Skip to content
nodeza.learn
DOCA course/7. AI factories/Lesson 7.4

Roadmap: BlueField-4, ConnectX-9, DOCA Memos

S7·E4Twenty minutes in the hotel lobby · Hotel lobby across from the customer's campus, 07:40

S7·E4Evaluate~25 minsources checked todayverified against NVIDIA press releases (CES Jan 5 2026, GTC Mar 16 2026, BlueField-4 storage), NVIDIA BlueField-4 blog and product pages, Dell press releases (May 2025, Jan 2026 blog, Mar 2026), NVIDIA/skills doca-bf4-deployment, DOCA 3.5.0 release notes and policies — fetched 2026-09-06

Builds on: The validated stack and version pinning, Out-of-band: BMC, Redfish, and the BlueField-4 path

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

  • State the publicly announced Vera Rubin platform components (Vera, Rubin, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6) with their published headline numbers and availability windows.
  • Explain how BlueField-4 day-one bring-up differs from BlueField-3 (BMC-driven ISO and PLDM path versus rshim and BFB) and what that means for a Dell lab.
  • Judge which DOCA developer assumptions carry forward to ConnectX-9 and BlueField-4 and which are not yet published.
  • Apply the FAE messaging rules to a customer roadmap question and phrase not-announced correctly.
  • Evaluate a draft customer briefing and remove every claim that goes beyond public NVIDIA or Dell statements.

Episode 4 — Twenty minutes in the hotel lobby

The situation · Hotel lobby across from the customer's campus, 07:40

Coffee, a laptop on your knees, and twenty minutes before the briefing. The Dell SE sits down with the slide the carrier’s CTO sent overnight: a two-year build plan with BlueField-4 in the storage tier, ConnectX-9 on every GPU, a purchase-order date, and a core count copied out of a trade article. It is already on the Promises tab. On the phone, dialled in from a briefing room two time zones away, the NVIDIA PM says the two words she says best: not announced. The network lead will be in that room too, with maybe four pages of notebook left, and he will want a counter for every number on the slide.

The risk here is not technical. NVIDIA announced the Rubin platform at CES on January 5, 2026 and said at GTC on March 16, 2026 that seven chips — Vera, Rubin, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 and Groq 3 LPU — are in full production, available from partners starting the second half of that year.[1][2] Dell published its own windows separately, and they are Dell’s to publish.[9] The checkable half of the CTO’s question already has an answer anyone can open: ConnectX-9 has a firmware line in the DOCA 3.5.0 tables and BlueField-4 has none.[13][14]

Say what the customer can open in a browser; everything else is an NDA conversation, not a slide. So you open both press releases and the release notes, and start separating the published from the merely repeated.

1What NVIDIA announced, and when

At CES on January 5, 2026, NVIDIA announced the Rubin platform: the Vera CPU with “88 NVIDIA custom Olympus cores, full Armv9.2”, the Rubin GPU with “50 petaflops of NVFP4”, NVLink 6 at “3.6TB/s of bandwidth” per GPU and “260TB/s” per NVL72 rack, and BlueField-4 with “Advanced Secure Trusted Resource Architecture, or ASTRA”; availability was given as “second half of 2026”, with Dell named among the partners.[1]

At GTC on March 16, 2026, NVIDIA said “seven new chips now in full production”: Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch and Groq 3 LPU, with five rack types including “BlueField-4 STX storage racks” and “Spectrum-6 SPX Ethernet racks”, and products “available from partners starting the second half of this year”.[2] The same release introduced DOCA Memos as “a new DOCA framework” for “dedicated KV cache storage processing to boost inference throughput by up to 5x”.[2]

Those two press releases, plus the product pages and Dell’s own announcements, are the whole public record you may speak from. Everything in this lesson is either quoted from them or explicitly marked as not published.

