The generation map: which SuperNIC is on the baseboard
S1·E5BlueField-3 is a family, not a speed · Final design review, Friday morning, two hours before the BOM freezes
Builds on: Which reference architecture governs this deal, Reading node nomenclature: 2-8-9-400
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
- Place a GPU platform on the generation map by naming its east-west adapter, its north-south adapter and its per-GPU speed.
- Distinguish a SuperNIC role from a DPU role by traffic direction rather than by part number.
- Explain why the Network Operator support matrix and the Spectrum-X platform page print different ConnectX-9 speeds without treating either as wrong.
- Predict the three field failures that follow from misreading the generation - aux power, transceiver type and doubled optics counts.
Episode 5 — BlueField-3 is a family, not a speed
The architect wants one confirmation before the freeze: his existing forecasting pod has BlueField-3 cards, so the converged network runs at 400 gigabit, correct? You are one syllable from agreeing. Then the SE reads the line off the old build sheet - one B3220 and eight B3140H per node, two 200 GbE ports to two separate switches - and the NVIDIA PM on the call, asked about anything past that, says “not announced” and goes back to muting himself.
That is a different row of the map. On the H100, H200 and B200 Enterprise RA, GPUs are connected in a rail-optimized topology through their respective BlueField-3 B3140H SuperNICs, while the converged network uses BlueField-3 B3220 DPUs with each compute and management node connected by two 200 GbE ports to two separate switches.[1] The 400 GbE answer belongs to the next row: on HGX B300 the east-west side is eight ConnectX-8 SuperNICs at 800 Gb/s per GPU delivered as 2 x 400 Gb/s Ethernet, and the north-south side is a single BlueField-3 B3240.[2]
The map exists because the row decides three lines of a quote at once. Transceivers, since the dual-plane design needs twin-port transceivers and the single-plane design single-port OSFP.[10] Fabric port speed, which doubles between the generations. And power, because some B3240 configurations can draw more than 75 W and need a PCIe power connector on top of slot power.[2]
“BlueField-3” names a family that spans two rows; ask which SuperNIC is on the baseboard. Start with one row per platform.
1One row per platform, two adapters per row
The generation map is a small table you should be able to draw from memory in a customer meeting. Four platforms, and each one uses two different adapter kinds: one for east-west GPU-to-GPU traffic and one for north-south storage, in-band management and tenant services.
On the H100, H200 and B200 Enterprise RA, GPUs are connected using a rail-optimized network topology through their respective BlueField-3 B3140H SuperNICs, while the converged network uses BlueField-3 B3220 DPUs in each compute node, with each compute and management node connected by two 200 GbE ports to two separate switches.[1] Out-of-band on that generation runs at 1 GbE.[1]
On HGX B300, the east-west side becomes eight NVIDIA ConnectX-8 SuperNICs per HGX B300 baseboard at up to 800 Gbps per adapter, published as 800 Gb/s consisting of 2 x 400 Gb/s Ethernet per GPU - a 1:1 ratio of SuperNICs to GPUs.[2] The north-south side is a BlueField-3 B3240 P-Series FHHL DPU in a single-slot FHHL PCIe Gen5 x16 or dual-slot FHHL form factor.[2]
The rack-scale platforms follow the same shape per tray. A DGX GB200 NVL72 compute tray integrates four ConnectX-7 NICs to support InfiniBand NDR at 400 Gbps, plus two BlueField-3 NICs at 2x200 Gbps for the in-band management and storage networks.[4] A GB300 NVL72 tray carries two mezzanine network boards with two ConnectX-8 silicon chips on each, for a total of four ConnectX-8 host channel adapters at up to 800 Gb/s, plus one dual-port QSFP112 BlueField-3 B3240 DPU with an aggregate of approximately 480 Gb/s.[5][6]
| 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.
2The map, and the column that is not on any NVIDIA page
| Platform | East-west adapter | North-south adapter | Per-GPU or per-port east-west speed |
|---|---|---|---|
| H100 / H200 / B200 | BlueField-3 B3140H SuperNIC[1] | BlueField-3 B3220 DPU, 2x200 GbE[1] | Converged side published as 200 GbE per port[1] |
| HGX B300 | 8x ConnectX-8 SuperNIC, 1:1 with GPUs[2] | BlueField-3 B3240 DPU, 2x400 Gb/s[2] | 800 Gb/s per GPU as 2x400 Gb/s Ethernet[2] |
| GB200 NVL72 | 4x ConnectX-7 per tray, NDR 400 Gbps[4] | 2x BlueField-3 per tray, 2x200 Gbps[4] | 400 Gbps NDR per adapter[4] |
| GB300 NVL72 | 4x ConnectX-8 per tray, up to 800 Gb/s[5] | 1x BlueField-3 B3240, approx. 480 Gb/s[5] | 800 Gb/s per GPU[12] |
The fifth column is the one that earns its keep in the field, and no NVIDIA page prints it: what breaks if you get this row wrong.
