Posts

My laptop has a ~/projects folder. Most of it is a graveyard. Not because the ideas were bad — I’d still build some of them if I sat down today. They’re dead because I get excited by a technical problem, work on it for two weekends, hit the part that stops being fun, and drift to the next thing. The codebase stays. The git log doesn’t. I’m 40, a Cloud Architect with ~18 years across IBM and AWS, and I have ADHD. Diagnosed late, lived with it longer. The pattern above isn’t laziness — it’s a specific shape of attention. Hyperfocus until the dopamine of novelty runs out, then gravitational pull toward whatever’s next. Anyone with this wiring recognizes the feeling: the moment a project transitions from “fun problem” to “ten unsexy decisions in a row,” part of your brain leaves the room. ...

What Are We Even Doing Here? It’s 2026. Most people run LLMs on NVIDIA H100s, AMD MI300X, or at least a decent gaming GPU. I’m running them on a 2016 IBM POWER8 server with 160 hardware threads and zero CUDA cores. Why? Because I can. And because nobody else has published POWER8 LLM benchmarks in 2026. And because alternative architectures deserve love too. This post covers: Building llama.cpp on ppc64le with GCC 16 Running Qwen 2.5 7B (text + vision) on POWER8 NUMA tuning discoveries (spoiler: conventional wisdom is wrong) Multimodal inference (yes, vision models work too) Full reproduceability (Gentoo USE flags, build commands, everything) TL;DR: Got 6.81 tokens/s on text generation and fully functional vision inference. POWER8 reads license plates better than some humans. ...

Apple Silicon vs IBM POWER8: A Tale of Two Architectures Running LLMs in 2026 Last week I published benchmarks of running Qwen 2.5 7B on a 2016 IBM POWER8. The results were surprisingly good — 6.81 tokens/s on CPU-only inference with 80 threads hammering away. But then came the inevitable question: How does it compare to modern hardware? So I ran the same benchmarks on my daily driver: a Mac Studio with Apple M2 Max. Same model (Qwen 2.5 7B Q4_K_M), same quantization, different decade. Here’s what I found. ...

Some boys arrive at the ranch and you hear them before you see them. This one arrived and we never saw him at all — not for half an afternoon, anyway. The fault was not his. The fault was a cable that lied. The forasteiro showed up on a Tuesday. ARM64. RK3588. Eight cores, an NPU nobody at the ranch had drivers for, dual 2.5G Realtek NICs, one M.2 slot, one SATA port, sixteen megabytes of SPI NOR flash. He’d been living in Debene Ranch for months already, under a different name — Rock NAS, dressed in Armbian, serving ZFS over NFS like a proper homestead servant. But Dom Felipe had retired him from that job. The kid wanted something else. ...

🎬 ABERTURA Câmera abre num galpão escuro. Luzes piscando. Som de cooler. Ao fundo, uma silhueta masculina com 24 cores se forma na fumaça. É ULTRA2. Plim plim. Entra a música tema: 🎵 “Detalhes tão pequenos de nóóóós… Um swap que ninguém pediu, um kernel que não rodóóóu…” 🎵 CAPÍTULO 1: A CHEGADA DO FORASTEIRO Era uma noite quente em Debene do Sul, pequena fazenda de servidores às margens do Rio MoCA 2.5. Dona Felipe, matriarca da família e dona de todas as terras (e dos /24 inteiros, que ela mesma fez questão de subnetar com as próprias mãos), recebia um forasteiro vindo de longe. ...

🎬 OPENING CREDITS Camera opens on a dark warehouse. Blinking lights. Sound of cooling fans. In the background, a masculine silhouette with 24 cores forms in the mist. It’s ULTRA2. Plim plim. (That’s the iconic Brazilian soap opera sound effect.) Theme song plays: 🎵 “Such tiny details of us two… A swap nobody asked for, a kernel that wouldn’t boot through…” 🎵 CHAPTER 1: THE STRANGER ARRIVES It was a hot night in Debene Ranch, a small server farm by the MoCA 2.5 River. Dona Felipe, matriarch of the family and owner of all the land (and all the /24 subnets, which she personally configured by hand), received a stranger from afar. ...

TL;DR — A new Intel Core Ultra 9 285H (“ultra2”) joined my homelab this week. First Geekbench 6 run scored 1471 single / 12864 multi. That’s slower than the i7-12900K next to it on the rack. Six hours, two Talos upgrades, three diagnostic jobs, and one DaemonSet later, single-core jumped to 3090 (+110%). The villains: an outdated kernel, UKI silently ignoring extraKernelArgs, and a vendor BIOS that thinks PL2=75W is a reasonable default for a chip rated for 115W. Here’s how the diagnosis went. ...

I needed 40 Gbps between two machines: intel9: Ubuntu 24.04, running K8s + misc services xeon2zfs: TrueNAS SCALE 25.10.3, serving storage via NFS The plan: Deploy InfiniBand FDR (40 Gbps), run IPoIB for network services, eventually layer NFS over RDMA if I felt fancy. The hardware: Two Mellanox ConnectX-3 Pro cards from eBay ($20 each), one unmanaged QDR/FDR switch ($35), some cheap QSFP cables. Should be plug-and-play, right? VPI cards, support both Ethernet and InfiniBand, auto-negotiate, just works™. ...

There’s a moment in every homelab evolution where you stop and ask: what am I actually buying when I buy infrastructure? For me that moment came two weeks ago, staring at a renewal email from a SaaS I won’t name, doing back-of-the-envelope math on a Synology I almost ordered, and pricing out an EKS cluster for a side project that — let’s be honest — was never going to leave my apartment. ...

The Question Nobody Asked (But I’m Answering Anyway) I have an IBM POWER8 server from 2014. It has 160 hardware threads. It draws 400W at idle. It sounds like a jet engine warming up. I compiled .NET 8 from source on it, then built Jellyfin 10.11 in nine increasingly desperate attempts, and now it’s streaming movies to my living room. My wife thinks I have a problem. She’s correct. But that’s not the point. ...