The XE9785 is Dell’s air-cooled 10U server for eight AMD Instinct MI355X or NVIDIA HGX B300 GPUs, run by two AMD EPYC 9005 processors in a pull-out host sled. It suits teams training or fine-tuning large models in rooms without a liquid loop. This sheet maps both GPU builds, what each unlocks or rules out, and the power and weight the chassis brings with it.
Our read of Dell’s documentation: where the XE9785 is strong, and the limits worth knowing before you spec one.
Strengths
+One chassis, either GPU vendor: eight 288 GB MI355X OAM modules on a Universal Base Board, or an eight-GPU HGX B300 NVL8 board fully linked by NVLink, its NVSwitch under a separate heat sink.
+Cooled entirely by air — five CPU fans on the mid tray and fifteen hot-swap GPU fans at the rear, N+1 redundant — with no coolant plumbing involved.
+The MI355X build opens twelve Gen5 x16 slots at 75 W each for network cards and DPUs; the B300 build keeps four slots and adds eight embedded OSFP ports.
+Twelve 3200 W Titanium supplies run 6+6 fully redundant or A/B grid redundant, and Dell quotes 96% efficiency at half load.
+Sixteen E3.S Gen5 NVMe bays at the front plus a hardware-RAID M.2 NVMe boot pair with its own heat sinks.
Watch-outs
!Very heavy and tall: 172.3 kg fully built with MI355X, 163.65 kg with B300, in a 439.5 mm-high 10U chassis that Dell says calls for a server lift.
!The 256 GB DIMM — and so the 6 TB ceiling — belongs to B300 builds only; an MI355X server stops at 3 TB with 24 × 128 GB.
!Picking B300 shrinks the PCIe budget to four 150 W slots: 3, 6, 9 and 12.
!High-line input only — Dell gives no low-line rating — and if PSU failures outrun redundancy in PSU-redundant mode, the GPUs are braked to a quarter of their clock.
!Launch software is narrow: Ubuntu Server LTS, with Red Hat and SUSE both flagged as unavailable at launch; the platform also lacks a user power cap and Sound Cap.
Processors8 SKUs
Cores / CPU48–192
DIMM slots24
Drive layouts2
Max bays16
GPUsup to 8
PSU options1 · to 3,200 W
02 — Processors
8 supported processors, mapped
Eight 5th Gen EPYC 9005 models are supported, always in pairs and none flagged as unavailable at launch, spanning 48 cores and 300 W (9455) up to 192 cores and 500 W (9965). Each is plotted by core count and TDP exactly as this guide prints them.
Processor landscape — cores × TDP
8 SKUs · 1 family
AMD EPYC 9005 8Bubble size = L3 cache · hover a bubble for the full SKU
Processor explorer
filter · sort · compare
Processor
Cores / threads
Base
L3 cache
TDP
W / core
Memory
AMD EPYC 9965
192 / 384
2.25 GHz
384 MB
500 W
2.6
DDR5-6400
AMD EPYC 9845
160 / 320
2.1 GHz
320 MB
390 W
2.4
DDR5-6400
AMD EPYC 9825
144 / 288
2.2 GHz
384 MB
390 W
2.7
DDR5-6400
AMD EPYC 9755
128 / 256
2.7 GHz
512 MB
500 W
3.9
DDR5-6400
AMD EPYC 9655
96 / 192
2.6 GHz
384 MB
400 W
4.2
DDR5-6400
AMD EPYC 9575F
64 / 128
3.3 GHz
256 MB
400 W
6.3
DDR5-6400
AMD EPYC 9475F
48 / 96
3.65 GHz
256 MB
360 W
7.5
DDR5-6400
AMD EPYC 9455
48 / 96
3.15 GHz
256 MB
300 W
6.3
DDR5-6400
All eight run memory at 6400 MT/s in Dell’s table, and each carries the same 6 TB memory-capacity entry.
Dell’s platform summary quotes up to 160 xGMI lanes, 5 nm process technology and a 500 W TDP ceiling.
In the processor feature list, each Zen 5 CCD shares up to 32 MB of L3 among 16 cores, with 1 MB of L2 per core.