For B300 the answer starts with power. The components page states that some BlueField-3 B3240 configurations can draw more than 75 W and therefore require both standard PCIe slot power and an additional PCIe power connector rated for at least 75 W, delivered over an 8-pin ATX 12V PCIe cable.[2] A PCIe slot supplies at most 75 W, so a card that fits mechanically can still be under-powered. That cable is a line item on a Dell build sheet, and it is the single most common surprise at rack-and-stack.
Next is optics. On HGX B300 the ConnectX-8 adapter supports Ethernet networking up to 800 Gb/s as 2x400 Gb/s via twin-port transceivers in the dual-plane design, or single-port OSFP transceivers in the single-plane design.[10] Plane count is therefore a transceiver purchasing decision and not just a drawing.
Last is the port count itself. Confuse a GB200 tray for a GB300 tray - both carry four east-west adapters - and every downstream number that depends on 400 versus 800 Gb/s is out by a factor of two.[4][5]
3A second document, a second set of numbers: the Network Operator matrix
The reference architectures tell you what is on a baseboard. The NVIDIA Network Operator platform-support matrix tells you something different: which adapters a given Network Operator release validates in Kubernetes. Re-fetched on 2026-09-09, the v26.4.0 matrix lists ConnectX-6 at 200 Gb/s with IB RDMA and RoCE; ConnectX-6 Dx at 200 Gb/s, Ethernet only, RoCE; ConnectX-7 at 400 Gb/s; ConnectX-8 SuperNIC at 800 Gb/s; ConnectX-9 SuperNIC at 800 Gb/s; BlueField-3 DPU at 200 Gb/s, Ethernet only; and BlueField-3 SuperNIC at 400 Gb/s, Ethernet only.[7]
Two details on that page change what you can promise. First, both BlueField-3 rows qualify RoCE as NIC mode only.[7] A design that wants the DPU running in DPU mode for infrastructure offload and wants RoCE on the same adapter is outside what that Network Operator version validates - not forbidden by physics, but unsupported by the operator.
Second, the ConnectX-9 row reads 800 Gb/s, while the Spectrum-X platform page describes ConnectX-9 as 1,600 Gb/s per GPU via 4x200 G SerDes with PCIe Gen6, for NVIDIA Vera Rubin NVL72 systems.[7][8] The same page places ConnectX-8 at 800 Gb/s total throughput via 2x400 G, PCIe Gen6, for NVIDIA Blackwell systems, and BlueField-3 for NVIDIA Hopper-era systems, optimized for RDMA/RoCE and multi-tenant isolation.[8]
Both numbers are current and neither is wrong. They answer different questions: what the operator validates today, versus what the silicon is specified to deliver on its target platform. The professional habit is to name the document every time you quote a speed.
The same discipline applies to model numbers. The public BlueField platform page describes BlueField-3 as a 400 Gb/s infrastructure compute platform and BlueField-4 as an 800 Gb/s infrastructure platform for gigascale AI factories, and says BlueField-4 STX combines the NVIDIA Vera CPU with ConnectX-9 networking - but it lists no B3140H, B3220 or B3240 model numbers, no port counts and no Arm core counts.[9] Those live in the RAs and the datasheets.
- Profile doca-all is "other profiles" in the matrix → Level 2. Every component is in cycle 25 (Oct 2025 → Jul 2026 (3.2 → 3.5)) → supported until the next October GA.
- This is exactly the Spectrum-X validated stack v2.3.1 combination (Sep 2026).
- ⚠ Any FW or mode change on PowerEdge needs a full power cycle, not a warm reboot (Dell KB 000300192; NVIDIA modes page).
Matrix (policy): doca-ofed ↔ FW/BF-FW-Bundle = L1 · doca-ofed ↔ BF-Bundle = L1 · other profiles ↔ FW or BF-Bundle = L2 · DOCA-DPU ↔ BF-FW-Bundle = L2 · DOCA Services ↔ BF-Bundle/FW = L2. source ↗
⚠ = not confirmed on a fetched primary source (hover for why). Facts as of DOCA 3.5.0 (Sep 2026). Selections are saved.
