Skip to content
nodeza.learn
DOCA course/9. FAE capstone/Lesson 9.4

Mock interview and 90-day plan

S9·E4Forty minutes, one whiteboard, and your own case on the board · NVIDIA briefing room, Friday of week five, the third of five interviews

S9·E4Evaluate~60 minsources checked todayverified against DOCA 3.5.0 docs, Spectrum-X whitepaper and validated stack v2.3.1, Dell KBs and manuals, NVIDIA forum threads, NVIDIA and Dell press releases, learning-science abstracts, 2026-09-06

Builds on: Scenario: a qualification request, Scenario: escalations, Scenario: sizing and positioning

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

  • Answer twenty technical and situational FAE interview questions from published sources, in under three minutes each, and grade your own answers for mechanism rather than keywords.
  • Run a timed, closed-book self-test protocol with confidence ratings, and schedule the second pass by the evidence on spacing and retrieval practice.
  • Structure a situational answer as symptom, questions, evidence, boundary, next step.
  • Write a 90-day onboarding plan for the Dell account with named deliverables and the sources each one rests on.

Episode 4 — Forty minutes, one whiteboard, and your own case on the board

The situation · NVIDIA briefing room, Friday of week five, the third of five interviews

The PoC report went out at 07:04, the Dell SE’s promise spreadsheet finally all closed. Six hours later, a different room. The third interviewer draws two switches and eight GPUs and asks why static hashing collapses AI throughput.

You answer the mechanism: elephant flows collide on the same egress port, adaptive routing picks the port with the minimal load over its egress queue per packet, and DDP reorders the sprayed packets in the SuperNIC at the RoCE transport layer, moving effective bandwidth from about 60 to 95 percent.[2] The follow-up is not a mechanism question but a situation: a consignment BlueField-3 in an R760, DSE enabled, a Dell ticket already open, and a customer waiting for you to say it will be fixed. You know this one: it is the clinical farm from week three, serial numbers filed off.

That pair is the job. The posting names both halves: a technical specialist supporting account managers, on-site bring-up, benchmarking and issue replication, and ownership from design-in to end-of-life.[1] Half of what they test is how the hardware works; the other half is whether you respect the boundaries in front of a customer.

Retrieval practice exists because rereading feels like learning and is not. Prior testing produced substantially greater retention at two days and one week than restudying, even though repeated studying raised confidence more.[26] Judgments of learning track fluency, not knowledge, so the smoothest answer is the one to check.[28] Confidence is the least reliable instrument in the room.

You take the marker back and start with what you would ask.

1What NVIDIA is hiring for, and how to practise for it

NVIDIA’s own posting for a senior networking FAE lists the duties: be “a technical specialist on networking products, directly supporting sales account managers”; “establish and nurture technical relationships with engineers, management, and architects”; “provide onsite support to do system bring-up, benchmarking/testing/issue replication”; and “lead the product through its entire lifecycle, from the design-in phase to end-of-life”.[1] The requirements name InfiniBand and Ethernet, RDMA and RoCE, BlueField and DOCA, high-speed board design and qualification testing, C/C++ and Python, and embedded Linux; experience with ODMs and EMs stands out.[1]

The interview will probe both halves: mechanism questions (how adaptive routing avoids ECMP collisions) and situational ones (a consignment card with DSE enabled and a Dell ticket open). The twenty questions in the next segment cover both. Practise them with this protocol, which follows the evidence rather than habit.

  1. Closed book. Twenty questions, three minutes each, sixty minutes total. Roediger and Karpicke found that at two days and one week, prior testing “produced substantially greater retention than studying, even though repeated studying increased students’ confidence”.[26] Dunlosky and colleagues rate practice testing and distributed practice high utility; rereading and highlighting rate low.[30]
  2. Before revealing each model answer, write a confidence rating from 1 to 5. Learners’ judgments of their own learning are “driven by current performance and the subjective sense of fluency” and are often wrong; the high-confidence errors are the items to review first.[28]
  3. Grade for mechanism, not keywords. An answer that names DDP but cannot say what it does scores zero.
  4. Interleave. Shuffle the mechanism, Dell, escalation and roadmap categories in one sitting. Interleaved practice beat blocked practice with effect sizes of 0.42 at one day and 0.79 at thirty days.[29] Varying the conditions of practice and using tests as study events are the “desirable difficulties” that pay off only when you already have the background, which is why this lesson comes last.[31]
  5. Space the second pass at least two days out, and widen the gap as the interview date recedes: across 839 assessments, the interval producing maximal retention increased with the retention interval.[27]

2The twenty questions

Each question has a model answer built from a published source. Cover the model, answer aloud against a three-minute timer, rate your confidence, then compare.

