P45: GPU Fleet — pooled inference across rack 3060 / rig 4090 / laptop 5070 Ti
Leverage every GPU in the homelab for batch LLM work instead of serializing everything through the rack 3060 (12GB, ~360 GB/s). Fleet: rack server RTX 3060 12GB (compose ollama, always-on), rig RTX 4090 24GB (~1008 GB/s), laptop RTX 5070 Ti 16GB GDDR7 (~896 GB/s).
Motivating case: the 2026-08-18 palliative-rfi PRISMA stage-2 screening ran ~5,000 title/abstract records through qwen2.5:14b-ctx8k on the 3060 alone at ~15-25 records/min (~3.5-4.5 h wall clock). The 4090 alone should do the same work ~3-4x faster, the 5070 Ti ~2.5-3.5x; pooled, the whole fleet is roughly 6-8x — the same corpus in well under an hour, and stage-3 full-text eligibility (much longer prompts) becomes tractable.
Building blocks that already exist: llm.pool.HostPool (least-loaded fan-out with health + model checks, currently embeddings-only), LLM_OLLAMA_HOSTS env override in llm.config, prisma.llm provider protocol (openai-compat points at any base URL).
Downstream beneficiaries: prisma screening/eligibility (P42-P44), llm RAG tagging batch runner (#575), post-close comment re-farm + re-index for CMS-2026-2377 (#615), golden-set evals (#577).
