The PowerEdge AI server line
S3·E1Nine adapters, four slots · Dell briefing centre, forty days to the customer's peak-season freeze
Before you read: what do you already know?
3 quick questions. Wrong answers are fine and expected; trying first makes the lesson stick.
After this lesson you can
- Identify the PowerEdge XE9680, XE9780 and XE9785 by form factor, accelerator options and network I/O.
- Explain where the east-west SuperNICs and the north-south DPU physically live in each chassis.
- Use the XE9780 slot-priority and thermal matrices to say which adapter is allowed in which slot.
- Describe the two documentation traps in Dell's own AI server material and state which source you would quote.
Episode 1 - Nine adapters, four slots
The Dell SE has row 41 of his promise spreadsheet on the screen and it does not fit. He costed nine network adapters per server, because that is the node the reference architecture describes: “8 on baseboard NVIDIA ConnectX-8 SuperNIC with a dual-port 400G configuration” plus “1 NVIDIA BlueField-3 B3240 DPU with 2x400G ports”.[7] The servers are XE9780s with eight B300 GPUs. The customer is a parcel carrier forty days from its peak-season change freeze, and its network lead, who trusts nothing, reads out the line that ends the argument: “Up to 4 x Gen5 x16 Full-height, Half-length PCIe slots with NVIDIA B300 GPU configuration”.[3] Nine adapters, four slots. The SE’s coffee goes cold.
You are the third person in the room, and the answer is not a smaller quote. The same features table lists Embedded OSFP as “8 x OSFP with B300 GPU”.[3] Eight of the nine adapters are already in the chassis, on the GPU baseboard - which is what the RA meant by “on baseboard”.[7][6]
That is why embedded OSFP exists. One 400G-class adapter per GPU used to mean one full-height card per GPU, and a chassis holding eight 1100 W accelerators has no room for eight more. Moving the east-west ports onto the baseboard deletes eight cards, eight risers and eight cable runs. The slot count did not shrink because the network shrank; the network moved.
Count the accelerator before you count the slots. The NVIDIA PM says the next generation is “not announced”. Which leaves three chassis, three generations and two tables.
1Three chassis, three generations
Dell’s AI server line is not a ladder of sizes, it is a ladder of accelerator generations, and the network I/O changes with each rung.
The PowerEdge XE9680 is a “6U rack server” and it is the widest-choice chassis in the line: “8 NVIDIA HGX H100 80GB 700W SXM5 GPUs”, “8 NVIDIA HGX H200 141GB 700W SXM5 GPUs”, “8 AMD Instinct MI300X 192GB 750W OAM accelerator” or “8 Intel Gaudi3 128GB 900W OAM accelerator with embedded RoCE ports for ethernet connectivity”.[1] It provides “Up to 10 x16 Gen5 (x16 PCIe) full-height, half-length” slots with the caveat “8 PCIe slots with Intel Gaudi3”, plus “1 x OCP 3.0 (optional)”, and it is powered by “2800W Titanium 200-240 VAC or 240 VDC redundant, hot swap” supplies.[1] Dell’s own comparison table describes the XE9680’s embedded OSFP as “6 x 800 Gb (available only with Intel Gaudi3)” - so on an NVIDIA build, every network adapter is a PCIe card.[3]
The PowerEdge XE9780 is “Dell’s latest 2-socket, 10U air-cooled rack server”, carrying “Two 6th Generation Intel Xeon Scalable processors with up to 86 cores per processor” and either “8 NVIDIA HGX B300 NVL8 270GB 1100W SXM6 GPUs” or “8 NVIDIA HGX B200 180GB 1000W SXM6 GPUs”.[3][2] Its power story is the visible jump: “12 x 3200W Titanium 200-240 VAC or 240 VDC, hot swap redundant” with a “Full redundant (6+6) FR (System can get full performance)” option.[2][3]
The PowerEdge XE9785 is the sibling chassis, also a “10U rack server”, offering “8 AMD Instinct MI355X 288GB 1400W OAM” or “8 NVIDIA HGX B300 NVL8 270 GB 1100W SXM6 GPUs”, with “1 x OCP 3.0 (Gen5 x8 PCIe lanes)” and “12 x 3200W Titanium 200-240 VAC, hot swap redundant” supplies.[4]
| PowerEdge | BF-3 offering | DPU mode | NIC mode | Aux power | vSphere DSE | Notable KB |
|---|---|---|---|---|---|---|
R660 16G | ||||||
R760 16G | ||||||
R760XA 16G (GPU) | ||||||
XE9680 16G HGX H100/H200 | ||||||
XE9680L 16G liquid | ||||||
R7725 17G (AMD) | ||||||
R770 17G (Intel) | ||||||
XE9780 / XE9785 17G HGX B300 — Dell AI Factory |
✓ yes · ✗ no · ◐ conditional · ? unknown · — n/a. Click a cell for the evidence.