Twenty questions: worked, faded, problem
  1. Explain how Spectrum-X adaptive routing avoids ECMP collisions. Model: per packet, the switch selects the port with the minimal load over its egress queue within the packet’s traffic class, informed by neighbour-switch status; packets of one flow take different paths and DDP in the SuperNIC reorders them at the RoCE transport layer, moving effective bandwidth from about 60 to 95 percent.[2] Yours: ____
  2. Why does adaptive routing not fix incast? Model: there is no alternate path around a single receiver; incast needs per-endpoint rate metering, which Spectrum-X does with in-band switch telemetry processed by the SuperNIC at millions of events per second.[2] Yours: ____
  3. SuperNIC versus DPU: what changes physically? Model: same silicon family; SuperNIC SKUs ship in NIC mode with the Arm inactive (8 cores, 16 GB); DPU SKUs ship in DPU mode with the Arm-hosted vSwitch owning the eSwitch (16 cores, 32 GB, 8-pin auxiliary power).[3][4] Yours: ____
  4. Which NIC goes with HGX B300, in what count and mode? Model: eight ConnectX-8 at 800G, 1:1 with GPUs, east-west; one BlueField-3 (B3240 recommended) in ECPF/DPU mode north-south, which needs 8-pin ATX 12V PCIe power.[5] Yours: ____
  5. What is in Spectrum-X validated stack v2.3.1? Model: Cumulus 5.18.1, BlueField-3 FW 32.50.1002, DOCA-Host 3.5.0-082, BFB 3.5.0, NetQ 5.1.0, NCCL 2.30.7, HPC-X 2.51, Network Operator 26.7.0, DTS 1.26.5 (September 2026).[6] Yours: ____
  6. A customer wants to upgrade DOCA-Host but keep last year’s BlueField firmware. Allowed? Model: only for the DOCA-OFED profile, which is Level 1 against firmware and BF-Bundle; other profiles and DOCA services are Level 2 and must stay within one October-to-July cycle.[7] Yours: ____
  7. What do DOCA_STABLE and semantic versioning guarantee? Model: stable APIs survive the major version; the C ABI is versioned and binary-compatible across minor and major releases; source compatibility across minor and patch; protocol compatibility permits rolling upgrades.[8] Yours: ____
  8. How do you switch a BlueField-3 to NIC mode on ESXi? Model: MFT on ESXi; esxcli system module parameters set -m nmlx5_core -p mst_recovery=1 and reboot -f; /opt/mellanox/bin/mlxconfig -d mt41692_pciconf0 set INTERNAL_CPU_OFFLOAD_ENGINE=1; a full power cycle; then mlxconfig -d mt41692_pciconf0 q | grep -i offload shows DISABLED.[9] Yours: ____
  9. A customer enabled DSE on a consignment BlueField-3 in an R760 and opened a Dell ticket. What happens? Model: Dell KB 000225111 says DSE on NVIDIA Channel cards is not validated and Dell cannot support it; route them to NIC mode or a Dell-SKU DPU and plan the ESXi reinstall.[10] Yours: ____
  10. Post-BIOS-update “No Memory Found” on an XE9680 with BlueField-3. Dangerous? Model: known Dell KB 000227031 state-machine timing issue at CPLD 1.1.5/1.1.7, iDRAC 7.10.50.00 and firmware 32.40.1000 in DPU mode; the server reboots itself after about a minute; the fix lands in a CPLD release.[11] Yours: ____
  11. The LC log shows “Device not detected: Nvidia Network Adapter” PR8 after every boot. Model: Dell KB 000379421, a BlueField-3 PCIe training bug below firmware 32.46.3048; power cycle, update firmware plus matching BFB, else a replacement card.[12] Yours: ____
  12. DPU firmware updated, host rebooted, fatal PCIe error in the LC log. Model: a warm reboot is insufficient; NVIDIA and Dell KB 000300192 require a full system power cycle after DPU firmware updates.[13] Yours: ____
  13. bfb-install succeeded but no ssh to 192.168.100.2. Model: open the rshim console, SW_RESET 1 if silent, read BL2/BL31/UEFI output, confirm firmware and BFB alignment, check the host tmfifo address and route.[14][32] Yours: ____
  14. mlx5_core “Firmware over 120000 MS in pre-initializing state” on a B3210. Model: check the 8-pin PCIe auxiliary cable; the 150 W card halts in the bootloader with “ATX power not detected”; never use an EPS12V cable.[15] Yours: ____
  15. An OVS mirror killed hardware offload. Bug? Model: NVIDIA’s forum answer called it expected for the OVS-DOCA offload provider (multi-output flows) and not a hardware limit, with tc-flower mirroring as the workaround; check the customer’s OVS-DOCA page version before repeating it.[16] Yours: ____
  16. rshim service runs but /dev/rshim0 is missing on a BMC-equipped DPU. Model: the card’s BMC owns RShim by default; release it on the BMC side and enable rshim on the host.[17][32] Yours: ____
  17. Which Dell 17G servers list BlueField-3, and where does it sit? Model: R7725: B3220 and B3240 in RC 2/3/4/5/8, slots 2 and 7, both 150 W FHHL PCIe Gen5 x16; the B3140H is in the same table at 75 W but its Compatible Risers column reads TBD and it carries the asterisk meaning it may not be available at launch. R770: B3220 and B3140H in RC 1 or RC 2 (slots 31 and 36) and RC 6-2 or RC 11-2 (slots 7 and 2); in RC 11-2 slot 5 also takes the B3140H when slots 7 and 2 hold RTX Pro 6000 GPUs.[18][19] Yours: ____
  18. What should you tell a customer about BlueField-4 today? Model: publicly 800 Gb/s, about 6x BlueField-3 compute, Grace plus ConnectX-9, the KV-cache storage platform with DOCA Memos, partners including Dell, second half of 2026, plus NVIDIA’s non-commitment notice; nothing beyond that, and it is not in the DOCA 3.5.0 supported list.[20][21][22][8][38] Yours: ____
  19. What does HBN require from the leaf and the DPU? Model: DPU or zero-trust mode, a 16 GB or larger SKU, and a BGP/EVPN-capable ToR such as Dell SN-series on Cumulus or SONiC.[23][24] Yours: ____
  20. A prospect says DOCA 4.0 will strand our ConnectX-5 fleet. Model: true; DOCA 3.6.0 (October 2026) is announced as the last release to support ConnectX-4 Lx and ConnectX-5; plan the split between LTS branches and new hardware.[25] Yours: ____

