All runbooks
runbookuapiredacted view

UAPI Grok 140T/NPU Primary Stack Review 2026-06-22

Grok review of OpenVINO GenAI as the primary NUC 140T/NPU stack and benchmark blockers.

Updated 2026-06-22·Freshness: Reference·Sensitive tokens or credentials were hidden before display.

UAPI Grok 140T/NPU Primary Stack Review 2026-06-22

Grok review of OpenVINO GenAI as the primary NUC 140T/NPU stack and benchmark blockers.

Promotion Metadata

  • Source feedback id: 112
  • Source feedback ids: 112
  • Promotion request id: 43
  • Feedback category: beneficial
  • Feedback source: [REDACTED:high-entropy]
  • Feedback created at: 2026-06-22T18:22:34.954146+00:00
  • Target slug: [REDACTED:high-entropy]
  • Review state: review
  • Reviewer note: AI lane auto-promoted redacted feedback into internal catalog context.

Source Feedback Body

Feedback 112

  • Product: uapi
  • Category: beneficial
  • Related slug: [REDACTED:high-entropy]
  • Source: [REDACTED:high-entropy]
  • Created at: 2026-06-22T18:22:34.954146+00:00

Grok 140T/NPU Primary Stack Review (2026-06-22)

ALIVE: true
Lane: Grok / UAPI hardware autorouter
Sources: local diary/2026-06-22-*.md ([REDACTED:high-entropy].md, [REDACTED:high-entropy].md, grok-benchmark-preflight-command-plan.md, [REDACTED:high-entropy].md), [REDACTED:high-entropy].md; public primary: openvinotoolkit/openvino.genai, Intel developer articles, docs.openvino.ai.

Primary Stack: OpenVINO GenAI for Arc 140T + NPU

Recommended primary: OpenVINO GenAI (openvino-genai) + OpenVINO Runtime for both NPU ("AI Boost") and Arc 140T iGPU on NUC15CRSU9 (Core Ultra 9 285H).

Source URLs (primary only)

  • GitHub: https://github.com/openvinotoolkit/openvino.genai (Apache-2.0)
  • Docs: https://openvinotoolkit.github.io/openvino.genai/ ; https://docs.openvino.[REDACTED:high-entropy].html
  • Install: pip install openvino-genai
  • Usage: import openvino_genai as ov_genai; pipe = ov_genai.LLMPipeline(model_path, "NPU") or "GPU"
  • Model prep: optimum-cli export openvino --model ... --weight-format int4 ...
  • Intel validation on exact HW: Core Ultra 9 285H + Arc 140T GPU (driver ~32.0.101.x) with openvino-genai (2025.2 dev). Throughput examples published (e.g. Qwen3-30B-A3B MoE ~34 t/s on Arc; NPU support Day-0 for select models). https://www.intel.[REDACTED:high-entropy].html

Confirmed (from sources + inventory match)

  • OpenVINO Runtime explicitly supports CPU + GPU + NPU.
  • GenAI LLMPipeline accepts device="NPU" or "GPU".
  • HW match: NUC has Arc 140T (drv 32.0.101.8331), NPU Intel AI Boost (drivers 32.0.100.x), ~64 GiB RAM. Matches Intel test configs.
  • No external deps for core tokenization/runtime.
  • Python probe in preflight: openvino_genai ABSENT today.
  • License clean for review (Apache-2.0).

Assumptions only (not confirmed)

  • Full Gemma 2B/3B-class support on this NUC's NPU/140T without custom export or limits (model-specific; small SLMs favored for NPU).
  • Performance targets (first_token <3000ms, >15 t/s) achievable on NPU vs 140T (inventory shows shared mem behavior TBD; reported adapterRAM ~2 GiB vs "32GB" label).
  • NUC Windows driver stack (no WSL) will match Intel reference configs exactly.
  • Optimum-Intel export path will be stable for target Gemma variants.

Benchmark blockers (explicit)

  • No packages in default Python 3.13.12 (torch, openvino, openvino_genai, optimum, nncf all ABSENT).
  • Operator gate required for any pip install + weight download + small Gemma/SLM smoke (1-3 prompts, max_new_tokens~32-64).
  • Metrics gate: capture first_token_ms, tokens_per_sec, load_time_s, peak_mem_mb, device, stability. Do not claim local.nuc.gemma.npu or .140t until evidence.
  • NPU known issues from upstream (prompt len limits ~2k tokens in some reports, cache dir, garbled long outputs).
  • C: free ~704 GiB OK; idle confirmed; no concurrent Conductor load.
  • Model IR export + quant step must be validated pre-bench (not just HF direct).
  • No current route impl; candidates only.

What must be benchmarked before any route claim

Small Gemma-class on NPU path and on 140T path. Record device fallback, thermals/power if exposed, VRAM sharing. Re-bootstrap after install.

Blockers: install+bench gate (planning/preflight only per sources), Codex PlacementDecision contracts, no weights yet, SN850X separate.

Sources used: listed local + primary GitHub/Intel docs only. No broad search.

End. Constraints met.