The two F-suffix parts top the maximum-clock column — 5.0 GHz for the 9575F and 4.8 GHz for the 9475F — with the 9475F printed at a 3.65 GHz base and 360 W.
03 — Memory
24 DIMM slots, 12 per processor
Twenty-four DDR5 RDIMM slots — twelve channels per socket, one DIMM on each — so a full server never shares a channel. Three module sizes are listed, 96, 128 and 256 GB, all rated at 6400 MT/s, and the GPU board decides which of them you may fit.
Capacity with every slot filled
by DIMM size
Rated memory speed
from the processor table
Processor family
Up to
AMD EPYC 9005
6400 MT/s
Dell’s 6 TB maximum is 24 × 256 GB, and that module is supported only in the B300 GPU configuration. With MI355X the biggest module is 128 GB, so 24 slots top out at 3 TB (3072 GB).
Only registered ECC modules at 1.1 V are accepted; unbuffered DIMMs are not supported.
Processor model and population can hold the running speed below 6400 MT/s, Dell notes.
The 96 GB and 128 GB modules are dual-rank x4; the 256 GB module is eight-rank x4.
04 — Storage
2 drive-bay layouts, up to 16 bays
Local storage is small and fast by design: 16 E3.S Gen5 NVMe drives attached without a RAID card (configuration C01), and a 10-drive 2.5-inch U.2 Gen5 layout (C02) that Dell marks as unavailable at the November 2025 launch. No PERC, no SAS or SATA, no rear bays.
Headline layouts
front panel view
Front panel
16 × EDSFF NVMe at the front — 16 bays in total.
SAS/SATANVMeSAS/SATA + NVMe
Controllers & boot
Software RAID S160 — Software RAID. Listed as the storage controller for both drive configs C01/C02 (PERC Qty 0); direct-attach PCIe SSDs only
Embedded BOSS — Boot. HWRAID 1, 2 x M.2 NVMe SSDs. Dedicated PERC: N/A
Embedded BOSS (2 x M.2 NVMe SSD, HWRAID 1, dedicated heat sinks and thermal pads) — boot
Dell names its S160 software RAID as the storage controller for both layouts, and both need the two processors fitted.
The operating system boots from an embedded BOSS: two M.2 NVMe SSDs in hardware RAID 1, each with a dedicated heat sink and thermal pad.
Plan new builds on the E3.S layout; Dell asks buyers to confirm the U.2 option on its configurator before counting on it.
Bulk datasets belong on external storage — the chassis offers no 3.5-inch or SAS path, only direct-attached PCIe SSDs.
Every documented layout
2 with drive bays
2.5-inch3.5-inchEDSFF (E1.S / E3.S)
05 — Accelerators & I/O
GPUs, PCIe and networking
The GPU sled is why this server exists. The AMD version carries a Universal Base Board with eight Instinct MI355X OAM modules, 288 GB and 1400 W each, linked by AMD Infinity Fabric. The NVIDIA version carries an HGX B300 NVL8 baseboard — eight 270 GB SXM6 GPUs at 1100 W joined by NVLink — with the NVSwitch under its own heat sink and two OSFP boards with cable assemblies in the sled. Eight GPUs either way: 8 × 1400 W is 11.2 kW of rated MI355X GPU power and 8 × 1100 W is 8.8 kW for B300, all air-cooled in 10U.
Dell credits each EPYC 9005 with as many as 128 high-speed I/O lanes; the expansion slots sit behind a Broadcom PCIe Gen5 switch, twelve x16 slots at most. How many you actually get is set by the GPU board: MI355X builds use a PCIe Switch Base Board (PSBB), B300 builds a PCIe Switch Retimer Base Board (PSRBB).