3Situational answers: the shape of a good one

Situational questions test whether you can act inside boundaries. Five parts, under three minutes:

  1. Restate the symptom in one line, with the versions you would need. “Not detected” and “does not boot” are different faults with different owners.[12][15]
  2. Say what you would ask: server and riser, SKU and sourcing, mode, firmware, BFB, DOCA-Host profile, BIOS/iDRAC/CPLD, what changed last.[11][7]
  3. Say what read-only evidence you would run before anything else: lspci -d 15b3:, mlxconfig q, flint q, cat /dev/rshim0/misc, the console capture, dmesg for mlx5_core.[32] Name what you would refuse to do first: no reflash, no RMA, no mode flip without a window and a rollback.[35]
  4. Name the boundary. Dell’s KB owns the DSE answer and the power-cycle rule; NVIDIA Enterprise Support owns a bricked card; the non-commitment notice owns every roadmap question; Dell product management owns dates and prices.[10][13][33][8]
  5. Commit to a next step and a date: “I would send the evidence request tonight and the runbook tomorrow.”

Applied to question 9: “A consignment BlueField-3 in an R760 has DSE enabled and a Dell ticket is open. I would first confirm it is an NVIDIA-channel card, because Dell KB 000225111 says Dell has not validated DSE on those and cannot support it if enabled. I would tell the customer that before Dell does, offer NIC mode via the ESXi mlxconfig flip and a power cycle, and explain that DSE requires a Dell-sourced DPU SKU and an ESXi reinstall on the mode change. Next step: a one-page note to the customer and the Dell SE today.”[10][9]

Applied to question 18: “I can tell the board what NVIDIA and Dell have published: 800 Gb/s, about six times BlueField-3 compute, Grace plus ConnectX-9, the KV-cache storage platform with DOCA Memos, second half of 2026, and Dell’s announced Dell AI Data Platform integration. NVIDIA’s documents state none of this is a commitment, and BlueField-4 is not in DOCA 3.5.0’s supported list. For a Dell ship date I would bring in Dell product management.”[20][22][37][8][38]

4The 90-day plan for the Dell account

A 90-day plan turns the posting’s duties into dated deliverables. The structure below is a template; the durations are yours to propose to your manager and the account manager, not a published figure.

