Skip to content
nodeza.learn
DOCA course/9. FAE capstone/Lesson 9.3

Scenario: sizing and positioning

S9·E3Three cards on the table and a printed press release · Dell briefing room, Round Rock, day twenty-two of the five-week window

S9·E3Evaluate~45 minsources checked todaylab mutates hardwareverified against DOCA 3.5.0 docs, Spectrum-X whitepaper and validated stack, HGX B300 RA, Dell KBs 000225111/000300192, Broadcom KB 379391, NVIDIA and Dell press releases Jan-Mar 2026, 2026-09-06

Builds on: Modes of operation: DPU, NIC, Zero-Trust, Offload economics, Roadmap: BlueField-4, ConnectX-9, DOCA Memos

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

  • Choose between DPU mode, NIC mode and a ConnectX-8 SuperNIC for a VMware farm, an AI training pod and a storage-heavy inference cluster, and defend the choice from published documents.
  • Answer the five standard objections (cost, Arm complexity, existing OVS, lock-in, wait for BlueField-4) with NVIDIA's or Dell's published position and nothing more.
  • Apply the honesty rules for unreleased products: quote announced numbers and windows, the non-commitment notice, and route dates and prices away from yourself.
  • Write a one-page positioning memo that a Dell SE can forward without editing.

Episode 3 — Three cards on the table and a printed press release

The situation · Dell briefing room, Round Rock, day twenty-two of the five-week window

Three cards lie on the table, literally: a BlueField-3 DPU, a BlueField-3 SuperNIC and a ConnectX-8. Across the table: the network lead, notebook now a third full, the storage lead, and procurement, who asks about lead time. The Dell SE has fresh coffee and will let it go cold; on the video tile the NVIDIA PM opens with his line: nothing unannounced gets discussed.

The network lead opens with the expected objection: a DPU is a NIC with extra power draw. The storage lead slides across a printed BlueField-4 announcement and asks whether to wait.

Both are configuration questions, not product ones. In DPU mode the Arm cores are active and run the services that manage NIC resources and the data path; in NIC mode the Arm is inactive, the device functions as an NVIDIA ConnectX adapter, and that reduces power consumption, improves network performance and minimises the host memory footprint.[1] The third card has no Arm at all: ConnectX-8, 800 Gb/s over PCIe Gen6 x16.[8]

The mode switch exists because two customers wanted opposite things from one piece of silicon. One wanted a fast, quiet adapter; the other wanted a small server inside the server, running the vSwitch or a BGP/EVPN router on cores the host cannot touch, which is why HBN needs DPU mode and rules out the 8 GB platforms.[1][16] So NVIDIA shipped one that can be told which to be.

The mode is a question about the workload, not about the card. So you answer neither objection. You ask what has to run on the card.

1Three answers and the one question that decides

Positioning a BlueField on a Dell platform is a choice among three configurations. In DPU mode the Arm cores are active and “the embedded Arm system runs services that manage the NIC resources and data path”; in NIC mode the Arm cores are inactive and “the device functions as an NVIDIA ConnectX network adapter”, which “reduces power consumption, improves network performance, and minimizes the host memory footprint”.[1] The third answer is a ConnectX-8 SuperNIC, 800 Gb/s over PCIe Gen6 x16, with no Arm at all.[8][7]

The deciding question is whether anything must run on the card. If not, the customer wants bandwidth and isolation and the choice is NIC mode or ConnectX-8. If yes, because OVS-DOCA, HBN, SNAP storage emulation or vSphere DSE must run on the Arm, the card is in DPU mode and the SKU must carry enough memory for the service; HBN, for instance, excludes 8 GB platforms.[16] DPU SKUs (B3240, B3220, B3210, B3210E) have 16 cores and 32 GB and need the 8-pin auxiliary connector; SuperNICs have 8 cores, 16 GB and no connector.[30]

