Asus esc8000 gpu bios tuning

Skill soulmachine/skills/asus-esc8000-gpu-bios-tuning

Apply a low-latency GPU-serving + VM-passthrough BIOS profile to ASUS ESC8000-E12P (ASMB12/AMI-BMC, Xeon 6/GNR) servers over Redfish — diff current settings, stage, reboot, verify. Use when asked to optimize/tune BIOS for GPU inference/serving, set performance power profile, enable Resize BAR / SR-IOV / disable SNC, or replicate one server's BIOS config onto another ESC8000.From its SKILL.md

Install

npx -y skills add soulmachine/skills --skill asus-esc8000-gpu-bios-tuning

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

3 things to look at

reads credentialsReads from 1 credential source: `BMC_PASSWORD`.
3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
runs commandsInstructs the agent to run 5 commands, including `export BMC_PASSWORD='...'` and 4 more.

SKILL.md

3.3 KB, 820 tokens by cl100k_base, as published. Nobody here has run it

ASUS ESC8000 GPU-Serving BIOS Tuning

Applies a 10-setting low-latency profile + verifies a 4-setting passthrough baseline, over Redfish. Validated on ESC8000-E12P (8× RTX PRO 6000 Blackwell, 2× Xeon 6730P, BIOS 0804). Settings are reversible until you reboot; the tool is dry-run by default.

Quick start

export BMC_PASSWORD='...'
# 1. Dry-run: show what would change on the target (no writes)
python3 scripts/tune_gpu_bios.py --target <BMC_IP>

# 2. Stage only (reversible, no reboot) — settings pending until next boot
python3 scripts/tune_gpu_bios.py --target <BMC_IP> --apply

# 3. Stage + reboot + verify active (the full apply)
python3 scripts/tune_gpu_bios.py --target <BMC_IP> --apply --reboot

Two source modes (pick one)

Profile (default): targets come from scripts/gpu_serving_profile.json (token strings validated on BIOS 0804 / registry S0317.8.4.0).
Template (most robust): copy the live values from an already-tuned reference server — immune to BIOS-version token drift:
```
python3 scripts/tune_gpu_bios.py --target <new_BMC_IP> --from-host <tuned_BMC_IP> --apply --reboot
```
Prefer --from-host when you have a known-good server (e.g. tuning a second box to match the first).

The profile (what & why)

10 tuned + 4 baseline-verified. Full rationale table in REFERENCE.md. Headline: Resize BAR on, EPP/EPB → Performance, Energy-Efficient-Turbo / C1E-promotion off, Workload → I/O-sensitive, Latency-Optimized on, SNC off (1 NUMA/socket), ASPM-IIO off, AC-loss → Power On. Baseline (must already hold): Above-4G on, SR-IOV on, ACS off (fast GPU P2P — flip to on only at VM adoption). VT-d has no BIOS toggle on this platform (always exposed; gate it OS-side).

Workflow checklist

Confirm firmware parity first — same BIOS version on source & target, else tokens won't match (use the asus-esc8000-firmware-upgrade skill). With --from-host, mismatched versions are flagged.
Dry-run; review the diff with the operator.
--apply --reboot. The reboot is GracefulShutdown → poll Off → On (this build has no GracefulRestart). POST takes ~5–6 min (SNC change forces full memory retraining).
Verify prints VERIFIED N/N active. The post-reboot host shows 0/N mid-POST — that's normal, it resolves when the host republishes its BIOS map.

Gotchas (details in REFERENCE.md)

PATCH target is …/Bios/SD, not /Bios/Settings; needs If-Match etag (else 428).
PATCH returns 503 ("inventory processing") while the host is mid-boot — wait for it to settle.
Pace gently / single connection; a freshly-rebooted AMI BMC locks out on rapid auth (AccessDenied).
This only touches BIOS. The intel_iommu=on iommu=pt kernel cmdline is a separate OS-side step.

See REFERENCE.md for the attribute table, rationale, and the full Redfish API notes.

What ships with it: 3 files

12.7 KB alongside SKILL.md, 1 of them executable

scripts/

gpu_serving_profile.json1.5 KB
tune_gpu_bios.pyruns7.0 KB

REFERENCE.md4.2 KB