Rendering diagram…
Diagram source (Mermaid)
flowchart LR; A[Days 1-30 Learn the account and inventory the lab]-->B[Days 31-60 Own one qualification or escalation end to end]-->C[Days 61-90 Lead a Dell-facing deliverable and propose next quarter]
Three phases; each ends in a written artefact Dell can use.
90-day plan: worked, faded, problem

Days 1 to 30: learn and inventory.

  • Meet Dell platform, BIOS, iDRAC and validation contacts; record who owns which KB.[1]
  • Inventory the lab: every BlueField-3 by SKU, mode, firmware, BFB, DOCA-Host profile, riser and aux power; every server by BIOS, iDRAC and CPLD level.[3][4][11]
  • Map the inventory to validated-stack rows and to the four Dell KBs (000227031, 000379421, 000300192, 000225111); list every card below firmware 32.46.3048.[6][12]
  • Deliverable: a one-page state-of-the-lab with KB exposure and a version-alignment proposal.

Days 31 to 60: own one workstream.

  • Take one qualification request or one live escalation from intake to closure using the scoping and triage runbooks; publish the ask/run/do-not sheets for the six field patterns to the Dell team.[14][15][17]
  • Run the twenty-plus power-cycle soak on one representative server and record the evidence rows.[11][13]
  • Deliverable: a closed case or a completed matrix with log excerpts, plus the runbook set.

Days 61 to 90: lead and propose.

  • Present a positioning memo for one Dell customer (VMware farm, training pod, or storage-heavy inference) with the objection table and honesty rules.[5][10][20]
  • Propose next quarter: the DOCA 3.6.0 October boundary (Level 2 lock-step; last release for ConnectX-4 Lx and ConnectX-5) and the Dell platforms whose BlueField-3 support is not yet confirmed by a page you can cite (XE9780 DPU table).[25][7]
  • Deliverable: the memo, a Q-plus-1 plan, and a list of open questions for Dell product management.

The answer you give, and the case you close

How it ended

You do not promise a fix. You confirm the card is NVIDIA-channel, quote Dell KB 000225111, which says DSE is not validated on those cards and Dell is unable to support it if enabled, offer NIC mode through the ESXi mlxconfig flip and a power cycle, and name the ESXi reinstall that any mode change costs.[10][9] Then a next step: a one-page note to the customer and the Dell SE today. The interviewer writes for longer than the answer took.

What you say: “I would rather tell them tonight than have Dell tell them tomorrow.”[10]

At 18:00 the customer signs the PoC report. The network lead’s notebook closes full; the night-shift operator sends a photo of the riser cable, still labelled “PCIE, NOT EPS”. Five weeks, one fabric, case closed.

Lab

Optional, with a colleague and the lab BlueField-3 available for a read-only walkthrough (no configuration changes).

  1. Whiteboard BlueField-3 internals from memory: Arm cores and DDR5, the ConnectX-7-generation NIC core and eSwitch, the DPA, RShim and the BMC, the PCIe Gen5 x16 edge and the 8-pin auxiliary connector on DPU SKUs. Then verify against the specifications page and correct the board.[4]
  2. Read-only confirmation on the card: lspci -d 15b3: -nn, sudo mlxconfig -d /dev/mst/mt41692_pciconf0 q INTERNAL_CPU_OFFLOAD_ENGINE, cat /dev/rshim0/misc; narrate which drawn block each output comes from and which mode the value implies.[3][32] Expected: your colleague can point at the block for each line. If not: redraw the block you skipped.
  3. Have the colleague play Dell PG and ask three of the situational questions from Segment 2; answer in the five-part shape from Segment 3 and have them mark any sentence you could not attribute.[10][12]
  4. Do not change any firmware parameter or mode during this exercise; if the colleague proposes it, practise the refusal with the hardware-safety reasons.[35]

Retrieval check

10 questions from memory. Answer before looking anything up; misses become flashcards.

Explain it to a Dell SE

Explain to a friend outside the industry, in four sentences, how you are preparing for the FAE interview and why you refuse to re-read your notes the night before.

14 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 v2.3.1, Dell KBs and manuals, NVIDIA forum threads, NVIDIA and Dell press releases, learning-science abstracts, 2026-09-06. Dates are when each page was fetched.