2024202520262027LTS23 (2.5.x) → 2026LTS24 (2.9.x) → 2027 ▸LTS25 (3.2.x) → 2028 ▸3-year LTS windows (Level 1 for doca-ofed ↔ FW/BFB)DOCA 2.5.0 · 2023-12 (approx.)2.5.0Oct GA ~DOCA 2.6.0 · 2024-02 (approx.)2.6.0Jan ~DOCA 2.7.0 · 2024-04 (approx.)2.7.0Apr ~DOCA 2.8.0 · 2024-08 (approx.)2.8.0Jul ~DOCA 2.9.0 · 2024-11 (approx.)2.9.0Oct GA ~DOCA 2.10.0 · 2025-02 (approx.)2.10.0Jan ~DOCA 3.0.0 · 2025-05 (approx.)3.0.0Apr ~DOCA 3.1.0 · 2025-08 (approx.)3.1.0Jul ~DOCA 3.2.0 · 2025-11 (approx.)3.2.0Oct GA ~DOCA 3.3.0 · 2026-02 (approx.)3.3.0Jan ~DOCA 3.4.0 · 2026-05 (approx.)3.4.0Apr ~DOCA 3.5.0 · 2026-093.5.0JulDOCA 3.6.0 · 2026-103.6.0Oct GADOCA 4.0 · 2027-014.0Janquarterly GAOctober LTS baseannounced, not released~ = month approximate

⚠ = not confirmed on a fetched primary source (hover for why). Facts as of DOCA 3.5.0 (Sep 2026). Selections are saved.

Timeline tab: Hopper (BlueField-3 SuperNIC, Spectrum-4), Blackwell (ConnectX-8, BlueField-3 DPU), Vera Rubin (ConnectX-9, BlueField-4, Spectrum-6). Hover a part for its published window and source.

2The networking pieces: ConnectX-9, Spectrum-6, BlueField-4, DOCA Memos

ConnectX-9 is the Vera Rubin SuperNIC at “up to 1.6 Tb/s throughput per GPU” with 200G SerDes and PCIe Gen6, handling the platform’s scale-out RDMA networking.[7][6] Spectrum-6 is “102.4 Tb/s per switch chip with 200 G SerDes”; the SN6600 is “128 ports of 800 G OSFP” and the SN6800 “512 ports of 800 G in a 5U form factor”, with co-packaged optics carrying NVIDIA’s reliability and power-efficiency claims.[6][1]

BlueField-4 is described by NVIDIA as an 800 Gb/s infrastructure platform with “6x the compute power” of BlueField-3, combining “NVIDIA Grace CPU and NVIDIA ConnectX-9 networking”, in “early availability as part of NVIDIA Vera Rubin platforms in 2026”.[4][5] The product page lists a BlueField-4 STX variant that “Includes NVIDIA Vera CPU”.[5] Note the two CPU names across NVIDIA’s own pages; report both as published and do not reconcile them for the customer. The storage announcement positions BlueField-4 as the engine of the “Inference Context Memory Storage Platform”, “up to 5x” tokens per second and “up to 5x greater power efficiency than traditional storage”, integrating with DOCA, NIXL and Dynamo, with Dell Technologies, DDN, VAST Data, WEKA, Pure, HPE and Supermicro as partners.[3]

What NVIDIA has not published on the fetched pages: the BlueField-4 Arm core count, its PCIe generation, memory size, and ConnectX-9 lane configuration. The widely repeated “64-core” figure comes from third-party press (ServeTheHome, October 2025) and is not on any NVIDIA page fetched for this course.[18] Say so.

3BlueField-4 bring-up: the BMC-driven path

The BlueField-3 lifecycle you learned in module 2 runs over rshim: the host pushes a BFB with bfb-install, and the Arm side is reachable at the tmfifo address. The NVIDIA/skills repository describes a different day-one path for BlueField-4: bring-up is “driven from its BMC”, where “Grace” is the Arm CPU complex on the card, built on the DMTF Redfish and PLDM standards and UEFI.[11][12]

Three OS-install methods land the BlueField/DOCA bundle ISO on Grace: Method A, UEFI HTTP Boot (recommended), from an operator-run HTTP server with no PXE, DHCP or TFTP; Method B, PXE Boot, for sites that already have PXE or need a custom bf.cfg; Method C, Redfish Virtual Media, fully out-of-band, with the ISO uploaded to the BMC eMMC via Redfish SimpleUpdate (a 5 GB local-payload limit is documented), attached via Redfish VirtualMedia, followed by a BootSourceOverride to USB and a reset of Grace, and requiring a recent dpu-bmc version.[12] In every method the console is reached over BMC SSH with obmc-console-client, and the install media must be detached afterwards “to avoid boot loops”.[12]