Every answer then gets pinned to a validated-stack row so the versions are not up for debate: v2.3.1 (September 2026) lists Cumulus 5.18.1, BlueField-3 firmware 32.50.1002, DOCA-Host 3.5.0-082, NetQ 5.1.0, NCCL 2.30.7.[19] The DOCA-OFED host profile is Level 1 against firmware; any other profile or service is Level 2.[18][29]

Offload arithmetic supports the argument but does not make it. NVIDIA’s own energy blog quotes partner measurements on BlueField-2 (Ericsson server CPU power down 24 percent; VMware “freeing up 20% of the CPU’s resources”; Red Hat networking demand on CPUs cut 70 percent) and labels them as partner tests; present them as such, never as your measurement on the customer’s workload.[28]

Scenario:

HGX H100/H200 pod on Spectrum-4. East-west wants a SuperNIC in NIC mode (BF-3 B3140H on Hopper, ConnectX-8 on Blackwell), 1:1 with GPUs; the RA still puts one BF-3 in DPU mode per node for north-south. ⚠ Core count (2× 56c Xeon) and 8% infra share are estimates.

Workload
Features needed
Assumptions and sources (defaults are estimates — every field is editable)
  • Nodes = 200 — Fleet size.
  • Host cores per node = 64 cores — Physical cores across both sockets (e.g. 2× 32c = 64).
  • Cores consumed by infrastructure = 20 %estimate — Share of host cores spent on vSwitch/SDN agents, storage initiator, encryption, telemetry. NVIDIA marketing says "up to 30%"; hyperscaler blogs quote 25–40% — both secondary. ESTIMATE, measure your own with perf/top. source ↗
  • Server cost per core-year = $120estimate — e.g. $30k 2-socket server ÷ 64 cores ÷ 4-year life ≈ $117. ESTIMATE.
  • Per-core software licence per year = $0estimate — Hypervisor / SDN stacks licensed per core (VMware VCF is per-core). Default 0 — enter the customer's quote; this lever dominates in virtualization farms. ESTIMATE.
  • Watts per core = 6 Westimate — ≈ 350–400 W TDP ÷ 64 cores ≈ 6 W at load. ESTIMATE.
  • Electricity price = 0.12 $/kWhestimate — US commercial average ≈ $0.12–0.13/kWh (EIA). ESTIMATE — use the site tariff. source ↗
  • PUE = 1.3estimate — Facility overhead multiplier on IT watts. 1.3 is a typical enterprise value. ESTIMATE.
  • DPU price delta vs ConnectX = $2500estimate — Dell configurator snippets: BF-3 B3140H SuperNIC $6,213.80 (R7725), Partner DPU dual-port 400GbE $8,347.83 (R670) → ≈ $2.1k delta; ConnectX-7 400G list not fetched. UNVERIFIED ESTIMATE. source ↗
  • Extra watts per DPU vs NIC = 60 Westimate — B3220/B3240 are 150 W cards vs 75 W for the B3140H SuperNIC (Dell R7725 ISM). Real delta depends on load; 60 W is an ESTIMATE. source ↗
  • DPUs per node = 1 — HGX B300 RA uses one north-south BF-3 per node; VMware hosts typically one. source ↗
  • OVS / vSwitch datapath offload: frees 20% of infra cores in NIC mode, 35% in DPU mode estimate — ASAP² (tc-flower / OVS-DOCA) offloads the datapath on ConnectX too; DPU mode additionally moves ovs-vswitchd and the SDN agents to the Arm. source ↗
  • Connection tracking (CT) offload: frees 10% of infra cores in NIC mode, 15% in DPU mode estimate — CT offload is an ASIC feature (ConnectX-6 Dx and later); OVS-DOCA 3.5.0 adds IPv6 CT. source ↗
  • IPsec / TLS inline crypto: frees 15% of infra cores in NIC mode, 20% in DPU mode estimate — BF-3 datasheet: "MACsec/IPsec/TLS data-in-motion" inline. Crypto is SKU-dependent ("selective on SuperNICs") — check the OPN. source ↗
  • NVMe-oF / SNAP storage emulation: frees 0% of infra cores in NIC mode, 20% in DPU mode estimate — BlueField SNAP (NVMe / virtio-blk emulation) and the storage initiator run on the Arm cores → DPU mode only. source ↗
  • Tenant isolation / zero-trust: frees 0% of infra cores in NIC mode, 5% in DPU mode estimate — Zero Trust (Restricted) is a DPU-mode variant: the host admin loses port ownership, HW counters, tracer, RShim and FW flashing. source ↗
  • HBN (BGP/EVPN router on the DPU): frees 0% of infra cores in NIC mode, 5% in DPU mode estimate — HBN runs on the Arm side; not supported on 8 GB DDR SKUs. DPU mode only. source ↗
  • Model: incremental cores = infra cores × (DPU share − NIC share); savings = cores × ($/core-year + licence) + (cores × W/core − DPUs × extra W) × PUE × 8.76 kWh × $/kWh; payback = premium ÷ monthly savings. Recommendation rules: any Arm-only feature → DPU mode; no features + east-west AI → SuperNIC/ConnectX-8 in NIC mode; no features → plain ConnectX; else DPU mode only if payback ≤ 36 months.
Full calculator. Switch scenarios (training-pod, vmware-farm, storage-inference) and watch which line item moves the answer between NIC mode, DPU mode and ConnectX-8.