4Reading the map backwards, from a card in your hand
In the Dell lab you rarely start from an RA page. You start from a card, and you have to decide which row of the map it sits on and whether it is playing the SuperNIC role or the DPU role.
Two questions settle it. Is there a general-purpose CPU complex on the card? A BlueField-3 DPU has Arm cores and its own management identity - the DGX B300 XDR SuperPOD RA describes the out-of-band network as connecting the management ports of all devices including DGX B300 compute trays including system BMCs and BlueField-3 BMCs, which only makes sense for a card that boots its own operating system.[16] A ConnectX-8 SuperNIC has a Data Path Accelerator but not a host you log into. Which direction does it face? East-west GPU-to-GPU adapters sit 1:1 with GPUs on the baseboard; north-south adapters sit at one or two per node in a PCIe slot.[2]
That second question is why the nomenclature string from lesson 2 works as a checksum. The Enterprise RA whitepaper defines HGX 2-8-9-400 as two CPUs balanced with eight GPUs plus nine NICs with east-west traffic of 400 GbE per GPU.[13] The B300 appendix lists eight on-baseboard ConnectX-8 SuperNICs in a dual-port 400G configuration and one BlueField-3 B3240 DPU with 2x400G ports - eight plus one, and the appendix never prints the string itself.[3] If your inventory of a node adds up to nine adapters in an 8+1 split, the generation row is confirmed from the hardware side.
One caution about firmware as an identifier: firmware versions move independently of the map. The Spectrum-X validated solution stack pins matched rows such as ConnectX-8 firmware 40.50.1002 and BlueField-3 firmware 32.50.1002 against a switch OS version, and those rows are revised on their own cadence.[14] Use firmware to answer “is this cluster on a validated row”, never to answer “what generation is this card”.
Given. A Dell quote describes a PowerEdge node with an HGX baseboard: eight GPUs at 288 GB HBM3e each, eight adapters listed as ConnectX-8 dual-port 400G on the baseboard, and one PCIe card described as “BlueField-3 dual-port 400 GbE”.
Step 1 - place the east-west row. Eight on-baseboard ConnectX-8 SuperNICs in a dual-port 400G configuration at 1:1 with eight GPUs is the HGX B300 row of the map.[2][3] Not the H100/H200/B200 row, which would show B3140H SuperNICs instead.[1]
Step 2 - place the north-south row. A BlueField-3 with two 400 Gb/s ports is a B3240, and the components page names the B3240 P-Series FHHL DPU for HGX B300 deployments.[2] On the earlier generation that card would be a B3220 at 2x200 GbE.[1]
Step 3 - write the nomenclature string. Two CPUs, eight GPUs, 8+1 = nine NICs, 400 GbE per GPU per interface: 2-8-9-400, following the whitepaper’s definition.[13]
Step 4 - derive the per-GPU east-west speed. The components page publishes 800 Gb/s per GPU as 2 x 400 Gb/s Ethernet.[2] The string says 400 because it names the per-interface speed; the platform says 800 because each GPU has two of them. Both are right, and stating only one of them in a customer meeting is how sizing arguments start.
Step 5 - flag what breaks. Dual plane means twin-port transceivers, not single-port OSFP.[10] And check the build sheet for an 8-pin ATX 12V PCIe cable, because some B3240 configurations can draw more than 75 W.[2]
Answer. HGX B300 row; 2-8-9-400; 800 Gb/s per GPU east-west; twin-port optics if dual plane; confirm the DPU aux power cable.
Given. A rack-scale quote: 18 compute trays, each with four Blackwell Ultra GPUs and two Grace CPUs, “4x CX-8 HCA” and “1x BF3 dual-port QSFP112”.
- Trays per rack and GPUs per tray put this on the ________ row of the map.[5]
- Four ConnectX-8 host channel adapters per tray run at up to ________ Gb/s each.[5]
- The single BlueField-3 is model ________ with an aggregate of approximately ________ Gb/s.[5]
- The GB300 RA overview page calls this tray a ________ (dual plane) node architecture, which decomposes as 2 Grace CPUs, 4 GPUs, ________ NICs and ________ Gb/s per GPU.[15]
- Name the one adapter fact that distinguishes this tray from the previous rack generation, whose trays carry four ________ at NDR ________ Gbps.[4]
Problem. A Dell account sends you two node descriptions with the adapter models redacted, leaving only these facts:
- Node A: eight GPUs; nine network adapters total; converged network described as “two 200 GbE ports to two separate switches”; out-of-band 1 GbE.
- Node B: eight GPUs; nine network adapters total; converged network described as “two 400 GbE ports”; one adapter needs an auxiliary power cable.