Platform firmware (BMC, NIC firmware, SBIOS, ERoT) updates through PLDM for Firmware Update driven over Redfish: a multipart POST of the .fwpkg to the UpdateService, a Redfish Task to monitor, pldmtool fw_update GetFwParams -m <eid> to verify pending images, and a power cycle to activate; a BMC factory reset is recovery-only.[12] After the ISO install, the build string in /etc/mlnx-release identifies the image, the same file you read on a BlueField-3.[12]

Two cautions about evidence. First, the skill itself says “BF4 is not BF3”: methods, Grace naming and the dpu-bmc Redfish surface are BlueField-4-specific and the BlueField-3 rshim path is a different skill.[12][11] Second, this is a skills repository, not the DOCA document set: DOCA 3.5.0’s release notes and firmware table do not list BlueField-4 at all.[13][14] For a Dell lab that means a BlueField-4 test bench needs an out-of-band-reachable BMC, an HTTP server on the OOB network for the ISO, and pldmtool on a management host, none of which the BlueField-3 rshim workflow required.

4What changes for DOCA developers

Start from the supported-hardware tables. DOCA 3.5.0 supports BlueField-3, BlueField-2 and BlueField, and ConnectX-9, -8, -7, -6 Lx, -6 Dx, -6, -5/-5 Ex, -4 Lx and -4; its firmware table pins ten device lines: BlueField-3 32.50.1002, BlueField-2 24.50.1002, ConnectX-9 82.50.1002, ConnectX-8 40.50.1002, ConnectX-7 28.50.1002, ConnectX-6 Lx 26.50.1002, ConnectX-6 Dx 22.50.1002, ConnectX-6 20.43.8004, ConnectX-5 and -5 Ex 16.35.8008, and ConnectX-4 Lx 14.32.1912.[13][14] So ConnectX-9 is a DOCA 3.5.0 target today: a host-side DOCA application written against DOCA Core, Flow, RDMA or Ethernet on ConnectX-8 has a published path to ConnectX-9. The DPA toolchain already knows the family: the 3.5.0 samples’ default dpacc_mcpu list is nv-dpa-bf3,nv-dpa-cx7,nv-dpa-cx8,nv-dpa-cx9,nv-dpa-cx10.[19]

BlueField-4 is absent from all of those tables, so there is no supported DOCA target for it in 3.5.0 and no published statement about which DOCA release adds it.[13][14] What a developer can rely on is the policy, not the roadmap: DOCA_STABLE APIs are supported for the major version and the C ABI is versioned across minor and major releases, but a new major may introduce incompatible API changes.[16] Level 2 dependencies (any host profile other than doca-ofed, DOCA-DPU, DOCA Services) may break at each October GA.[17] NVIDIA has already used that mechanism: DOCA 3.6.0 (October 2026) is the last release for ConnectX-4 Lx and ConnectX-5, and DOCA 4.0 (January 2027) removes them.[15]

DOCA Memos is a name and a claim so far: a framework for KV-cache storage processing with an up-to-5x inference throughput figure, tied to the BlueField-4 storage platform and to NIXL and Dynamo.[2][3] No API, library quality level or DOCA release number is published on the fetched pages. A developer asking whether to design against it should hear exactly that.

5Dell's windows and the messaging rules

Dell publishes its own dates and its own support statements. On March 16, 2026 Dell announced the PowerEdge XE9812 (Vera Rubin NVL72) “globally available 2H 2026”, the XE9880L, XE9882L and XE9885L (HGX Rubin NVL8) for “Q3 2026”, PowerSwitch SN6000-series (Spectrum-6) “starting in July”, and PowerSwitch SN5610 and SN2201 “supporting Cumulus Linux and SONiC”.[9] At CES Dell said its AI Data Platform would integrate the BlueField-4 “data and storage processor” providing “800 Gb/s networking performance”, that PowerSwitch would expand with Spectrum-6 at “102.4 Tb/s” and “up to 512 ports of 800G”, and that HGX Rubin NVL8 PowerEdge systems would carry “800 Gb/s NVIDIA ConnectX-9 SuperNICs”.[8] Earlier windows follow the same pattern: SN5600 and SN2201 and the XE9780 and XE9785 were Dell’s “2H 2025”.[10] Dell also owns its support boundaries; for example Dell states that VMware DSE is unsupported on NVIDIA-channel DPUs in PowerEdge.[20]

The rules an FAE on the Dell account follows:

