Dell's switch catalog: SN, Q and Z in one price list
S3·E2The pallet that was four times too big · a partner's staging warehouse, the morning the optics land
Builds on: The PowerEdge AI server line
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
- Name every switch family Dell sells for AI networking and the role each one plays.
- Convert a Dell catalog port string into cages and logical ports before quoting optics.
- Explain the Spectrum-4 feature set a Dell rep can claim from the datasheet.
- Identify where Dell's and NVIDIA's published strings for the same switch disagree and say which one you would quote.
Episode 2 - The pallet that was four times too big
The pallets are open and the count is wrong by a factor of four. The order says 256 transceivers per leaf, there are four leaves, and they are SN5610s. Procurement is certain the number came off a datasheet, and she is half right: NVIDIA’s comparison table counts that switch as 256 ports of 200GbE at 51.2 Tb/s in 2U.[2] Dell’s own catalogue line for the same box counts the sheet metal instead - “64x 800 GbE OSFP, 2x 25 GbE SFP28”.[1] Somebody read the bigger number. The freeze is thirty-three days out and three quarters of that pallet has nowhere to go.
This is a vocabulary problem with an engineering cause. An OSFP cage is one opening in the sheet metal with one set of electrical lanes behind it, and the transceiver decides how those lanes are presented. The reference architecture says it in design terms: a dual-plane build uses “twin-port transceivers” and a single-plane build uses “single-port OSFP transceivers”, out of the same cage.[7] A vendor writing for a fabric architect counts what the switch can present; a vendor writing for a purchasing system counts what the sheet metal has. Both numbers are true. Only one of them is the number you buy against.
The night-shift operator wanders over with his label maker, looks at the pallet and asks the only useful question: how many holes are there. Cages are what you buy; ports are what you present. Dell’s page prints four switch families in that same shorthand. Learn to read the cage out of every one.
1One price list, four families
Dell’s AI networking page is a single catalogue with four distinct families on it, and a Dell rep can quote every one of them. Knowing the whole page keeps you honest when a customer says “we already looked at Broadcom”.
Spectrum-6. PowerSwitch SN6800-LD at “512 x MMC-12” and “409.6 Tbps”, SN6810-LD at “128 x MMC-12” and “102.4 Tbps”, SN6600-LD at “64 x OSFP224” and “102.4 Tbps”.[1]
Spectrum-4. PowerSwitch SN5610 at “64x 800 GbE OSFP, 2x 25 GbE SFP28” and “51.2 Tbps”, SN5600D at “64x 800 GbE OSFP, 1x 25 GbE SFP28” and “51.2 Tbps”, SN5400 at “64x 400 GbE QSFP-DD, 2x 25GbE SFP28” and “25.6 Tbps”.[1]
Broadcom Tomahawk, same page. PowerSwitch Z9864F-ON at “64x 800 GbE OSFP112, 2x SFP+” and “51.2 Tbps”, Z9664F-ON at “64x 400 GbE QSFP56-DD, 2x 10 GbE SFP+” and “25.6 Tbps”.[1]
InfiniBand. Quantum-X800 “Q3401-RD and Q3400-RA” at “144 x 800Gb/s over 72 OSFP cages” and “115.2 Tbps”; “Q3200-RA” at “72x 800 Gb/s over 36 OSFP cages” and “57.6 Tbps”; Quantum-2 “QM9790 / QM9701 / QM9700” at “64x 400 Gb/s over 32 OSFP cages” and “25.6 Tbps”.[1]
Out-of-band. PowerSwitch SN2201 at “48x RJ45 1 G BASE-T, 4x QSFP28” and “448 Gbps”, or Dell’s own S3248T-ON at “48x RJ45 10/100/1000 Mb, 4x 10 G SFP+, 2x 100 G QSFP28” and “576 Gbps”.[1]
| Product | Speed | PCIe | Role | GPU generation |
|---|---|---|---|---|
NIC | ||||
SuperNIC (no Arm) | ||||
SuperNIC (no Arm) | ||||
DPU | ||||
DPU | ||||
SuperNIC (Arm inactive) | ||||
DPU / storage processor | ||||
Ethernet switch | ||||
Ethernet switch | ||||
InfiniBand switch | ||||
InfiniBand switch |
⚠ = not confirmed on a fetched primary source (hover for why). Facts as of DOCA 3.5.0 (Sep 2026). Selections are saved.
