E1.S)
Mezzanine2 × CX8, 2 × 800 Gb/s OSFP each · InfiniBand NDR/XDR, GPU direct Gen6 x16
Other LAN1 GbE LOM · RJ45 BMC management

Power (rack-level)

Sled PSUsNone — up to 54 VDC from the rack busbar
Power shelfPS33: 33 kW, 674 A, 6 × 5500 W AC PSUs · 97.5% peak / 96.5% full load
Shelf input7-pin 60 A AC · 60 A per-phase upstream breaker recommended
Capacitance shelfDelta, 1RU, 52 V · 15 kW charge/discharge for 5 s · BK500 500 A
VoltagesSled input up to 54 VDC · shelf output 54 V (overview) or 49/50 V (Table 4) · 48 V busbar clip · 52 V capacitance shelf

Rack (IR9048)

Size48 RU · 2286 × 750 × 1200 mm (1340 mm deep with stabilisers) · spec sheet height 2294 mm
Rack standardORv3 in the guide’s feature table; the spec sheet calls the IR9048 a 19-inch rack, and the guide gives 1 RU widths of 482.6 mm and 538.98 mm
Wet weight1590 kg
CDUsIn-rack: Vertiv XDU100B or 121, Delta LCHD124A4QA · in-row: Motivair MCDU-50, Vertiv XDU 1350
Other itemsDell SN2201-ON management switch · Vertiv VP4UU1A0 rPDU · Integrated Rack Controller

Ports, environment & warranty

Front1 × USB 3.0 Type-A · BIOS serial out · RJ45 BMC · Mini-DisplayPort · 1 GbE LOM · 4 × OSFP
VideoMatrox G200, 16 MB
EnvironmentCategory A2, 10–35 °C · 40 °C supported for the GB300 board · coolant: Recochem PG25 only
WarrantyThree years hardware, standard for PowerEdge XE
11 — FAQ

XE9712 questions, answered from the documentation

What is the Dell PowerEdge XE9712?

A 1U liquid-cooled compute sled for NVIDIA’s GB300 NVL72 architecture, supplied inside Dell’s 48 RU IR9048 Open Rack v3 cabinet. Each sled carries two GB300 boards, each pairing one Grace CPU with two Blackwell Ultra GPUs, and it depends on the rack for DC power, liquid cooling and GPU-to-GPU NVLink.

How many XE9712 sleds are in an NVL72 rack?

Neither Dell’s technical guide nor its spec sheet prints a sled count. The spec sheet puts up to 72 Blackwell Ultra GPUs in one NVLink rack, and at four GPUs per sled that works out to 18 sleds — our arithmetic, not a Dell figure. NVLink switch trays, power shelves, a capacitance shelf, CDUs and a management switch fill the rest of the rack.

How much GPU and CPU memory does each sled have?

Each of the four GPUs has 288 GB of HBM3e, 1,152 GB per sled by multiplication. Dell’s feature tables and spec sheet give 480 GB of soldered LPDDR5X with ECC per Grace, which would be 960 GB per sled, but the guide’s summary quotes 480 GB without saying per CPU and never prints a sled total, so confirm the figure with Dell. There are no DIMM slots.

How much power does an XE9712 sled use?

Dell’s TDP table gives 2900 W as the total for each GB300 board (it also lists 300 W for the CPU and 1400 W for each of the two GPUs, which add up to more), and a sled holds two boards, so 5,800 W by our multiplication before fans, drives and network cards. The sled has no power supplies; it takes up to 54 VDC from the IR9048 busbar, fed by 33 kW shelves of six 5500 W AC supplies each.

Is the XE9712 liquid- or air-cooled?

Both. The CPUs and GPUs sit under combined cold plates with leak sensors, plumbed through inner manifolds to quick-disconnects at the rear that join the rack manifold, and a liquid-to-liquid CDU exchanges the heat with the facility. Eight fans cool the remaining components.

What networking does the XE9712 have?

Two CX8 mezzanines, each with two 800 Gb/s OSFP ports, support InfiniBand NDR and XDR with GPU direct over Gen6 x16. The front riser takes either one BlueField-3 B3240 DPU at up to 400 Gb/s or two CX8 C8180 cards, and there is a 1 GbE LOM plus a dedicated BMC port.

What storage does the XE9712 support?

Eight hot-swap EDSFF E1.S NVMe bays at the front and one M.2 2280 NVMe boot drive on a riser, without RAID. Every E1.S slot must hold a drive or a blank, and Dell’s thermal table pairs two drives in S0–S1 with two C8180 cards, or eight Solidigm PS1010 drives with the B3240.

Which operating system and management does the XE9712 use?

The guide lists GraceOS, an NVIDIA-qualified OS, and Dell’s spec sheet lists Ubuntu. The standard sled is managed through OpenBMC on an Aspeed AST2600 BMC with an NVIDIA HMC link to the GB300 boards; the spec sheet also describes an “XE9712 with iDRAC” version with iDRAC 10.

How does the XE9712 differ from the XE8712?

The XE8712 is a GB200 NVL4 tray for Dell’s 21-inch IR7000 rack, with 192 GB GPUs, up to four FH HL PCIe slots, iDRAC10 and two E3.S bays. The XE9712 is a GB300 NVL72 sled for the IR9048, with 288 GB GPUs, rear NVLink5 connectors to the rack’s NVS array, CX8 mezzanines at 800 Gb/s and eight E1.S bays.

Is the XE9712 end of life?

No end-of-sale or end-of-service date appears in Dell’s November 2025 guide. PowerEdge XE systems carry a standard three-year hardware warranty, and Dell recommends ProSupport Plus for them. Our end-of-life checker will list dates once Dell publishes any.

Sources & verification

Where these figures come from

  • Dell PowerEdge XE9712 Technical Guide (Rev. A00, November 2025) — Tables 1–2 (features), Figures 1–3 and Table 3 (views; gallery images from PDF pages 8–9), power shelf (Table 4), Integrated Rack Controller (Table 5), switch and rPDU (Tables 6–7), memory, M.2 (Table 8), PCIe (Tables 9–11), DPU, CDU, capacitance shelf (Table 12), networking, thermal and TDP (Tables 13–16), OS, processor, dimensions and weight (Tables 17–19), GPU, video (Table 20) and environmental (Tables 21–22) appendices, service appendix.
  • Dell PowerEdge XE9712 Specification Sheet (April 2026) — Up to 72 GPUs per rack and 1.8 TB/s GPU-to-GPU NVLink, 1:1 NIC-to-GPU, multi-rack positioning, “XE9712 with iDRAC” variant, OS (Ubuntu), PCIe and storage-controller wording, sled depth, rack height and 19-inch wording.
  • Dell PowerEdge IR7000, XE8712 Technical Guide (Rev. A01, January 2025) — XE8712 facts used in the comparison FAQ, family text and verdict.

Specifications are extracted from Dell’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 Dell PowerEdge XE9712

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.

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

Enquiring about: Dell PowerEdge XE9712

Response within 4 business hours · No obligation

Quote for
Dell PowerEdge XE9712
0800 987 4111

Talk to a UK specialist

Get expert advice or a no-obligation quote — servers, storage, networking, maintenance, finance and cloud. We reply the same working day.

or call 0800 987 4111