  1. Say only what is in public NVIDIA or Dell statements: chip names, headline numbers, and the published window (“second half of 2026”; Dell’s “Q3 2026” or “2H 2026”).[1][2][9]
  2. Quote the notice when asked for guarantees: every NVIDIA document states it “is not a commitment to develop, release, or deliver any Material … code, or functionality” and that specifications may change “at any time without notice”.[16]
  3. Never quote unpublished specifications: BlueField-4 core count, PCIe generation or memory, ConnectX-9 lane counts. If a number exists only in third-party press, name the press and say you cannot confirm it.[18][4]
  4. Never promise compatibility across a DOCA major or across an October GA for Level 2 pairs.[17][15]
  5. Never state Dell availability for an NVIDIA part or NVIDIA availability for a Dell system; attribute each window to the company that published it.[9][2]
  6. Safe phrasing: “publicly announced for”, “NVIDIA has stated”, “Dell’s press release says”; route firm dates, pricing and anything beyond the public record to the account manager and Dell product management.

How to say not announced, in one breath: “NVIDIA has not published that. What it has said is X, on date Y. If the decision depends on the unpublished part, I will ask the account manager to arrange the right conversation under NDA.”

Audit a draft roadmap briefing

Draft sentence from a colleague: “BlueField-4 ships in Q3 2026 with 64 Arm cores and PCIe Gen6, DOCA 3.6 will support it on day one, and it will be a drop-in upgrade for your BlueField-3 DPUs in the XE9780.”

  1. “ships in Q3 2026”: NVIDIA said “second half of 2026” and “early availability as part of NVIDIA Vera Rubin platforms in 2026”; Q3 is Dell’s window for the HGX Rubin NVL8 PoweEdge systems, not for the BlueField-4 part. Rewrite with the two attributed windows.[1][4][9]
  2. “64 Arm cores and PCIe Gen6”: neither figure is on an NVIDIA page fetched; 64 cores is ServeTheHome. Remove or mark as unconfirmed press.[18]
  3. “DOCA 3.6 will support it on day one”: no DOCA document states which release adds BlueField-4; DOCA 3.5.0 does not list it. Replace with the factual gap.[13][15]
  4. “drop-in upgrade for your BlueField-3 DPUs in the XE9780”: the BlueField-4 bring-up path is BMC-driven (ISO, PLDM), not rshim and BFB, and Dell has published no XE9780 BlueField-4 option. Replace with what is published: Dell’s BlueField-4 statements concern the AI Data Platform and Rubin-era systems.[12][8]
  5. Rewritten: “NVIDIA has announced BlueField-4 as an 800 Gb/s DPU with early availability alongside Vera Rubin platforms in 2026; Dell has announced BlueField-4 in its AI Data Platform and Rubin-based PowerEdge systems for Q3 and 2H 2026. NVIDIA has not published the core count or PCIe generation, and DOCA 3.5.0 does not yet list BlueField-4; provisioning is documented as BMC-driven rather than the rshim path used on BlueField-3.”

Episode 4 — case closed: the pod, and the slide

How it ended

The briefing slide ends with three columns: the claim, who published it, and the date. Vera Rubin components in full production as of GTC, March 16, 2026; Dell’s PowerEdge and PowerSwitch windows attributed to Dell’s own release; BlueField-4 at 800 Gb/s with early availability alongside Vera Rubin.[2][9][4] The core count from the trade article comes off, named as third-party press you cannot confirm.[18] To the CTO you say: “ConnectX-9 is a DOCA target today and I can show you the table. Which DOCA release adds BlueField-4 has not been published — for that I will get you the NDA conversation.”[13] The pod trained through storm season on thirty-two corrected quotes. The notebook is full. The operator’s last label, stuck to the spine rack, reads NEITHER OF THESE IS BUSY.

Lab

Read-only. Baseline the Dell-lab BlueField-3 so a future BlueField-4 bench can be compared against it, and record which BlueField-4 provisioning prerequisites the lab already has.