2Cages, not terabits
The number that decides a bill of materials is the cage count. Everything else is arithmetic on top of it.
An SN5600, SN5600D or SN5610 presents 64 OSFP cages. Dell prints the two it carries in its AI catalogue as “64x 800 GbE OSFP, 2x 25 GbE SFP28” for the SN5610 and “64x 800 GbE OSFP, 1x 25 GbE SFP28” for the SN5600D; NVIDIA’s comparison table counts the same 64 cages as “256” ports of 200GbE at 51.2 Tb/s in 2U, and shows the SFP28 count as the only connector-string difference across the three models - “+ 2 SFP28 25GbE” on SN5610 against “+ 1 SFP28 25GbE” on SN5600 and SN5600D.[1][2] A QM9700 presents 32 OSFP cages carrying 64 logical ports of NDR 400 Gb/s, with 25.6 Tbps of switching throughput.[4] Dell’s catalogue says the same thing in one string: “64x 400 Gb/s over 32 OSFP cages”.[1]
The RA connects cage type to design choice directly: the dual-plane option uses “twin-port transceivers” while the single-plane option uses “single-port OSFP transceivers”.[7] So the plane decision made in module 2 is what fixes the optics part number, and the cage count is what fixes how many of them fit.
There is one published string conflict to handle carefully. NVIDIA’s comparison table lists SN6600 as “64x OSFP 2x 800 Gb/s” with “512” 200Gb/s ports at 102.4 Tb/s in 3U, while Dell lists “SN6600-LD” as “64 x OSFP224” at “102.4 Tbps”.[2][1] Both agree on 64 cages and on 102.4 Tbps; they differ in whether they count the cage or the logical twin ports it carries. What neither page decodes is Dell’s -LD suffix, so state the agreement and stop there.
3What the Spectrum-4 datasheet lets you claim
When a Dell rep asks what they can say about the SN5600 series, the datasheet is the answer and it is unusually specific.
Scale: “512,000 forwarding entries flexibly shared across access control list (ACL), longest prefix match (LPM) routes, host routes, media access control list (MAC), equal-cost multi-path (ECMP), and tunnel applications”, “512,000 on-chip flow counters” and “256-way equal-cost multi-path (ECMP)”.[3] Buffer: a “160 MB fully shared packet buffer”.[3] Pipeline: a “deep packet parser and editor that can process payloads up to the first 512 bytes” and “single-pass VXLAN routing and bridging”.[3] Visibility: What Just Happened, “Hardware-accelerated histograms powered by NetQ” that “track and summarize queue depths at submicrosecond granularity”, INT and streaming telemetry.[3] Security: “Hardware, firmware, and software are authenticated by a built-in root of trust, from the basic input and output system (BIOS) to the network operating system (NOS)”.[3]
The physical differences between the three SN5600-series models are the ones that get missed in a design review. SN5610 has four power supplies in 2+2 with five fans at N+1; SN5600 and SN5600D have two supplies in 1+1 with four fans at N+1.[3] SN5610 is rated “0 - 40 degrees C” while SN5600 and SN5600D are rated “0 - 35 degrees C”.[3] SN5610 and SN5600 take “200 - 240 VAC” while the SN5600D is a “DC bus bar” switch with an “Input range: 40-60 VDC”.[3] All three are 2U and 788 mm deep with airflow listed as “Reverse”, and all three carry 8x 100G PAM4 lanes per port with 32 GB of system memory.[3]
On the InfiniBand side the equivalent facts are power and airflow. QM9700 draws a typical 747 W with passive cables and a maximum 1720 W with active cables; QM9790 is 640 W and 1610 W; QM9701 is 720 W and 1660 W.[4] Airflow is directional and temperature-bounded: “Forward air flow: 0 to 35 degrees C” and “Reverse air flow: 0 to 40 degrees C”.[4] The Quantum-X800 generation above it advertises “144 ports of 800 Gb/s connectivity per port” with “SHARP v4, adaptive routing, and telemetry-based congestion control”, and the switching page names Q3200 as the 2U air-cooled member of the family.[6][5]
4How the catalogue shows up in a real quote
Dell’s own published AI Factory solution uses exactly three of these lines: “3x NVIDIA Spectrum-4 SN5600 400G BE + 200G FE”, “2x PowerSwitch S5232F-ON Storage Cluster 100G BE” and “1x NVIDIA SN2201 ToR”.[8] Three observations follow from that single fragment.