2Scenario (a): a VMware farm

Default answer: a plain ConnectX or a BlueField-3 in NIC mode. Broadcom’s KB describes NIC mode as “BlueField behaves exactly like an adapter card from the perspective of the external host”, and NVIDIA’s modes page adds the power and memory footprint benefits.[2][1] NVIDIA’s ESXi HowTo gives the exact flip: MFT on ESXi, esxcli system module parameters set -m nmlx5_core -p mst_recovery=1 with reboot -f, /opt/mellanox/bin/mlxconfig -d mt41692_pciconf0 set INTERNAL_CPU_OFFLOAD_ENGINE=1, then a power cycle, then verify the offload engine shows DISABLED.[31]

DPU mode is worth positioning only when the customer wants vSphere Distributed Services Engine, which runs ESXio on the Arm cores to offload NSX switching, routing and firewalling. Broadcom’s rules: the card leaves the factory in DPU mode with DSE disabled; DSE needs vSphere 8.0 U3b or later; and “when you switch BlueField-3 from NIC mode to VMware Distributed Services Engine enabled DPU mode, you must re-install ESXi” (both directions).[2]

On PowerEdge the DSE conversation has a Dell gate. KB 000225111: “VMware’s Distributed Services Engine is not supported on NVIDIA Channel cards by Dell. Dell has not validated this feature and is unable to support it, if it is enabled.”[3] So a consignment or NVIDIA-channel BlueField-3 in an R760 farm is a NIC-mode card; DSE needs a Dell-sourced DPU SKU. Dell also delivers BlueField NIC firmware for ESXi hosts as an iDRAC/Lifecycle-Controller-only DUP, at least for BlueField-2; executing it from the host OS is not supported.[4]

Position it in one sentence: “NIC mode today; DSE only if Dell qualifies the DPU SKU for your platform, and plan the ESXi reinstall.”

3Scenario (b): an AI training pod

East-west is decided by the GPU generation. On HGX B300 the reference architecture specifies eight ConnectX-8 SuperNICs per baseboard at up to 800 Gb/s with a 1:1 GPU-to-NIC ratio; the guidance lists eight 400G NICs as the minimum and sixteen 400G via breakout as recommended.[5] On Hopper with Spectrum-4, the BlueField-3 SuperNIC at 400 Gb/s in NIC mode carried east-west, as in the Israel-1 reference on Dell XE9680.[9][7] The SuperNIC page places the three generations: BlueField-3 up to 400 Gb/s, ConnectX-8 up to 800 Gb/s, ConnectX-9 up to 1.6 Tb/s per GPU.[7]