  1. sudo mst start && sudo mlxfwmanager --query. Record Part Number, PSID and FW version (pin for BlueField-3 in DOCA 3.5.0: 32.50.1002).[14]
  2. Record the BlueField-3 provisioning path evidence: ls -l /dev/rshim* on the host and, if reachable, ssh ubuntu@192.168.100.2 cat /etc/mlnx-release. These are the rshim and BFB artifacts that BlueField-4 replaces with BMC-driven ISO install; /etc/mlnx-release is the one file both paths share.[12]
  3. If the card’s 1GbE OOB port is cabled and you know the BMC address, run read-only Redfish GETs from a management host: curl -sk https://<bmc-ip>/redfish/v1/ (expect JSON with a RedfishVersion field) and curl -sk https://<bmc-ip>/redfish/v1/UpdateService (expect an UpdateService resource). Do not POST. These are the endpoints the BlueField-4 PLDM flow uses; record whether the lab BMC exposes them.[12] If not reachable: record “BMC OOB not cabled” and stop.
  4. Write the two-column comparison, BlueField-3 today versus BlueField-4 as documented: install transport (rshim BFB versus UEFI HTTP Boot, PXE or Redfish Virtual Media), firmware update path (BFB or BF-FW-Bundle versus PLDM over Redfish UpdateService), console (rshim console versus BMC SSH with obmc-console-client), and DOCA support status (3.5.0 listed versus not listed).[11][12][13]
  5. List what the lab would still need for a BlueField-4 bench: an HTTP server on the OOB network for the bundle ISO, pldmtool on a management host, and a BlueField-4 with its BMC reachable out-of-band. State that no configuration was changed.[12]

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, what you can and cannot tell a customer about BlueField-4 today.

12 flashcards for this lesson — 0 in deck. Spaced review lives at /review.

Sources

Facts in this lesson were checked against NVIDIA press releases (CES Jan 5 2026, GTC Mar 16 2026, BlueField-4 storage), NVIDIA BlueField-4 blog and product pages, Dell press releases (May 2025, Jan 2026 blog, Mar 2026), NVIDIA/skills doca-bf4-deployment, DOCA 3.5.0 release notes and policies — fetched 2026-09-06. Dates are when each page was fetched.

  1. NVIDIA Kicks Off the Next Generation of AI With Rubin (CES, Jan 5 2026) · fetched 2026-09-06
  2. NVIDIA Vera Rubin Opens Agentic AI Frontier (GTC, Mar 16 2026) · fetched 2026-09-06
  3. NVIDIA BlueField-4 Powers New Class of AI-Native Storage Infrastructure (Jan 5 2026) · fetched 2026-09-06
  4. NVIDIA Launches BlueField-4: The Processor Powering the Operating System of AI Factories · fetched 2026-09-06
  5. BlueField Networking Platform (BlueField-3 / BlueField-4 DPU) · fetched 2026-09-06
  6. NVIDIA Spectrum-X Ethernet Platform for Giga-Scale AI · fetched 2026-09-06
  7. NVIDIA SuperNICs (BlueField-3 / ConnectX-8 / ConnectX-9) · fetched 2026-09-06
  8. Dell Uplevels AI Infrastructure With NVIDIA at CES (Jan 2026) · fetched 2026-09-06
  9. Dell AI Factory with NVIDIA Delivers Proven Path to Enterprise AI ROI (Mar 16 2026) · fetched 2026-09-06
  10. Dell Technologies Unveils Next Generation Enterprise AI Solutions with NVIDIA (May 19 2025) · fetched 2026-09-06
  11. NVIDIA/skills — doca-bf4-deployment SKILL.md · fetched 2026-09-06
  12. NVIDIA/skills — doca-bf4-deployment CAPABILITIES.md · fetched 2026-09-06
  13. DOCA Release Notes v3.5.0 · fetched 2026-09-06 · DOCA 3.5.0
  14. DOCA General Support (firmware table) · fetched 2026-09-06 · DOCA 3.5.0
  15. DOCA 3.5.0 Changes and New Features · fetched 2026-09-06 · DOCA 3.5.0
  16. DOCA Backward Compatibility Policy (PDF) · fetched 2026-09-06 · DOCA 3.5.0
  17. DOCA Dependency Compatibility Policy (DOCA 3.5.0) · fetched 2026-09-06 · DOCA 3.5.0
  18. ServeTheHome: NVIDIA BlueField-4 with 64 Arm Cores and 800G Networking Announced for 2026 (third-party press) · fetched 2026-09-06
  19. doca-samples 3.5.0 — applications/meson_options.txt (dpacc_mcpu targets) · fetched 2026-09-06 · DOCA 3.5.0
  20. Dell KB 000225111 — NVIDIA Channel DPU VMware DSE Support · fetched 2026-09-06

The same idea elsewhere

Other lessons that cover this ground, sometimes from another course's angle.