Network adapters Dell lists
15 options
Adapter
Ports
Form
Broadcom 2 x 100 GbE Q56 (QSFP56) OCP 3.0
2 × 100GbE
OCP 3.0
Broadcom 2 x 200 GbE QSFP OCP 3.0
2 × 200GbE
OCP 3.0
NVIDIA 2 x 100 GbE QSFP OCP 3.0
2 × 100GbE
OCP 3.0
Intel 2 x 100 GbE Q56 (QSFP56) OCP 3.0 (*not at initial launch)
2 × 100GbE
OCP 3.0
Broadcom 4 x 25 GbE QSFP OCP 3.0
4 × 25GbE
OCP 3.0
Intel 4 x 25 GbE S28 (SFP28) OCP 3.0 (*not at initial launch)
4 × 25GbE
OCP 3.0
Intel 2 x 25 GbE S28 (SFP28) OCP 3.0 (*not at initial launch)
2 × 25GbE
OCP 3.0
NVIDIA 2 x 25 GbE QSFP OCP 3.0
2 × 25GbE
OCP 3.0
Broadcom 2 x 25 GbE QSFP OCP 3.0
2 × 25GbE
OCP 3.0
Intel 2 x 10 GbE BT (base-T) OCP 3.0 (*not at initial launch)
2 × 10GbE
OCP 3.0
Broadcom 2 x 10 GbE QSFP OCP 3.0
2 × 10GbE
OCP 3.0
DPU card 2 x 400 GbE (*not at initial launch) / 2 x 200 GbE (Table 27; no vendor/model named)
2 × 400GbE / 200GbE
PCIe
PCIe add-in card NIC 1 x 400 GbE / 2 x 200 GbE / 2 x 100 GbE (Table 27; no vendor/model named)
× 400GbE / 200GbE / 100GbE
PCIe
Embedded 8 x OSFP ports (B300 GPU configuration only; OSFP boards with cable assembly on the GPU sled)
8 ×
embedded
Front I/O board 2 x RJ45 1 GbE dedicated iDRAC Ethernet ports
2 × 1GbE
embedded
The PCIe chapter gives 75 W of card-edge power per slot, yet the B300 build is specified for 150 W cards in its four usable slots — confirm the power arrangement with Dell for any card above 75 W.
There is a single riser configuration (0, no riser): two processors, no PERC and no rear storage.
Table 14 lists the OCP 3.0 slot as Gen4 x8 with a 35 W budget, up to four ports and NC-SI, SNAPI and WoL; the NIC comparison table elsewhere in the guide says OCP 3.0 runs up to x16.
Dell’s port table covers OCP cards from 10 to 200 GbE, PCIe NICs at 1 × 400, 2 × 200 or 2 × 100 GbE, and DPUs at 2 × 200 GbE, with 2 × 400 GbE DPUs marked as not available at launch.
There is no LOM: apart from the B300 board’s OSFP ports, the only built-in network ports are the two 1 GbE iDRAC RJ45 ports on the front I/O.
06 — Power & cooling
1 power supply, 3,200 W
Twelve 3200 W Titanium supplies sit below the rear fan sled, configured as 6+6 full redundancy — which Dell says preserves full performance — or as A/B grid redundancy chosen in iDRAC. Input must be high-line: 3200 W from 216.1–240 VAC or 240 VDC, 2900 W from 200–216 VAC, and no low-line rating at all.
AC
Each supply takes a C19 lead, has a 5280 W peak rating and is quoted at 90, 94, 96 and 91% efficiency at 10, 20, 50 and 100% load.
In PSU-redundant mode, losing more supplies than the redundancy covers engages the GPU power brake: clocks fall to a quarter and performance to roughly a fifth. When PSU redundancy is not the chosen policy, the brake does not fire automatically.
IT-type power distribution is supported too, provided the phase-to-phase voltage stays at or below 240 V. Dell offers no user power cap, and PSUs from two vendors may not share one chassis.
Heat dissipation is 11,850 BTU/hr per supply at 3200 W. For real system draw Dell points to its Enterprise Infrastructure Planning Tool and the live figures in iDRAC.
07 — Management, security & OS
Run it, secure it, choose the OS
Management
iDRAC10 (Datacenter licence in the support matrix)
iDRAC Direct on USB-C (Dell USB-C to 2.5 Gbps Ethernet adapter, DP/N HYRGV)
iDRAC RESTful API with Redfish
RACADM CLI and IPMI
iDRAC Service Module
OpenManage Enterprise with Power Manager, Service and Update Manager plug-ins
CloudIQ for PowerEdge plug-in
BMC TrueSight, ServiceNow, Red Hat Ansible and Terraform integrations
Security
AMD Secure Encrypted Virtualization (SEV) and Secure Memory Encryption (SME)
Source: Dell PowerEdge XE9785 Technical Guide (Rev. A03, May 2026), Table 2 (comparison of PowerEdge XE9785 and XE9680).