North-south is one BlueField-3 DPU per node, B3240 recommended, in ECPF/DPU mode, for storage, management and tenant isolation; it “requires 8-pin ATX 12V PCIe” power.[5]

The proof points are in the Spectrum-X whitepaper. Static ECMP hashing collapses AI elephant flows to about 60 percent of throughput; adaptive routing picks “the port with the minimal load over its egress queue” per packet, sprays, and DDP in the SuperNIC “transforms any out-of-order data at the RoCE transport layer, transparently delivering in-order data to the application”, lifting effective bandwidth to 95 percent.[6] Incast is different: “RoCE Adaptive Routing will not solve incast congestion since there is no way to route around it”; the answer is congestion control driven by in-band switch telemetry, executed in the SuperNIC at millions of events per second.[6] The platform page adds 1.6x over standard Ethernet and Multiplane scaling to 128,000 GPUs.[25]

If the customer is single-tenant and only wants a fast NIC, a SuperNIC (B3140H or ConnectX-8) in NIC mode is the right answer and a DPU is the wrong one; the DPU earns its place only for offloaded services on the north-south card.[1][7]

4Scenario (c): storage-heavy inference

Here the DPU runs in DPU mode as a storage controller. DOCA 3.5.0 adds “support for B3220SH Storage Controller” in OVS-DOCA, and HBN 3.5.0 supports it too.[10][16] Dell’s own story is ObjectScale “with BlueField-3 DPU and Spectrum-4” for S3 over RDMA, with Dell quoting “up to 230% higher throughput, up to 80% lower latency”.[11] If the customer wants BGP/EVPN on the server for a multi-tenant L3 fabric, HBN runs on the Arm and therefore needs the card in DPU or zero-trust mode, not NIC mode where the Arm cores are inactive[1], and HBN 3.5.0 does not support BlueField platforms with 8 GB on-board DDR, so a 16 GB or larger SKU is the floor.[16] The other side of the link must speak BGP/EVPN; Dell’s PowerSwitch SN5610 and SN2201 run Cumulus Linux and SONiC.[34]

Roadmap belongs in this scenario because customers raise it. What is public: BlueField-4 powers the “Inference Context Memory Storage Platform”, a KV-cache tier, with “up to 5x” tokens per second and Dell named among partners, available in the second half of 2026.[12] NVIDIA’s GTC release names DOCA Memos as “a new DOCA framework” for KV-cache storage processing and lists BlueField-4 STX storage racks.[13] The BlueField-4 blog states 800 Gb/s, “6x the compute power” of BlueField-3 and a Grace CPU with ConnectX-9.[14] Dell has said it will integrate BlueField-4 into the Dell AI Data Platform.[15]

What is not public, and what you do not say: BlueField-4 core count and PCIe generation (no fetched NVIDIA page states them), a Dell ship date for a BlueField-4 part, or BlueField-4 support in DOCA 3.5.0, whose release notes list BlueField-3, BlueField-2, first-generation BlueField and ConnectX-9 through ConnectX-4 only.[24] BlueField-4 bring-up is also a different skill: BMC-driven via Redfish, PLDM and UEFI HTTP boot rather than the BlueField-3 RShim and BFB path.[32]

Position it: “BlueField-3 in DPU mode is the shipping storage controller for this design; BlueField-4 is announced for later this year and I can tell you what NVIDIA and Dell have published about it.”

