The Reclaimed Node Blog
Practical guides to homelabs and local AI: picking hardware and software, putting reclaimed gear to work, and benchmarking LLMs on real machines.
PCIe Gen2 for the CMP 170HX in a Proxmox VM
A field guide to getting an unlocked 64 GB CMP 170HX to link at Gen2 x16 when it is passed through to a VM: a small driver patch in the VM and a link retrain on the host.
Oct 30, 2026 · 14 min read Local AISame Model, Three Engines: vLLM Puts the CMP 170HX Ahead
The same 8-bit model on one CMP 170HX with vLLM and an M5 Ultra Mac Studio with MLX and llama.cpp. The CMP finished first at every prompt size and kept scaling to eight users.
Oct 27, 2026 · 10 min read Local AIllama.cpp vs MLX on an M5 Ultra Mac Studio
Same model, same quantization, different engine. MLX reads long prompts 55 to 63% faster, generation speed is close to a tie, and two simultaneous users split the two engines.
Oct 25, 2026 · 6 min read Local AISame Model, Two Machines: Do the Answers Match?
The same 27B model on a Mac Studio and a CMP 170HX server, checked on four published tests. Scores matched within two points. The reply limit moved them by up to 14.
Oct 24, 2026 · 8 min read Local AIRace Your Local AI Machines and Trust the Numbers
LLM-Super-Bench, our free open-source tool, sends the same prompt to two local LLM servers at the same instant, times them fairly, and turns the results into charts you can share.
Oct 23, 2026 · 6 min read Local AIMac Studio vs. CMP 170HX: Which Serves a 27B LLM Better?
Same Qwen 27B model, same llama.cpp build and flags. The Mac Studio won 28 of 31 metrics and pulls ahead as more users and agents share it. Plus cost, power, and where the CMP still fits.
Oct 21, 2026 · 12 min read HomelabBuilding a 4-GPU EPYC Workstation on DDR4
Why we built on EPYC 7402P and the ROMED8-2T: the newest EPYC platform on DDR4, four full x16 slots, and an open case built for airflow.
Oct 6, 2026 · 5 min read Local AIFour CMP 170HX Cards, One Server: 256GB of VRAM
Four mining cards unlocked to 64GB each, running GLM-5.3-Flash on vLLM at up to 563 tokens per second. Build, first benchmarks and thermals.
Oct 6, 2026 · 3 min read Local AIMac Studio vs. a 64GB NVIDIA CMP 170HX: Local AI Image Generation Head-to-Head
Same model, same prompts, same seeds. An Apple M5 Ultra Mac Studio takes on a mining-only NVIDIA card unlocked to 64GB, and wins every image.
Oct 5, 2026 · 6 min read Local AILocal AI Image Generation on a 64GB NVIDIA CMP 170HX
Qwen Image 2512 at full bf16 precision, entirely in VRAM. No quantization, no offloading.
Sep 26, 2026 · 6 min read HomelabThe Ultimate Homelab BMC Showdown: Dell iDRAC vs. HPE iLO vs. Lenovo XCC
Comparing out-of-band management platforms for enterprise servers—licensing, interfaces, firmware updates, and fan noise.
Sep 15, 2026 · 10 min read HomelabUnlocking a “Mining-Only” GPU on Enterprise Server Hardware
Two CMP 170HX cards, a Dell firmware wall that no setting could move, and the fix that sidestepped it entirely.
Sep 10, 2026 · 7 min read HomelabWhat Exactly Is a Homelab? (And What Can You Use It For?)
Your personal, dedicated space to play, break, and build things with computer hardware and software.
Sep 06, 2026 · 5 min read HomelabHomelab Hardware Breakdown: How to Choose Your Gear
Balancing power bills, room noise, and project goals across old PCs, mini PCs, enterprise servers, and Pis.
Sep 06, 2026 · 6 min read HomelabHomelab Software Breakdown: Operating Systems, Virtualization, and Containers
Proxmox, TrueNAS, Ubuntu Server, and Docker — picking the software foundation that matches your goal.
Sep 06, 2026 · 6 min read