⚠ = not confirmed on a fetched primary source (hover for why). Facts as of DOCA 3.5.0 (Sep 2026). Selections are saved.
2Where the network actually lands
The reference architecture describes a node by roles, not by slots. The HGX B300 Enterprise RA specifies “Eight NVIDIA ConnectX-8 SuperNICs per NVIDIA HGX B300 baseboard” at “Up to 800 Gbps per adapter”, giving “800 Gb/s (2 x 400Gb/s Ethernet) per GPU” - a one-to-one GPU-to-SuperNIC ratio - plus one “NVIDIA BlueField-3 B3240 DPU” for the north-south side.[6] The appendix states the same node as “8 on baseboard NVIDIA ConnectX-8 SuperNIC with a dual-port 400G configuration” and “1 NVIDIA BlueField-3 B3240 DPU with 2x400G ports”.[7] The phrase on baseboard is the whole point.
Dell realises that shape as embedded I/O. The XE9780 features table lists Embedded OSFP as “8 x OSFP with B300 GPU”, and the XE9785 lists Embedded OSFP as “8x CX8 OSFP (default)” for B300 and “N/A” for MI355X.[3][4] What remains for PCIe cards is small on purpose: “Up to 4 x Gen5 x16 Full-height, Half-length PCIe slots with NVIDIA B300 GPU configuration”, against “12 (8 x 75 W, 4 x up to 150 W) Gen5 x16 FHHL slots” when the same chassis is built with B200 GPUs.[3] The XE9785 prints the same fork as “4 x 150W Gen5 x16 FHHL cards” with B300 versus “12 x 75W Gen5 x16 FHHL cards” with MI355X.[4]
That north-south DPU carries one physical requirement worth memorising: the RA says the BlueField-3 B3240 is “Single-slot FHHL PCIe Gen5 x16” or “Dual-slot FHHL” and requires an “8-pin ATX 12V PCIe” connector because configurations “can draw more than 75 W”.[6] A slot that is electrically fine for a NIC is not automatically fine for this DPU.
The generation before it is packaged the other way round. The H100/H200/B200 Enterprise RA uses “BlueField-3 B3140H SuperNICs” for the compute fabric and “BlueField-3 B3220 DPUs” in compute nodes for the converged network at “two 200 GbE ports to two separate switches”.[8] On an XE9680 there is no embedded OSFP to help, so those nine adapters are nine cards in the ten available x16 slots - which is exactly what a real Dell BOM shows as “NVIDIA Bluefield-3 (1x3220/8x3140H)”.[9]
3Slot priority and thermal restriction: the two tables that decide a build
The XE9780 Technical Guide publishes a riser table and per-configuration slot lists, and they are the closest thing Dell gives you to an answer for “can I put this card there”.
In the PSRBB+B300 configuration the eight accelerators occupy slot priority “21, 22, 23, 24, 25, 26, 27, 28” with a maximum of eight, and every adapter class shares one short slot list: “NVIDIA DPU FH 200GB” at “6, 9, 3, 12” maximum four; “NVIDIA DPU FH 400GB (post RTS)” at “6, 9, 3, 12” maximum four; “NVIDIA OSFP NIC: 400G NDR” at “6, 9, 3, 12” maximum four; and the same list for the Broadcom, Intel and NVIDIA 100/200 GbE cards.[3] The riser table confirms why: in the RC0-B300 PSRBB row only OCP slot 1 and PCIe slots 3, 6, 9 and 12 are populated, while RC0-B200 PSBB populates all thirteen.[3]
The PSBB+B200 configuration is the roomy one. “NVIDIA Mellanox ConnectX-7, NDR 400G InfiniBand adapter card, 1x OSFP Direct Supply Method (single port)” and “NVIDIA Mellanox ConnectX-7 200 GB QSFP112 Gen 5.0 (dual port)” both list slot priority “4, 8, 7, 11, 2, 10, 5, 13, 6, 9, 3,12” with a maximum of twelve cards, while OCP 3.0 cards are restricted to slot “1” with a maximum of one.[3]
Then read the second table, because the first one can still be wrong for your build. The thermal restriction matrix lists DPU “B3220” and “B3240” as “STD fan” in the “B300 + 4 x PCIe card + 8 x OSFP” column and as “-” in the “B200 +12 x PCIe card” column.[3] Support is per configuration. A slot that exists physically is not a slot that is thermally qualified for that card in that build.