5Objections, published answers, honesty rules

Objection Published answer Honesty rule
“A DPU is a NIC with extra power draw.” Choose the mode: NIC mode drops Arm power and host memory footprint; DPU mode is for offloaded services (HBN, OVS-DOCA, storage, DSE).[1] Do not quote configurator prices or consignment lines; route pricing to the account manager.
“We cannot run Arm.” NIC mode needs no Arm management. At scale DPF provisions BFBs, firmware parameters and services from Kubernetes; DOCA 3.5.0 pairs with DPF v26.4.0.[27] Say which DPF tag exists; do not invent later ones.
“We already have OVS.” OVS-DOCA is the datapath offload layer (DPIF) for OVS using DOCA Flow; NVIDIA advises it over OVS-kernel and OVS-DPDK.[21] Name the limits on the same page: ct-ct-nat unsupported, mirror configurations default to 128 and are raised with doca-mirror-max, Geneve one option; the forum shows a mirror once dropped offload to software.[20][22]
“Lock-in.” DOCA_STABLE APIs are supported for the major version; the C ABI is versioned and binary-compatible across minor and major releases; protocol compatibility allows rolling upgrades; the DOCA-OFED profile is Level 1.[17][18] Also say: DOCA 3.6.0 is announced as the last release for ConnectX-4 Lx and ConnectX-5; Level 2 breaks at October GA.[10][18]
“Wait for BlueField-4.” 800 Gb/s, about 6x BlueField-3 compute, Grace plus ConnectX-9, DOCA Memos, second half of 2026, Dell integration announced.[14][13][15] Quote the notice that NVIDIA documents are “not a commitment to develop, release, or deliver any Material”; decline core counts and dates.[17][26]
“Dell will not support it.” Dell’s Partner DPU is “supported by Dell Technical Support when used with a Dell system”; the exceptions are DSE and NVIDIA-channel cards.[23][3] Quote Dell’s page, not your reading of it.
“Ethernet cannot do AI.” Adaptive routing plus DDP plus telemetry-based congestion control: about 60 to 95 percent effective bandwidth; 1.6x over standard Ethernet.[6][25] Whitepaper numbers are NVIDIA’s test; offer to measure on the customer’s fabric.

Three phrasings keep every answer inside the evidence: “NVIDIA has stated”, “Dell’s page says”, and “not confirmed by a page I can cite”. Dates and prices go to the account manager or Dell product management; Dell publishes its own windows and its own support statements.[11][3]

Positioning: three customers, one afternoon · decision 1/7A 0 · P 0 · S 0

Brief — Back-to-back calls: a VMware farm (200× R760), an AI training pod (64× XE9780 on Spectrum-X) and a storage-heavy multi-tenant inference cluster. Each has objections.

Dell account team + three end customers: VMware farm: "200 R760s on vSphere 8; NSX eats about 20% of our cores. Should we go DPU?"

Run the positioning scenario against a simulated customer who raises the objections above; the score rewards cited answers and penalises unpublished claims.

6Positioning memo: worked, faded, problem

Positioning memo: worked, faded, problem

To: Dell SE, account team. Re: BlueField configuration for the customer’s HGX B300 training pod (XE9780 class).

Recommendation. East-west: eight ConnectX-8 SuperNICs per node at 800 Gb/s, 1:1 with GPUs, per the HGX B300 reference architecture. North-south: one BlueField-3 B3240 in DPU mode for storage, management and tenant isolation; it requires 8-pin ATX 12V PCIe power.[5]

Why not BlueField-3 SuperNICs east-west. That was the Hopper design (Israel-1 on XE9680 with Spectrum-4); for B300 the RA and the validated-stack B300 rows track ConnectX-8 firmware.[9][19]

What the fabric buys. Adaptive routing plus DDP lifts effective bandwidth from about 60 percent to 95 percent; telemetry-based congestion control handles incast that routing cannot.[6]

Versions. Pin to validated stack v2.3.1: Cumulus 5.18.1, DOCA-Host 3.5.0-082, NetQ 5.1.0, NCCL 2.30.7; BlueField-3 firmware 32.50.1002 on the north-south card. Services on the DPU are Level 2: firmware, BFB and services move together.[19][18]

Objections raised. OVS: OVS-DOCA is the advised datapath; ct-ct-nat is not offloaded.[21][20] Lock-in: DOCA_STABLE APIs for the major version and a versioned ABI; DOCA 3.6.0 is the last release for ConnectX-4 Lx and ConnectX-5, which does not affect this design.[17][10] BlueField-4: NVIDIA has stated 800 Gb/s, about 6x BlueField-3 compute and second-half-2026 availability; Dell has announced Dell AI Data Platform integration; none of this is a commitment and BlueField-4 is not in DOCA 3.5.0’s supported list.[14][15][24]

Dell support. Partner DPU is supported by Dell with a Dell system; no DSE in this design.[23]

Not confirmed by a page I can cite: Dell auxiliary-power cable kit part numbers for the XE9780 riser; the XE9780 DPU table.