Decide which generation row each node sits on, name the east-west and north-south adapter models you expect on each, and state the per-GPU east-west figure you would quote for Node B.
Acceptance criteria. (1) You cite the converged-network port speed as the deciding evidence, not the GPU count, since both nodes have eight GPUs and nine adapters. (2) You name B3140H plus B3220 for one node and ConnectX-8 plus B3240 for the other. (3) Your Node B figure is 800 Gb/s per GPU delivered as 2x400 Gb/s, and you say which page publishes it. (4) You flag the aux power cable on Node B as an item to confirm on the build sheet rather than assume.
Case closed at noon
You do not correct him on the card. You ask for the converged-network port speed, he reads back two 200 GbE ports per node, and the old pod lands on the H100/H200/B200 row with B3140H SuperNICs east-west and B3220 DPUs north-south - while the new nodes sit a row down on ConnectX-8 and a B3240.[1][2] What you actually say: “BlueField-3 is a family. Tell me the converged port speed - 200 GbE puts you on the Hopper-era row, 400 GbE puts you on B300 or GB300.” The BOM freezes on time: twin-port optics for dual plane, the aux power cable procurement added on Tuesday, two OOB labels per node.[10] Row 41 through row 47, all closed, and the SE’s coffee is cold again.
Lab
Goal: place the cards you actually have on the map you just built. Every step is read-only - no firmware writes, no mode changes, no configuration.
- Pre-flight inventory. On each BlueField-3 and ConnectX host in the Dell lab, record hostname, PowerEdge model and slot placement before you run anything. Note which cards sit in OCP 3.0 slots, because those are the ones people miscount.
- Run
mst status -vand record the device list with PCI addresses. Expected: one line per adapter with anmtdevice name. If the command is not found, DOCA-Host or the MLNX_OFED tools are not installed - stop and record that rather than installing anything on a shared lab host. - Run
mlxfwmanager --queryand record Part Number, PSID, Description and firmware version per adapter. Expected: a Description string that names the family. Place each card on your map row using the Description, not the firmware version - firmware moves independently of the generation.[14] - Decide DPU or SuperNIC for each card. A BlueField-3 DPU has Arm cores and a management identity of its own; check for
/dev/rshim0on the host and, if present, readcat /dev/rshim0/miscto confirm the Arm side is reachable. Expected on a DPU: a misc file withDEV_NAMEand boot fields. Expected on a ConnectX SuperNIC: no rshim device at all. If rshim is absent on a card you believe is a DPU, note it as a finding - do not load modules to force it. - Cross-check with
lspci -nn | grep -i -E 'mellanox|nvidia'and confirm the count matches step 2. A mismatch usually means an OCP 3.0 card enumerating differently, not a fault. - Check the aux-power question without touching the chassis. Run
mlxfwmanager --queryand record the Part Number and Description for each BlueField-3, then map the PN against the HGX AI Factory components page, which states that some BlueField-3 configurations for north-south connectivity can draw more than 75 W and therefore require both standard PCIe slot power and an additional PCIe power connector rated for at least 75 W.[2] Expected: a PN you can map to a B3240 or B3220 line, and a note of whether that configuration is one of the ones above 75 W. If not: record the PN as unmapped and ask the Dell build owner for the slot and power drawing rather than opening the chassis. Physical inspection of the card’s 8-pin connector belongs in a scheduled maintenance window on a host you own, outside this read-only lab. - Write the result as one sentence per host, in this shape: “This node carries N adapters of family X in role R, which places it on the Y row of the generation map, so its east-west per-GPU figure is Z.” Check the last clause against the RA page rather than memory.[1][2]
- Acceptance: for every card you can name the map row, the role, the document that publishes its speed, and whether its aux power connector is populated. Nothing on the hosts changed - confirm by re-running
mlxfwmanager --queryand seeing identical firmware versions.
Goal: build the generation map yourself, one cell at a time, each cell sourced to a specific page. No hardware needed. Budget 35 minutes.
- Open the H100/H200/B200 Enterprise RA physical-topologies page and record the east-west adapter model, the north-south adapter model, the converged port count and speed per node, and the out-of-band speed.[1] Expected: B3140H SuperNIC east-west, B3220 DPU north-south, two 200 GbE ports to two separate switches, 1 GbE out-of-band. If the page shows different model numbers, it has been revised - record what it now says with today’s date.
