jinojose/aipackage
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
main
aipackage/specs/001-enterprise-experience/contracts/workstation-hardware-model-baseline.md
T

8.1 KiB
Raw Permalink Blame History

Offline Workstation Hardware and Starter-Model Baseline

Status: Evidence-backed approval draft; no implementation authority.
Evidence date: 2026-07-14.

The first release explicitly excludes AMD GPU acceleration, ROCm, and Strix Halo acceleration. An AMD64 host may still use the truthful CPU-only profile, but no AMD GPU capability is promised.

Hardware Architecture Decision Required

Hardware family CPU architecture Memory / VRAM Storage Current conclusion
NVIDIA DGX Spark ARM64 Grace Blackwell, 20-core CPU 128 GB unified LPDDR5x 1 TB or 4 TB NVMe Not compatible with an AMD64-only Ubuntu ISO; 1 TB variant also contradicts the current 2 TB baseline
GeForce RTX 5090 Workstation Host CPU remains an approved AMD64 system decision; GPU is Blackwell 32 GB GDDR7 GPU memory Repo baseline says 2 TB NVMe; official GPU page does not define host storage Candidate for the first AMD64 validation profile
RTX PRO 6000 Blackwell Workstation Edition Host CPU remains an approved AMD64 system decision; GPU is Blackwell 96 GB GDDR7 ECC Repo baseline says 2 TB NVMe; GPU documentation does not define host storage Candidate for the second AMD64 validation profile

NVIDIA documents DGX Spark as a 20-core ARM64 Grace Blackwell system with 128 GB unified memory and 1 TB or 4 TB NVMe. NVIDIA documents RTX 5090 with 32 GB GDDR7 and RTX PRO 6000 Blackwell Workstation Edition with 96 GB GDDR7 ECC. GPU specifications do not establish the surrounding workstation CPU, RAM, storage, firmware, or motherboard support contract.

Recommended release decision

Keep one customer-facing Offline Workstation product with a universal, hardware-adaptive payload inside each CPU architecture:

  • Offline Workstation amd64 ISO first, supporting compatible AMD64 systems from CPU-only or low-VRAM NVIDIA GPUs through RTX 3050/4090/5090 and RTX PRO-class GPUs. GPU marketing names do not select media; detected VRAM, driver/runtime compatibility, RAM, disk, and capabilities select the safe runtime profile.
  • Offline Workstation arm64 ISO later for DGX Spark, using its supported OS/runtime/container/model stack and independent evidence gate. It is the same Workstation product, not a new commercial tier.

An alternative single multi-architecture boot image is not recommended until a reproducible Ubuntu boot/build design proves it can boot and install both architectures without duplicating most payloads. Removing DGX Spark from the supported hardware list is the other coherent option. Merely relabelling the existing AMD64 artifact as DGX Spark-compatible is prohibited.

Universal AMD64 Capability Profiles

The ISO MUST boot and install the same core platform across profiles. Hardware discovery chooses which local inference services and starter-model backends are enabled. Unsupported capabilities remain installed-but-disabled or omitted with a truthful reason; the installer does not fail merely because a high-end GPU is absent.

Profile Detected evidence Required behavior
core Less than 32 GB RAM or no supported local inference capacity, including AMD GPU hosts in the first release Portal, administration, document/workflow surfaces supported within their own RAM/disk prerequisites; local GPU model serving not promised
cpu-ai At least 32 GB RAM and 8 CPU cores, no usable NVIDIA VRAM Small quantized Ollama inference may use CPU; vLLM and GPU-only functions disabled with explanation
gpu-lite NVIDIA GPU with at least 8 GB detected VRAM Small quantized Ollama chat/embedding enabled; vLLM/fine-tuning disabled; this covers qualifying RTX 3050 variants based on measured VRAM, not name
gpu-starter At least 24 GB detected VRAM Ollama plus approved vLLM/fine-tuning starter capabilities; representative RTX 4090/5090-class validation
gpu-standard At least 48 GB detected VRAM Higher-capacity single-GPU capabilities; representative RTX PRO-class validation

These thresholds reflect the current feasibility checker and are planning baselines, not final support guarantees. The Workstation category remains fixed; these are capability profiles, not Server S/M/L/Max tiers and not commercial entitlements.

System RAM and disk minimums MUST be expressed per capability. A universal ISO may install the core platform on smaller hardware while blocking local models when RAM, VRAM, or free disk is insufficient. The media itself still needs enough target disk for its embedded payload and recovery headroom.

Universal Payload Behavior

  • Embed the common platform and smallest approved starter models once for AMD64.
  • Prefer Ollama small-model inference for gpu-lite and supported CPU fallback.
  • Enable the embedded vLLM Phi-3 snapshot only when the validated VRAM/runtime profile permits it.
  • Do not download a different payload during offline installation.
  • Do not expose different Workstation license tiers based on GPU size.
  • Do not install ROCm or claim AMD GPU acceleration in the first release.
  • Allow separately signed optional model packs after installation, subject to hardware, storage, runtime, license, and compatibility checks.
  • Record the chosen profile and every disabled capability in readiness/support evidence.

Immutable Starter-Model Candidates

Runtime use Current mutable input Approval pin License evidence Status
Ollama chat phi3:mini Ollama manifest sha256:4f222292793889a9a40a020799cfd28d53f3e01af25d48e06c5e708610fc47e9; model layer sha256:633fc5be925f9a484b61d6f9b9a78021eeb462100bd557309f01ba84cac26adf MIT Approved under CT-G1-2026-008; packaged file/notice evidence pending
Ollama embeddings nomic-embed-text Ollama v1.5 manifest sha256:0a109f422b47e3a30ba2b10eca18548e944e8a23073ee3f3e947efcf3c45e59f Apache-2.0 upstream Approved under CT-G1-2026-008; packaged license blob/attribution evidence pending
vLLM chat microsoft/Phi-3-mini-4k-instruct Hugging Face commit f39ac1d28e925b323eae81227eaba4464caced4e MIT Approved under CT-G1-2026-008 for validated 24 GB-or-higher profiles; exact downloaded-file inventory/digests and redistribution notice pending

Pins are approval candidates, not permission to modify payload.env or build media. The release manifest MUST record every actual file/layer digest. Tag and commit identity alone does not replace post-download object verification.

Acceptance Still Required

  1. Product owner confirms universal hardware-adaptive behavior per CPU architecture and chooses a later ARM64 DGX Spark artifact, a proven multi-architecture artifact, or removal of DGX Spark.
  2. Delivery owner approves AMD64 low/mid/high validation machines, including at least CPU-only or sub-8 GB behavior, an 8 GB-class GPU, a 24 GB-class GPU, and a 48 GB-or-higher GPU.
  3. Platform owner approves workstation-runtime-compatibility-baseline.md, resolves its branch-level baseline to exact immutable package/image/object locks, and validates it separately for AMD64 and any future ARM64 artifact.
  4. AI/platform and legal owners approve model revisions, redistribution, attribution, safety/use boundaries, and local inference acceptance prompts.
  5. Firmware/UEFI/Secure Boot, no-GPU, driver-missing, 6 GB, 8 GB, 24 GB, 32 GB, 48 GB-or-higher, insufficient RAM/storage, multi-GPU, and driver rollback cases are added to the disconnected matrix.

Official Evidence

Reference in new issue View Git Blame Copy Permalink