What ends up in the memo

How it ended

The clinical farm resolves first: NIC mode, because Dell has not validated vSphere DSE on NVIDIA-channel cards and cannot support it if enabled; switching between NIC mode and DSE-enabled DPU mode costs an ESXi reinstall in both directions.[3][2] The storage lead gets three sentences of BlueField-4: 800 Gb/s, about six times BlueField-3 compute, Grace with ConnectX-9, second half of 2026, and none of it a commitment to deliver anything.[14][17] Every version in the memo is pinned to validated stack v2.3.1.[19] The notebook gains two pages; nobody asks for a counter.

What you say: “Tell me what has to run on the card and I will tell you which of these three you are buying.”[1]

The report is due Friday. So is the interview for the job that owns this account.

Lab

Mutating steps ahead. Needs a maintenance window, out-of-band access (BMC/iDRAC/rshim console) and a rollback path. Record the pre-flight inventory before changing anything. Never on a production host.

Goal: a five-minute narrated demo for a Dell audience showing the difference between DPU mode and NIC mode on the lab BlueField-3, with an OVS-DOCA offload check, as the visual for the positioning memo. Mode changes require a power cycle; use a written window and rollback.[1][33]

  1. Pre-flight (read-only): sudo mst start && sudo mst status -v; sudo mlxconfig -d /dev/mst/mt41692_pciconf0 q INTERNAL_CPU_OFFLOAD_ENGINE (expect ENABLED(0)); sudo flint -d /dev/mst/mt41692_pciconf0 q; cat /dev/rshim0/misc.[1] If not ENABLED(0): the card is already in NIC mode; adapt the script to start from step 4.
  2. In DPU mode on the Arm, show what “services on the card” means: sudo ovs-appctl dpctl/offload-stats-show and sudo ovs-appctl coverage/show; narrate that this is the datapath OVS-DOCA offloads and where its limits are documented.[20] If no OVS-DOCA bridge exists: say so on camera; do not build one inside the window.
  3. Flip to NIC mode (mutating): sudo mlxconfig -d /dev/mst/mt41692_pciconf0 s INTERNAL_CPU_OFFLOAD_ENGINE=1; full power cycle. Rollback: =0 plus a power cycle.[1]
  4. Verify DISABLED(1); show the host now sees a ConnectX-class adapter and the rshim console is silent because the Arm is inactive; narrate the power and memory-footprint argument from the modes page.[1][31]
  5. Roll back to DPU mode, power cycle, wait for INFO[MISC]: DPU is ready, re-run step 2. Expected: offload statistics return. If not: capture the console and follow the escalation lesson.[1]
  6. Close by stating, on camera, which scenario each mode fits and which Dell KB governs the power cycle you just performed.[33]

Retrieval check

9 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, how you decide between NIC mode, DPU mode and a ConnectX-8 for a new customer, and what you refuse to say about BlueField-4.

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

Sources

Facts in this lesson were checked against DOCA 3.5.0 docs, Spectrum-X whitepaper and validated stack, HGX B300 RA, Dell KBs 000225111/000300192, Broadcom KB 379391, NVIDIA and Dell press releases Jan-Mar 2026, 2026-09-06. Dates are when each page was fetched.