4Two documentation traps to carry into the field
Trap one: the XE9780 guide contradicts itself on the SXM generation. The system overview and the specification table both read “Eight NVIDIA HGX B300 NVL8 270 GB 1100W SXM6 GPUs”, while the XE9780-versus-XE9680 features comparison table in the same document reads “8 NVIDIA HGX B300 NVL8 270 GB 1100W SXM7 GPUs”.[3] The Dell product page, re-fetched on 2026-09-07, lists “8 NVIDIA HGX B300 NVL8 270GB 1100W SXM6 GPUs”.[2] Two sources out of three say SXM6, including the customer-facing product page, so quote the product page, name the discrepancy rather than hiding it, and let the account team raise the doc bug.
Trap two: the campaign page is not the catalogue. The Dell AI Factory with NVIDIA landing page names XE7740, XE9780 and R770 servers and PowerScale F710 storage, and it names PowerSwitch only generically - no XE9680, no XE9785, no SN5600, no SN2201.[5] Dell demonstrably sells all of those: the XE9680 and XE9785 have their own product pages, and a published Dell AI Factory solution is built from “2x Dell PowerEdge XE9680”, “3x NVIDIA Spectrum-4 SN5600 400G BE + 200G FE” and “1x NVIDIA SN2201 ToR”.[1][4][9] Absence from a landing page is a marketing decision, never a product statement.
Row 41, rewritten
The quote loses eight line items and gains one sentence. Eight east-west ports are embedded OSFP on the B300 sled, so the only PCIe adapter still to buy is the north-south DPU, and the slot-priority matrix puts the network cards at 6, 9, 3, 12 with a maximum of four in that configuration.[3] What you actually say: “You are not hitting a four-slot limit - eight of your nine adapters are already on the baseboard.”
Then procurement, who has said nothing all morning, looks up from her lead-time sheet. The optics for the new leaves were ordered last week, off the catalogue string, and the pallets land tomorrow.
Lab
Goal: match the paper matrices to a physical Dell-lab PowerEdge. Everything here is read-only - no firmware changes, no reboots, no BIOS edits.
- Pre-flight inventory. Record the service tag, model and iDRAC version before touching anything:
Expected: system model, service tag, iDRAC firmware. If not:racadm getsysinforacadmis not installed on the host - read Service Tag and iDRAC version from the iDRAC web UI under Dashboard, or runsudo dmidecode -s system-serial-number. Note that this is your case-opening evidence in lesson 5 as well. - Pull the PCIe slot inventory:
Expected: a slot-by-slot listing with vendor, device and slot number. If the grep is too noisy, runracadm hwinventory | grep -A12 -i "PCIeSSD\|Slot\|NIC"racadm hwinventoryalone and page through it, or use the iDRAC web UI under System, Inventory, Hardware Inventory. - Map each populated slot number to the guide’s priority list for that GPU configuration.[3] Expected: adapters sit in the earliest slots of the published priority list. Record any card that sits in a later slot and ask why - it is usually a previous rebuild, and it is worth knowing before you blame the card.
- Locate the BlueField-3 physically and confirm the auxiliary power cable is connected to the 8-pin ATX 12V PCIe header.[6] Expected: cable present. If it is absent, do not connect it now - record the finding, because that is a chassis-open change and belongs in a scheduled window.