The XE9785 family side by side
The XE9785 has a liquid-cooled twin, the XE9785L: a 3 OU node for Dell’s IR7044 and IR7050 racks with the same EPYC processors, both of these GPU boards and an HGX B200 option. The XE9780 and XE9780L are the 6th Gen Intel Xeon Scalable equivalents, sold with NVIDIA B300 or B200 boards only, and the 6U XE9680 is the model Dell measures it against.
This modelSiblings and successorEach value from that model’s own verified spec file · bars scale per column
Four eight-GPU builds, one for each decision this chassis forces: GPU vendor, host memory, core count and slot budget. Every part is permitted on the XE9785 in our configurator, and each build opens there ready to adjust.
Large-model training
Eight MI355X on 3 TB of host memory
The AMD board gives the bigger GPU memory pool of the two — 8 × 288 GB, 2304 GB in all — and leaves all twelve Gen5 slots free for adapters. Two 64-core 9575F processors boost to 5.0 GHz to keep data moving, and 24 × 128 GB is the most memory an MI355X build can hold.
2 × AMD EPYC 9575F — 64 cores @ 3.3 GHz, 400 W each
8 × AMD Instinct MI355X 288 GB OAM (1400 W each) on the UBB
24 × 128 GB DDR5-6400 RDIMM — 3 TB, one per channel
Recommendation engines and large language workloads are on Dell’s own list for this server, and only the B300 build accepts 256 GB modules, so 24 of them reach the 6 TB ceiling. Two 96-core 9655s at 400 W handle preprocessing, and the eight embedded OSFP ports arrive with the board rather than using a slot.
2 × AMD EPYC 9655 — 96 cores @ 2.6 GHz, 400 W each
Dell lists HPC compute and simulation among this server’s workloads. Two 192-core 9965s put 384 cores beside the GPUs for codes that split work between host and accelerator, at the 500 W top of the processor range, and 24 × 96 GB keeps one module on every channel.
2 × AMD EPYC 9965 — 192 cores @ 2.25 GHz, 500 W each
8 × AMD Instinct MI355X 288 GB OAM
24 × 96 GB DDR5-6400 RDIMM — 2304 GB
8 × 3.84 TB E3.S Gen5 NVMe — 30.72 TB raw
Embedded BOSS boot pair (RAID 1)
Broadcom 4 × 25 GbE OCP 3.0
12 × 3200 W Titanium (6+6)
Red Hat Enterprise Linux 9
Red Hat Enterprise Linux was not available when the XE9785 launched; confirm its status, or build on Ubuntu Server LTS.
Fine-tuning, chatbots and speech models appear in Dell’s workload list too. Two 48-core 9475F parts, printed at a 3.65 GHz base and 4.8 GHz boost, keep the host side quick at 360 W each, while 3 TB of DDR5 and eight E3.S drives cover datasets and checkpoints.
2 × AMD EPYC 9475F — 48 cores @ 3.65 GHz, 360 W each
NVIDIA HGX B300 NVL8 — 8 × 270 GB SXM6
24 × 128 GB DDR5-6400 RDIMM — 3 TB
8 × 7.68 TB E3.S Gen5 NVMe — 61.44 TB raw
Embedded BOSS boot pair (RAID 1)
Broadcom 2 × 25 GbE OCP 3.0
12 × 3200 W Titanium (6+6)
Canonical Ubuntu Server 24.04 LTS
Only four PCIe slots are usable with B300, so plan any NIC or DPU around slots 3, 6, 9 and 12.
NIC 1 × 400, 2 × 200 or 2 × 100 GbE · DPU 2 × 200 GbE, 2 × 400 GbE not at launch
Power
Supplies
12 × 3200 W AC Titanium, redundant · 6+6 full redundant or A/B grid redundant
Input
3200 W at 216.1–240 VAC or 240 VDC · 2900 W at 200–216 VAC · no low line · C19
Peak and heat
5280 W peak per supply · 11,850 BTU/hr
Limits
No user power cap · GPU power brake if PSU failures exceed redundancy (PSU-redundant mode)
Ports & video
Front
iDRAC Direct USB-C · 2 × RJ45 iDRAC · USB Type-A (USB 3.0) · Mini-DisplayPort
Rear
None — GPU fans and power supplies only
Internal
1 × USB 3.1 Type-A
Video
Integrated Matrox G200, 16 MB
11 — FAQ
XE9785 questions, answered from the documentation
What is the maximum memory of the PowerEdge XE9785?