  1. NVIDIA: Senior Field Application Engineer, Networking System Hardware Design (posting) · fetched 2026-09-06
  2. NVIDIA Spectrum-X Network Platform Architecture whitepaper (July 2024, third-party mirror) · fetched 2026-09-06
  3. BlueField Modes of Operation · fetched 2026-09-06 · DOCA 3.5.0
  4. NVIDIA BlueField-3 Networking Platform User Guide: Specifications · fetched 2026-09-06
  5. NVIDIA HGX AI Factory Enterprise Reference Architecture: Components · fetched 2026-09-06
  6. NVIDIA Spectrum-X Validated Solution Stack · fetched 2026-09-06 · DOCA 3.5.0
  7. DOCA Dependency Compatibility Policy · fetched 2026-09-06 · DOCA 3.5.0
  8. DOCA Backward Compatibility Policy (PDF) · fetched 2026-09-06 · DOCA 3.5.0
  9. HowTo Configure NVIDIA BlueField-3 to NIC Mode on VMware vSphere 8.0 · fetched 2026-09-06
  10. Dell KB 000225111: NVIDIA Channel DPU VMware DSE Support · fetched 2026-09-06
  11. Dell KB 000227031: No Memory Found event on NVIDIA BlueField-3 enabled PowerEdge servers · fetched 2026-09-06
  12. Dell KB 000379421: BlueField-3 DPU PCIe Initialization Failure · fetched 2026-09-06
  13. Dell KB 000300192: XE9680L power cycle required after updating DPU firmware · fetched 2026-09-06
  14. NVIDIA Developer Forums: BF3 RShim not working (bfb-install done, no ssh to 192.168.100.2) · fetched 2026-09-06
  15. NVIDIA Developer Forums: BlueField-3 (DPU mode) stuck in FW pre-initializing, no host netdevs · fetched 2026-09-06
  16. NVIDIA Developer Forums: Hardware offload lost when applying mirror in DOCA · fetched 2026-09-06
  17. NVIDIA Developer Forums: BlueField visible in lspci, rshim running, but no /dev/rshim and no tmfifo_net0 · fetched 2026-09-06
  18. Dell PowerEdge R7725 Installation and Service Manual: DPU specifications · fetched 2026-09-06
  19. Dell PowerEdge R770 Installation and Service Manual: DPU specifications · fetched 2026-09-06
  20. NVIDIA Launches BlueField-4 (NVIDIA blog) · fetched 2026-09-06
  21. NVIDIA BlueField-4 Powers New Class of AI-Native Storage Infrastructure (CES, Jan 2026) · fetched 2026-09-06
  22. NVIDIA Vera Rubin Opens Agentic AI Frontier (GTC, Mar 16 2026) · fetched 2026-09-06
  23. HBN Service Release Notes (HBN 3.5.0) · fetched 2026-09-06 · DOCA 3.5.0
  24. Dell AI Factory with NVIDIA Delivers Proven Path to Enterprise AI ROI (Mar 2026) · fetched 2026-09-06
  25. DOCA 3.5.0 Changes and New Features · fetched 2026-09-06 · DOCA 3.5.0
  26. Roediger and Karpicke 2006: Test-enhanced learning (abstract) · fetched 2026-09-06
  27. Cepeda et al. 2006: Distributed practice in verbal recall tasks (abstract) · fetched 2026-09-06
  28. Bjork, Dunlosky and Kornell 2013: Self-regulated learning (abstract) · fetched 2026-09-06
  29. Rohrer, Dedrick and Stershic 2015: Interleaved practice improves mathematics learning (abstract) · fetched 2026-09-06
  30. Dunlosky et al. 2013: Improving students' learning with effective learning techniques (abstract) · fetched 2026-09-06
  31. Bjork and Bjork 2011: Making things hard on yourself, but in a good way · fetched 2026-09-06
  32. BlueField Platform Software Troubleshooting Guide: SoC Management Interface (RShim) · fetched 2026-09-06
  33. NVIDIA Developer Forums: BlueField-3 stuck at PSC BL1 START, RShim BFB transfer stalls · fetched 2026-09-06
  34. Dell Partner DPU, NVIDIA BlueField-3 Dual Port 400GbE (352-BBFH) · fetched 2026-09-06
  35. NVIDIA/skills: doca-hardware-safety · fetched 2026-09-06
  36. NVIDIA Kicks Off the Next Generation of AI With Rubin (CES, Jan 5 2026) · fetched 2026-09-06
  37. Dell Uplevels AI Infrastructure With NVIDIA at CES (Jan 2026) · fetched 2026-09-06
  38. DOCA Release Notes v3.5.0 · fetched 2026-09-06 · DOCA 3.5.0

The same idea elsewhere

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