<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
        <title>I Turned a £200 Trash Can Mac Pro Into a Linux AI Server in 2026</title>
        <link>https://tube.devwithzachary.com/videos/watch/a5dd0bb9-c68b-4c95-9bed-4136ba45af7d</link>
        <description>Can a £200 2013 "Trash Can" Apple Mac Pro (MacPro6,1) running Linux actually beat modern hardware for local AI, vision alt-text, and daily coding in 2026? Out of the box, modern local AI tools hit three brick walls on this machine: AMD dropped ROCm support for GCN 1.0 (FirePro D500) years ago, Ubuntu defaults to the legacy radeon kernel driver with zero Vulkan support, and the 2013 Ivy Bridge Xeon lacks AVX2 and FMA instructions so stock binaries crash immediately with SIGILL. In this video, we unlock all three layers from scratch: forcing amdgpu.si_support=1 for Mesa RADV Vulkan 1.3 across both 3GB FirePro D500 GPUs (6GB GDDR5 total @ 240 GB/s per card), compiling llama.cpp for Ivy Bridge with Link-Time Optimization, building a custom dual-GPU telemetry tool (trash-top), and enabling Linux Transparent Huge Pages (+60% CPU speedup!). Then we put the £200 Trash Can head-to-head on exact commit 0f35d6d7b against three modern contenders: a Star Labs StarFighter Ultra 5, an 8GB MacBook Neo, and a £2,000 M4 Max Mac Studio. Want to support the channel and my open-source projects? Patreon: https://www.patreon.com/DevWithZachary, Buy Me a Coffee: https://www.buymeacoffee.com/devwithzachary, Code, Setup Scripts &amp; Full Benchmarks (trash-top + trash-ai): Project TrashCanAI Repo: https://github.com/devwithzachary/TrashCanAI, Chart &amp; Infographic Generator Tool: https://github.com/devwithzachary/graphic-generator, Website &amp; All Project Links: https://devwithzachary.com, The 4 Test Machines &amp; Models Tested (llama.cpp commit 0f35d6d7b): 2013 Mac Pro ("Trash Can", ~£200 / £210 CEX D500): 12C Xeon E5 (-t 12 optimal), 64GB DDR3 ECC, Dual FirePro D500 (6GB GDDR5), Ubuntu Server (Vulkan), Star Labs StarFighter Ultra 5 (~£1,500-£2,500): Core Ultra 5 125H (28W PL1), 32GB LPDDR5/X, Intel Arc iGPU (21.4 GiB Vulkan UMA), CachyOS Linux, MacBook Neo (~£600-$1,000): Apple A18 Pro (6C CPU / 5C GPU), 8GB Unified Memory (5,726 MB Metal cap), macOS 27.2 (Metal), M4 Max Mac Studio (~£2,000): Apple M4 Max (14C CPU / 32C GPU), 36GB Unified Memory @ 410 GB/s, macOS 27.0.1 (Metal), Models: Qwen 2.5 Coder 1.5B (Q8_0), Qwen 2.5 3B (Q4_K_M), Qwen 2.5 Coder 7B (Q4_K_M), Qwen 2.5 Coder 14B (Q4_K_M), Qwen 2.5 Coder 32B (Q4_K_M), and Moondream2 Vision (F16)., Key 2013 Mac Pro (Dual D500 Vulkan + THP) Benchmark Highlights: Qwen 2.5 Coder 1.5B (Q8): 392.5 tok/s prompt | 36.6 tok/s generation (beats StarFighter 21.2 &amp; MacBook Neo 31.9), Qwen 2.5 3B (Q4_K_M): 178.3 tok/s prompt | 22.1 tok/s generation, Qwen 2.5 Coder 7B (Q4_K_M): 74.4 tok/s prompt | 13.0 tok/s generation (beats StarFighter 11.7 &amp; MacBook Neo 11.74), Qwen 2.5 Coder 14B (Q4_K_M): 27.6 tok/s prompt | 6.40 tok/s generation (Hybrid 24/48 GPU+CPU; MacBook Neo OOM), Qwen 2.5 Coder 32B (Q4_K_M): 9.06 tok/s prompt | 2.53 tok/s generation (64GB ECC RAM; MacBook Neo OOM), Moondream2 Vision: 1.75s ViT encode | ~7.8s total IDE alt-text, Hardware &amp; Upgrade Links (Affiliate Links - Supports the Channel at No Extra Cost): Star Labs Linux Laptops (5% off