  1. BlueField Modes of Operation · fetched 2026-09-06 · DOCA 3.5.0
  2. Broadcom KB 379391: Configuring BlueField-3 into DPU mode for vSphere · fetched 2026-09-06
  3. Dell KB 000225111: NVIDIA Channel DPU VMware DSE Support · fetched 2026-09-06
  4. Dell driver 4RR0M: NIC firmware for NVIDIA BlueField-2 DPUs (DUP) · fetched 2026-09-06
  5. NVIDIA HGX AI Factory Enterprise Reference Architecture: Components · fetched 2026-09-06
  6. NVIDIA Spectrum-X Network Platform Architecture whitepaper (July 2024, third-party mirror) · fetched 2026-09-06
  7. NVIDIA SuperNICs (BlueField-3 / ConnectX-8 / ConnectX-9) · fetched 2026-09-06
  8. NVIDIA ConnectX-8 SuperNIC User Manual: Introduction · fetched 2026-09-06
  9. NVIDIA Spectrum-X available from Dell, HPE, Lenovo (Nov 2023) · fetched 2026-09-06
  10. DOCA 3.5.0 Changes and New Features · fetched 2026-09-06 · DOCA 3.5.0
  11. Dell Technologies Unveils Next Generation Enterprise AI Solutions with NVIDIA (May 2025) · fetched 2026-09-06
  12. NVIDIA BlueField-4 Powers New Class of AI-Native Storage Infrastructure (CES, Jan 2026) · fetched 2026-09-06
  13. NVIDIA Vera Rubin Opens Agentic AI Frontier (GTC, Mar 16 2026) · fetched 2026-09-06
  14. NVIDIA Launches BlueField-4 (NVIDIA blog) · fetched 2026-09-06
  15. Dell Uplevels AI Infrastructure With NVIDIA at CES (Jan 2026) · fetched 2026-09-06
  16. HBN Service Release Notes (HBN 3.5.0) · fetched 2026-09-06 · DOCA 3.5.0
  17. DOCA Backward Compatibility Policy (PDF) · fetched 2026-09-06 · DOCA 3.5.0
  18. DOCA Dependency Compatibility Policy · fetched 2026-09-06 · DOCA 3.5.0
  19. NVIDIA Spectrum-X Validated Solution Stack · fetched 2026-09-06 · DOCA 3.5.0
  20. OVS-DOCA Hardware Acceleration · fetched 2026-09-06 · DOCA 3.5.0
  21. OpenvSwitch Acceleration: OVS in DOCA · fetched 2026-09-06 · DOCA 3.5.0
  22. NVIDIA Developer Forums: Hardware offload lost when applying mirror in DOCA · fetched 2026-09-06
  23. Dell Partner DPU, NVIDIA BlueField-3 Dual Port 400GbE (352-BBFH) · fetched 2026-09-06
  24. DOCA Release Notes v3.5.0 · fetched 2026-09-06 · DOCA 3.5.0
  25. NVIDIA Spectrum-X Ethernet Platform · fetched 2026-09-06
  26. NVIDIA Kicks Off the Next Generation of AI With Rubin (CES, Jan 5 2026) · fetched 2026-09-06
  27. DOCA Platform Framework (DPF) · fetched 2026-09-06 · DOCA 3.5.0
  28. Tests Show Energy Efficiency of BlueField DPUs (NVIDIA blog, Nov 2022) · fetched 2026-09-06
  29. DOCA Profiles · fetched 2026-09-06 · DOCA 3.5.0
  30. NVIDIA BlueField-3 Networking Platform User Guide: Specifications · fetched 2026-09-06
  31. HowTo Configure NVIDIA BlueField-3 to NIC Mode on VMware vSphere 8.0 · fetched 2026-09-06
  32. NVIDIA/skills: doca-bf4-deployment CAPABILITIES · fetched 2026-09-06
  33. Dell KB 000300192: XE9680L power cycle required after updating DPU firmware · fetched 2026-09-06
  34. Dell AI Factory with NVIDIA Delivers Proven Path to Enterprise AI ROI (Mar 2026) · fetched 2026-09-06

The same idea elsewhere

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