First, a Spectrum-4 switch is doing double duty - back end at 400G and front end at 200G - which the 64-cage breakout radix makes possible.[8][2] Second, the storage cluster network is carried by a Dell PowerSwitch, not by the NVIDIA line, at 100G.[8] Third, out-of-band is one SN2201, which matches the RA pattern of a 1 GbE out-of-band fabric for BMC and management ports.[8][7]
Three quarters of a pallet, going back
The surplus transceivers go back and nobody blames the datasheet, because the datasheet was right in both directions: 64 cages, counted by NVIDIA as 256 ports of 200GbE.[1][2] The number missing from that order was never on a switch page at all - it was the plane decision, which is what picks twin-port against single-port optics for those cages.[7] What you actually say: “Give me cages, cage type and how many planes, and I will give you the optics count. The terabits per second never enter that arithmetic.”
Procurement is not finished. She puts the pilot’s one-page solution sheet on the pallet and asks the question the SE has been avoiding: is that thing the NVIDIA architecture or not.
Lab
There is no NVIDIA switch in the Dell lab, so work the half you do have: the adapter end of the link decides which leaf SKU can terminate it. Read-only throughout.
- Pre-flight inventory. On each host record model, adapter count and PCIe placement:
Expected: one line per adapter with its PCI address and device name. If not:sudo lspci | grep -i mellanox sudo mst start && sudo mst status -vmst status -vprints nothing when the mst driver is not loaded; runsudo mst startagain and checklsmod | grep mst. If the host has no NVIDIA adapter, record the row as “no adapter” rather than leaving it blank. - Read the adapter identity and firmware:
Expected: part number, PSID, description and current firmware per device. If not:sudo mlxfwmanager --querymlxfwmanageris part of the firmware tools package - if the command is missing, installmftor read the same part number fromsudo lspci -vvand the label on the card bracket. The description string is what tells you the cage type the card expects. - Read the installed transceiver or cable on a live port:
Expected: identifier (QSFP112, OSFP, SFP28), vendor name and part number, and the connector type. If the port has a DAC attached you will see a copper cable identifier instead of optical diagnostics.sudo ethtool -m ens1f0np0 | head -20 - Map each adapter to the leaf SKUs in your table that can terminate it: an OSFP-caged adapter can land on an SN5600-series or QM97xx cage; a QSFP112 adapter cannot terminate directly in an OSFP cage without the correct transceiver and breakout.[1][4] Expected: a two-column list of adapter to compatible leaf SKUs, with any unknown marked unknown.
- Optional, only in a customer or partner lab with a live Spectrum switch and permission: run the read-only identity commands and compare them against the datasheet row you built.
Expected: platform model and NOS version. Change nothing on a customer switch.nv show platform nv show system - Deliverable: the adapter-to-leaf compatibility list plus one line naming which fact in your catalogue table could only be confirmed from hardware and not from either vendor’s page.
Goal: build the two-vendor catalogue table you can quote from memory, and rehearse the discovery workflow on the SONiC nodes you already have.