- Confirm whether the OCP 3.0 slot is populated:
Expected: the adapter list plus driver and firmware for one port. If not: no Ethernet device is listed, orsudo lspci | grep -i ethernet sudo ethtool -i ens1f0np0ethtool -ireports “No such device” because the interface name differs - runip -br linkfirst and use the name it prints. Cross-check the count against the chassis’s OCP maximum of one card.[3] - Deliverable: a filled chassis card for the lab machine - model, GPU or accelerator, populated slots with card identity, OCP populated yes or no, DPU aux cable yes or no - and one sentence saying whether this host could carry the RA node shape as built.[6]
Goal: build the slot-and-thermal table you will use in front of a customer, and document the SXM contradiction yourself.
- Open the XE9780 Technical Guide and the XE9785 product page. Record for each chassis: form factor, GPU options with wattage and SXM string, the Embedded OSFP line, PCIe slot count per GPU option, OCP provision and PSU string.[3][4] Expected: a six-row table with no blank cells. If a cell is blank, the fact is not published and you write “not published” rather than guessing.
- From the guide’s
PSRBB+B300table, build the slot-priority table: card type, slot priority list, maximum number of cards. Include the accelerators themselves (slots 21 to 28), the NVIDIA DPU FH 200GB and 400GB rows, the OSFP NIC 400G NDR row and the OCP row.[3] Expected: every network card class shares slots 6, 9, 3, 12 with a maximum of four. If not: if a card class you expect is absent from the PSRBB+B300 table, check whether it appears only in the PSBB+B200 table - absence from one configuration is the finding, not an error in your reading. - Repeat for
PSBB+B200and put the two side by side.[3] Expected: twelve-card maximum and the long priority list “4, 8, 7, 11, 2, 10, 5, 13, 6, 9, 3,12” - and a note that the B300 build is the constrained one. If not: if your two tables show the same maximum, you have read the same table twice - the B200 rows are the ones with the thirteen-slot riser. - Add the thermal restriction matrix as a third column: for B3220 and B3240, which configuration says “STD fan” and which says “-”.[3] Expected: supported in the B300 plus 4 PCIe card plus 8 OSFP column; not listed in the B200 plus 12 PCIe card column. If not: a dash is not a “no” you can quote as a prohibition - record it as “not listed for that configuration” and take the question to the account team.
- Find the SXM contradiction. Search the guide for both
SXM6andSXM7and record the section each appears in, then check the product page.[3][2] Expected: SXM6 in the overview and specification table, SXM7 in the features comparison table, SXM6 on the product page. Write the one sentence you would say to a customer who quotes SXM7 at you. If not: if the guide you downloaded shows SXM6 everywhere, Dell has revised it - record the revision date you fetched and note that the contradiction is closed rather than assuming your search missed it. - Deliverable: a one-page chassis card for XE9680, XE9780 and XE9785 with the RA node shape mapped onto it - eight east-west adapters and one north-south DPU - showing for each chassis whether east-west is embedded or in slots.[6][7][8] If not: if a chassis has no published embedded OSFP line for the accelerator you picked, write “none published” - that is the answer, and it means every adapter is a PCIe card.
Retrieval check
10 questions from memory. Answer before looking anything up; misses become flashcards.
Explain it to a Dell SE
Explain to a Dell SE in five sentences why the number of PCIe slots on an AI server tells you almost nothing about how many network adapters the reference architecture needs.
Sources
Facts in this lesson were checked against Dell PowerEdge XE9680, XE9780 and XE9785 product pages and the XE9780 Technical Guide re-fetched 2026-09-07; NVIDIA HGX AI Factory Enterprise RA components and appendix, 2026-09-07. Dates are when each page was fetched.
- PowerEdge XE9680 | Dell USA · fetched 2026-09-07
- PowerEdge XE9780 tech specs | Dell · fetched 2026-09-07
- Dell PowerEdge XE9780 Technical Guide · fetched 2026-09-07
- PowerEdge XE9785 | Dell USA · fetched 2026-09-07
- Dell AI Factory with NVIDIA (landing page) · fetched 2026-09-07
- Components - NVIDIA HGX AI Factory Enterprise RA · fetched 2026-09-07
- Appendix B - Node Configurations - NVIDIA HGX AI Factory · fetched 2026-09-07
- Networking Physical Topologies - NVIDIA HGX AI Factory (H100 H200 B200) · fetched 2026-09-07
- Dell AI Factory with NVIDIA Solution ID 19845005.1 · fetched 2026-09-07
The same idea elsewhere
Other lessons that cover this ground, sometimes from another course's angle.