HOMELAB · SEPTEMBER 2026 INDEX
Pick the box class, then the build.
Dual DGX Spark is a Flash-Next NVFP4 pair, not a merged 256 GB GGUF pool. Dual used 3090s are the 48 GB Ampere path. Mini PCs are a different memory class.
Spark / 128 GB class
Flash-Next and other 128 GB-class SKUs. Two Sparks are tensor-parallel, not one 256 GB llama.cpp box.
- How Do You Run Qwen 3.8 Flash-Next on Two DGX Sparks in 2026? Follow Mia's public dual-DGX-Spark vLLM NVFP4 runbook for Qwen3.8-Flash-Next. Two GB10 boxes are tensor-parallel serving, not one 256GB pool. Public benches only — LocalRig has no first-party tok/s for this stack.
- NVIDIA DGX Spark vs 2× RTX 3090 for Local AI: Buy the Box or Build the Rig? NVIDIA DGX Spark and a dual-RTX-3090 workstation solve the same 70B local-AI problem in very different ways. Compare memory, software, power, multi-device scaling, cost, and the two purchase paths before choosing a turnkey Spark or a DIY 3090 rig.
Used 3090 rigs
The 24–48 GB Ampere path for dense 27B. PSU and cooling before a second card.
- Dual RTX 3090 Build Guide 2026: 48GB of VRAM for the Price of One New Card A complete, defensible parts list for a dual RTX 3090 rig: board and PCIe lane selection, PSU sizing for two 350W cards, spacing and thermals, and the honest runtime-complexity cost of going dual-GPU — priced against buying one new flagship card instead.
- PSU Sizing for Multi-GPU AI Rigs: The Wattage Math That Keeps Your House Safe PSU sizing for multi-GPU AI workloads is a safety issue, not a performance optimization. This guide covers transient-spike headroom, the 15A wall-circuit ceiling, and per-build wattage math for single 5090, dual 4090, and dual 3090 setups.
- Quiet Cooling for a Home GPU Server: Taming Blower Cards and 24/7 Noise Home GPU servers are silent until they are not. A guide to the escalation ladder: undervolting, fan curves, passive datacenter shrouds, case selection, and closet airflow math — with honest expectations about dB trade-offs at each step.
- GPU Passthrough on Proxmox for Local AI: IOMMU, VFIO, and the Gotchas Getting a GPU into a Proxmox VM or LXC for Ollama is the most common homelab-AI integration task. A step-by-step guide through IOMMU grouping, VFIO setup, real failure modes, and honest guidance on when bare metal is the simpler answer.
Mini PC / NUC class
Strix Halo, Framework, and mini-PC boxes. Not a 200k 27B default.
- Best Mini PC for Local LLMs in 2026: Unified Memory Changed the Game AMD's Ryzen AI Max ("Strix Halo") put large unified memory pools into mini PCs the size of a paperback, and that changed what "best mini PC for local LLM" means. This guide ranks Beelink, Minisforum, and GMKtec models by the constraint that actually matters — what model size each memory configuration serves at usable speed, not what fits on the spec sheet.
- Minisforum vs Beelink vs GMKtec: Which Mini-PC Brand Can You Actually Trust? Hardware reviewers compare specs. This compares the three mini-PC brands on after-sales support, firmware cadence, and community trust — because a 24/7 inference box failing at year two is a real cost. Honest breakdown of support history and sentiment.
- Beelink GTR9 Pro for Local LLMs: 128GB Unified Memory Under $2,000, Examined The GTR9 Pro puts 128GB of unified memory in a silent mini-PC at $1,899. What does "70B Q5 ready" mean at this bandwidth tier? Honest spec breakdown, comparison to Mac Studio and discrete-GPU builds, and Beelink's support track record for the 70B-curious buyer.
- Framework Desktop as a Local AI Server: Strix Halo, ROCm, and the 128GB Question The Framework Desktop with Strix Halo CPU offers up to 128GB unified memory, full repairability, and x86 Linux native support. But unified memory bandwidth is not GDDR6X, and ROCm tooling is still catching up. This is a guide to what you actually get: the constraints, the community-reported 70B feasibility, and when it trades off speed for serviceability.
- Strix Halo vs Mac Studio for Local LLMs: The Unified-Memory Showdown Strix Halo (AMD Ryzen AI Max, 128GB unified memory, x86) and Mac Studio (M3 Ultra, up to 256GB unified memory, MLX) are the two open paths to running big models without a rack of GPUs. Neither wins outright — this scores both on bandwidth, ecosystem maturity, price-per-GB, and homelab fit.
Infrastructure
Power, storage, rack, and passthrough. Not a SKU selector.
- Threadripper vs EPYC for a Multi-GPU AI Workstation: Lanes, Memory, and Money At 2-4 GPUs the bottleneck stops being the GPU and becomes PCIe lanes, memory channels, and your electrical panel. This is the lane math, the 15A circuit reality, and the honest Threadripper-vs-EPYC tradeoff: new-platform warranty versus used-server value.
- Tesla V100 Budget AI Homelab: Datacenter Cast-Offs as the Value Multi-GPU Path 2× Tesla V100 32GB with HBM2 bandwidth and NVLink capability at half the cost of new consumer cards — but earned only with cooling mods, blower management, and SXM2 adapter tinkering. For builders, not plug-and-play buyers.
- NAS for AI Model Storage: Sizing a Library of GGUFs, Checkpoints, and Datasets A GGUF collection plus fine-tuning checkpoints hits terabytes in months, not years. This guide sizes NAS capacity, explains why network speed (not caching) is the real constraint, and covers backup strategy for irreplaceable checkpoints.
- Do NPUs Actually Matter for Local AI? The TOPS Marketing vs Reality Gap As of mid-2026, Ollama, llama.cpp, and LM Studio ignore the NPU and route inference through the iGPU or CPU. A 50-TOPS badge on a mini-PC tells you almost nothing about local LLM speed. Learn what to actually measure.
- Best UPS for a Home AI Server: Sizing for GPU Loads, Not Just Runtime GPU rigs need UPS sizing for peak transient draw and pure sine wave inverters, not just runtime hours. This guide sizes by load tier (575W single-GPU, 1000–1200W dual-GPU), explains VA vs watts, and ranks by safety and workload constraint.
- Best Server Racks and Rack Accessories for a Home AI Lab GPU servers need 30+ inches of usable depth, weight ratings above 500 lbs, and often-forgotten accessories like PDUs and cable management. A constraint-driven guide to 12U through 32U racks, sized for the builds LocalRig actually recommends.
- 10GbE vs 2.5GbE for an AI Homelab: When Model Loading Justifies the Upgrade The math on pulling a 40GB 70B-Q4 GGUF from NAS: ~2.6–2.8 minutes at 2.5GbE versus ~42–43 seconds at 10GbE. A constraint-first decision tree: local NVMe first, network second, and why used enterprise SFP+ gear is the honest value path.