with code ZACHARY5): https://starlabs.systems/zachary5, Smart Plug Power Monitor (Stop 120W 24/7 idle!): https://www.amazon.co.uk/s?k=Smart+Plug+Energy+Monitor+UK&amp;tag=devwithzach0f-21, Sintech M.2 NVMe SSD Adapter (2013 Mac Pro): https://www.amazon.co.uk/s?k=Sintech+M.2+NVMe+Adapter+Mac+Pro+2013&amp;tag=devwithzach0f-21, 1TB / 2TB M.2 NVMe SSD: https://www.amazon.co.uk/s?k=1TB+M.2+NVMe+SSD&amp;tag=devwithzach0f-21, Renewed Apple Mac Pro (Late 2013): https://www.amazon.co.uk/s?k=Apple+Mac+Pro+2013&amp;tag=devwithzach0f-21, My Open-Source Projects Featured in This Video: LinuxOnAndroid: https://github.com/devwithzachary/LinuxOnAndroid, LinuxOnVita: https://linuxonvita.com, MineServe (Android Minecraft Server): https://github.com/devwithzachary/MineServe, Follow Me Online: Discord (Channel + All Projects): https://discord.gg/QdT9yCrKmV, Website: https://devwithzachary.com, LinkedIn: https://www.linkedin.com/in/zachary-mg-powell/, Bluesky: https://bsky.app/profile/devwithzachary.bsky.social, PeerTube (Ad-Free Mirror): https://tube.devwithzachary.com, Mastodon: https://social.devwithzachary.com/@zachary, Timestamps: 00:00 Intro 00:32 My Love for the Trashcan (And What We Tried Before) 05:25 What We Are Comparing It Against (Star Labs StarFighter, MacBook Neo &amp; Mac Studio) 06:37 Why Running LLMs on the Trashcan Didn't Work in the Past 07:32 Getting the GPUs to Work With llama.cpp 10:08 Tuning the System for LLM Optimisation 14:00 Models We Used for Testing 17:10 Performance Numbers After Optimisation 20:05 Trashcan vs Star Labs StarFighter (Intel Ultra 5 125H) 22:10 Trashcan vs MacBook Neo 23:32 Trashcan vs M4 Max Mac Studio 25:18 Mac Studio Is Magnitudes Better... But £2,000 More 27:32 Live Real-Time Examples of LLMs Running on the Trashcan 30:00 What About the Mac Pro With D700 GPUs? 30:51 So Is It Worth Getting a Trashcan for LLMs?MacPro</description>
        <lastBuildDate>Tue, 17 Nov 2026 09:59:38 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>PeerTube - https://tube.devwithzachary.com</generator>
        <image>
            <title>I Turned a £200 Trash Can Mac Pro Into a Linux AI Server in 2026</title>
            <url>https://tube.devwithzachary.com/lazy-static/avatars/98d38527-3182-4149-b430-6872710ae157.jpg</url>
            <link>https://tube.devwithzachary.com/videos/watch/a5dd0bb9-c68b-4c95-9bed-4136ba45af7d</link>
        </image>
        <copyright>All rights reserved, unless otherwise specified in the terms specified at https://tube.devwithzachary.com/about and potential licenses granted by each content's rightholder.</copyright>
        <atom:link href="https://tube.devwithzachary.com/feeds/video-comments.xml?videoId=a5dd0bb9-c68b-4c95-9bed-4136ba45af7d" rel="self" type="application/rss+xml"/>
        <item>
            <title><![CDATA[I Turned a £200 Trash Can Mac Pro Into a Linux AI Server in 2026 - artyom]]></title>
            <link>https://tube.devwithzachary.com/w/mtVRJmHpBTz2Gg2X5ky5L6;threadId=6666</link>
            <guid>https://tube.devwithzachary.com/w/mtVRJmHpBTz2Gg2X5ky5L6;threadId=6666</guid>
            <pubDate>Mon, 16 Nov 2026 18:30:30 GMT</pubDate>
            <content:encoded><![CDATA[<p>🚫AI</p>
]]></content:encoded>
            <dc:creator>artyom</dc:creator>
        </item>
    </channel>
</rss>