Software EBOM + Interface Spec — AI-Tray-2U Rev B
A complete, filled portfolio artifact (no placeholders) — the ODM-facing contract for a concrete node, as you'd hand to Pegatron. Derived from the template in
../EBOM-AND-PROTOCOL-DOCS.md. Status: Approved, owner: Rack SW Platform, last validated: 2026-06-13, applies to rack classrack-ai-01.
1. Node summary
- Model:
AI-Tray-2URev B · 2U · 1× host CPU socket · 4× AI accelerators · 1× 400G NIC · 1× BMC. - Power: typical 1.48 kW, nameplate 2.0 kW (cap), liquid-cooled (cold plates on accelerators + CPU).
- Mgmt: BMC on mgmt VLAN 10; data NIC on tenant VLANs 100–199.
2. Firmware baseline (the reconciliation target — Phase 01/08)
| Component | Baseline | Min (anti-rollback) | Signed by | A/B | Update path | Owner |
|---|---|---|---|---|---|---|
| BIOS/UEFI | 2.4.1 | 2.4.0 | ODM key PGT-BIOS-2026 | yes | Redfish UpdateService (Task) | Pegatron |
| BMC (OpenBMC) | obmc-1.8.0 | obmc-1.7.2 | Platform key | yes | Redfish + JTAG recovery | Us + Pegatron (upstream) |
| Accelerator FW | accel-fw-2.1.0 | accel-fw-2.0.5 | Accel vendor | yes | vendor tool / Redfish | Accel vendor |
| 400G NIC FW | nic-fw-3.2.0 | nic-fw-3.1.0 | NIC vendor | no | in-band agent | NIC vendor |
| PSU FW | psu-1.0.3 | psu-1.0.3 | ODM | no | BMC (PMBus) | Pegatron |
| PCIe switch | pex-4.1 | pex-4.0 | Switch vendor | no | BMC | Microchip |
| CPLD | cpld-0.7 | cpld-0.7 | ODM | no | BMC/JTAG | Pegatron |
3. Software components (Phase 11 compatibility)
| Component | Version | Compat window |
|---|---|---|
| rack-agent (DaemonSet) | 1.5.3 | control-plane ≥ 1.4, ≤ 1.6 |
| metrics exporter | 1.3.1 | Prometheus ≥ 2.40 |
| RackNode operator | 0.9.0 | k8s/RKE2 ≥ 1.27 |
4. Config & security baselines (Phase 09)
- Secure Boot: enabled; PK =
RACKMGR-PK-2026; KEK ={RACKMGR-KEK, MS-UEFI-KEK}; db includes the signed shim/GRUB/kernel; dbx revokesgrub-2.06-cve-2023-pre. - Attestation: required before tenant scheduling; golden PCR set published per (BIOS, BMC, OS) triple in the attestation service; AK enrolled at commissioning.
- Network: mgmt VLAN 10 (control-plane only — hard boundary); tenant VLANs 100–199; QoS HTB with a reserved 10 Gbit management class.
- Power/thermal: derated feed 17.3 kW; rack CDU 130 kW; N+1 power, N+1 pumps.
5. Interface spec (Redfish 2023.2) — the build-to-it contract
- Required resources:
ServiceRoot;Systems/{id}(PowerState,ProcessorSummary,Actions/#ComputerSystem.Reset);Chassis/{id}(Power,Thermal);Managers/{id};UpdateService;EventService;TaskService.Oem/Rackmgron Chassis MUST exposeRackPosition(U),PowerBudgetW. - Sensors (Thermal.Temperatures[]): MUST include
Inlet,CPU1 Temp,Accel0..3 Temp, each withReadingCelsius(units Celsius) andUpperThresholdCritical. A sensor with no reading MUST returnReadingCelsius: null(not omit the entry, not0). - Reset: MUST advertise
ResetType@Redfish.AllowableValues ⊇ {On, ForceOff, GracefulShutdown, ForceRestart, GracefulRestart, PowerCycle}; aGracefulShutdownto a non-responsive host MUST NOT report success. - Firmware:
UpdateService.SimpleUpdateMUST return202+ aTask;TaskStateNew→Running→Completed|Exception; BMC update MUST be A/B with a recovery path. - Auth/transport: HTTPS only; cert chains to
RACKMGR-MGMT-CA; Session tokens; least-privilege roles (ReadOnlyfor telemetry). - Errors:
4xx/5xx+@Message.ExtendedInfo; never200with an error body. - Events: EventService subscriptions + SSE; PSU, thermal, and PCIe-fatal events MUST be emitted.
6. Conformance checks (executable spec — Phase 11)
| Req | Check | Status (2026-06-13) |
|---|---|---|
| §5 resources | GET each; assert presence + @odata.type | PASS |
| §5 sensors | assert Inlet/CPU/Accel0..3 present, units, null for no-reading | PASS |
| §5 Reset | AllowableValues ⊇ set; illegal type → 4xx | PASS |
| §5 firmware Task | submit → 202 + Task; poll to terminal | PASS |
| §5 errors | force error → 4xx/5xx + extended info | FAIL → finding RM-2026-014 |
| §5 events | subscribe + trigger fault → delivery | PASS |
7. Open findings (raised to Pegatron)
- RM-2026-014 (sev: medium): BMC
obmc-1.8.0returns200with an error body on an invalid PATCH instead of4xx(violates §5). Owner: Pegatron FW. Target:obmc-1.8.1. Workaround: client inspects body for@Message.ExtendedInfo(defensive — Phase 03 Ch. 9). - RM-2026-019 (sev: low):
CPU2 Temppresent but unpopulated on single-socket SKUs; agreed to returnnull(done) rather than omit.
8. Change log
- Rev B (2026-06-13): added attestation requirement; bumped accel-fw baseline 2.0.5→2.1.0; added conformance finding RM-2026-014. Rev A (2026-05-20): initial.