- Build one table with these columns: Dell SKU string, NVIDIA SKU string, cages and cage type, logical ports as each vendor counts them, throughput, rack units, role.[1][2] Cover SN6800-LD, SN6810-LD, SN6600-LD, SN5610, SN5600D, SN5400, Z9864F-ON, Z9664F-ON, Q3401-RD, Q3400-RA, Q3200-RA, QM9790, QM9701, QM9700, SN2201 and S3248T-ON. Expected: sixteen rows, every cell sourced. If not: if a cell has no published string on either vendor page, write “not published” rather than carrying a number over from a neighbouring model - the suffix rows are where that temptation bites.
- Mark every row where the two vendors’ strings disagree, and write one sentence per marked row explaining the disagreement.[1][2] Expected: the SN6600 versus SN6600-LD row is marked as a counting difference over the same 64 cages; the -LD and -RD/-RA suffix rows are marked as not decoded on any fetched page. If not: if you find yourself writing a decode you cannot quote, delete it and mark the row unknown - an invented decode is the failure this step exists to prevent.
- Add a power and thermal column from the datasheets: PSU count and redundancy, operating temperature, power input, airflow direction for the SN5600 series; typical and maximum wattage plus airflow ranges for QM9700, QM9790 and QM9701.[3][4] Expected: SN5610 at 0-40 C with 2+2 PSUs, SN5600 and SN5600D at 0-35 C with 1+1, SN5600D on 40-60 VDC bus bar. If not: the Dell-hosted datasheet covers the SN5600 series only - for the Quantum-2 rows switch to the QM97XX hardware manual, and leave the Spectrum-6 and Z-series rows blank rather than borrowing figures.
- In
~/containerlab, start the Dell Enterprise SONiC 4.5.1 node and run the discovery commands you would run on a real switch:
Expected: platform, HW SKU and ASIC strings, then a port table. Write down which of your catalogue columns you could have filled from this output alone and which you could not - cage type is usually the missing one. If not: if the container is not up,show platform summary show interface statussudo containerlab inspectwill say so; ifshow platform summaryis rejected, you are at the Linux shell rather than the SONiC CLI - runsonic-clifirst. - Deliverable: a one-page catalogue card, plus the two sentences you would say when a customer claims Dell and NVIDIA disagree about SN6600 port counts.[1][2] If not: if your two sentences assert that one vendor is wrong, rewrite them - the published agreement is 64 cages and 102.4 Tbps, and that is the whole defensible claim.
Retrieval check
10 questions from memory. Answer before looking anything up; misses become flashcards.
Explain it to a Dell SE
Explain to a Dell rep in four sentences why the first number you read on a switch datasheet should be the cage count and not the terabits per second.
Sources
Facts in this lesson were checked against Dell AI Networking Switches catalogue and NVIDIA Spectrum Ethernet switch comparison table re-fetched 2026-09-07; NVIDIA InfiniBand switching page 2026-09-07; Dell-hosted SN5600 datasheet and QM97XX specifications, 2026-09-07. Dates are when each page was fetched.
- AI Networking Switches | Dell USA · fetched 2026-09-07
- NVIDIA Spectrum Ethernet Switches · fetched 2026-09-07
- NVIDIA Spectrum SN5600 Series Switches Datasheet (Dell-hosted) · fetched 2026-09-07
- Specifications | QM97XX 1U NDR 400Gbps InfiniBand Switch Systems User Manual · fetched 2026-09-07
- NVIDIA InfiniBand Switches · fetched 2026-09-07
- NVIDIA Quantum-X800 InfiniBand Platform · fetched 2026-09-07
- Networking Hardware - NVIDIA HGX AI Factory Enterprise RA · 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.
- Choosing the switch: SN2201 through SN6800-LDSpectrum-X course · Same ground: spectrum6, power and E-switch
- The NVIDIA networking portfolioDOCA course · Same ground: quantum, infiniband and family
- Scenario: a 256-GPU Dell AI Factory design reviewElsewhere in this course · Same ground: oob, optics and power