Dell quotes 6 TB across 24 DDR5 RDIMM slots, which means 24 × 256 GB modules at one per channel. The 256 GB module is supported only with the NVIDIA B300 GPU board, so an MI355X server tops out at 3 TB using 128 GB modules.
Which GPUs does the XE9785 support?
One of two boards: eight AMD Instinct MI355X OAM modules (288 GB, 1400 W) on a Universal Base Board with Infinity Fabric, or an NVIDIA HGX B300 NVL8 board of eight 270 GB, 1100 W SXM6 GPUs connected by NVLink. The choice also fixes the PCIe switch board, the usable slots and whether the OSFP ports are present.
Is the XE9785 air-cooled or liquid-cooled?
Air-cooled. Five cold-swap fans cool the host sled and fifteen hot-swap fans in a rear fan sled cool the GPUs, with N+1 fan redundancy. For direct liquid cooling in Dell’s IR7000 racks, the XE9785L is the model to look at.
Which processors does the XE9785 take?
Two 5th Gen AMD EPYC 9005 processors from a list of eight: 9455, 9475F, 9575F, 9655, 9755, 9825, 9845 and 9965, spanning 48 to 192 cores and 300 to 500 W. Both sockets are always populated.
How many drive bays does the XE9785 have?
Sixteen E3.S Gen5 NVMe bays at the front, direct-attached and managed by S160 software RAID. A ten-bay 2.5-inch U.2 Gen5 layout is listed but marked as not available at launch. There are no SAS, SATA or rear bays, and boot uses a separate M.2 NVMe pair.
How many PCIe slots are available?
Up to twelve Gen5 x16 full-height, half-length slots with MI355X (slots 2–13, 75 W each), but only four with B300 (slots 3, 6, 9 and 12, up to 150 W). One OCP 3.0 slot comes on top, and the B300 board adds eight embedded OSFP ports.
What power does the XE9785 need?
Twelve 3200 W Titanium supplies, each fed by a C19 power cable, in 6+6 or A/B grid redundancy. Each delivers 3200 W at 216.1–240 VAC or 240 VDC and 2900 W at 200–216 VAC; Dell gives no low-line rating. IT-type power distribution is acceptable up to 240 V between phases.
How big and heavy is the XE9785?
It is 10U: 439.5 mm high, 448 mm wide (482.3 mm across the ears) and 1044.7 mm deep with the bezel. Fully populated it weighs 172.3 kg with MI355X or 163.65 kg with B300, and Dell recommends a server lift. It mounts on static rails with no cable management arm.
How does the XE9785 differ from the XE9780 and XE9785L?
The XE9780 is also a 10U air-cooled two-socket server, but built on 6th Gen Intel Xeon Scalable processors, with NVIDIA B300 or B200 boards and no AMD GPU option. The XE9785L keeps the EPYC processors and the MI355X and B300 boards, adds B200, and becomes a 3 OU liquid-cooled node installed in Dell’s IR7044 or IR7050 racks.
Is the Dell PowerEdge XE9785 end of life?
Not according to Dell’s Rev. A03 guide of May 2026, which publishes no retirement or support-end date for the model; several options in it are still marked as unavailable at the November 2025 launch. Dates will appear in our end-of-life checker when Dell releases them.
Sources & verification
Where these figures come from
Dell PowerEdge XE9785 Technical Guide (Rev. A03, May 2026) — Chapter 1 and Tables 1–2 (overview, new technologies, XE9785 vs XE9680); Tables 3–9 and Figures 1–7 (chassis views, HPM and GPU sleds; images from Figures 1, 3 and 5); Tables 10–17 (processors, memory, storage, OCP, PCIe); Tables 18–20 and 30–31 (power); thermal and acoustics chapter; Tables 22–29 (iDRAC, management, dimensions, weight, ports, video, USB).
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.