- Open the HGX AI Factory components page and record the SuperNIC count per baseboard, the per-adapter speed, the per-GPU speed and the BlueField-3 model with its form factors.[2] Expected: eight ConnectX-8 SuperNICs, up to 800 Gbps per adapter, 800 Gb/s as 2 x 400 Gb/s per GPU, B3240 P-Series FHHL. Also copy the sentence about the 8-pin ATX 12V PCIe connector verbatim - you will quote it.
- Open the DGX GB200 SuperPOD key-components page and record the tray NIC complement.[4] Expected: four ConnectX-7 for InfiniBand NDR at 400 Gbps, two BlueField-3 at 2x200 Gbps for in-band management and storage.
- Open the NVL72 AI Factory components page and record the GB300 tray NIC complement.[5] Expected: four ConnectX-8 host channel adapters at up to 800 Gb/s on two mezzanine boards, one dual-port QSFP112 B3240 at approximately 480 Gb/s. If you see two BlueField-3 per tray, you are reading the GB200 page by mistake - check the URL.
- Open the HGX AI Factory networking-hardware page and add a transceiver column: twin-port transceivers for dual plane, single-port OSFP transceivers for single plane.[10] This is the column that turns the map into a BOM check.
- Open the Network Operator 26.4.0 platform-support page and record, as a separate table, the max speed and protocol notes for ConnectX-6 through ConnectX-9 and both BlueField-3 rows.[7] Expected: ConnectX-9 at 800 Gb/s and both BlueField-3 rows carrying a NIC mode only qualifier on RoCE. Keep this table physically separate from the RA map so you never merge the two sets of numbers.
- Open the Spectrum-X platform page and record its ConnectX-8 and ConnectX-9 sentences verbatim.[8] Expected: 800 Gb/s total throughput via 2x400 G for Blackwell systems, and 1,600 Gb/s per GPU via 4x200 G SerDes for Vera Rubin NVL72. Write one sentence explaining why this does not contradict step 6.
- Add the fifth column, “what breaks if you get this wrong”, with one concrete consequence per row - aux power, transceiver type, or a 2x error in fabric port speed.[2][10]
- Acceptance: hand someone the map and a redacted node description that names only the converged port speed. They should reach the right row without asking you anything.
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 four sentences, how you tell which NVIDIA adapter generation a GPU node belongs to and what you would get wrong on the quote if you guessed.
Sources
Facts in this lesson were checked against HGX AI Factory components and networking hardware, HGX AI Factory (H100/H200/B200) physical topologies, NVL72 AI Factory components, DGX GB200 SuperPOD key components, NVIDIA Network Operator 26.4.0 platform support and the Spectrum-X Ethernet platform page, all re-fetched 2026-09-09. Dates are when each page was fetched.
- Networking Physical Topologies - NVIDIA HGX AI Factory (H100, H200, B200) · fetched 2026-09-09
- Components - NVIDIA HGX AI Factory (B300) · fetched 2026-09-09
- Appendix B - Node Configurations - NVIDIA HGX AI Factory · fetched 2026-09-07
- Key Components of the DGX SuperPOD - DGX GB200 Reference Architecture · fetched 2026-09-09
- System Hardware and Components - NVIDIA NVL72 AI Factory (GB300) · fetched 2026-09-09
- Networking Hardware - NVIDIA NVL72 AI Factory (GB300) · fetched 2026-09-07
- Platform Support - NVIDIA Network Operator 26.4.0 · fetched 2026-09-09
- NVIDIA Spectrum-X Ethernet Platform · fetched 2026-09-09
- NVIDIA BlueField Networking Platform · fetched 2026-09-07
- Networking Hardware - NVIDIA HGX AI Factory · fetched 2026-09-07
- Networking Physical Topologies - NVIDIA HGX AI Factory · fetched 2026-09-07
- Networking Physical Topologies - NVIDIA NVL72 AI Factory (GB300) · fetched 2026-09-07
- Key Building Blocks of Enterprise Reference Architectures · fetched 2026-09-07
- NVIDIA Spectrum-X Validated Solution Stack · fetched 2026-09-07
- Overview - NVIDIA NVL72 AI Factory (GB300) · fetched 2026-09-09
- Network Fabrics - DGX SuperPOD with DGX B300, Quantum-X800 InfiniBand and AC Power RA · fetched 2026-09-09
The same idea elsewhere
Other lessons that cover this ground, sometimes from another course's angle.
- Scenario: a 256-GPU Dell AI Factory design reviewElsewhere in this course · Same ground: planes, optics and converged
- The other three networks: converged, storage, out-of-bandElsewhere in this course · Same ground: gb200, converged and gb300
- PowerEdge support matrixDOCA course · Same ground: matrix, power and 12v