{"id":17359,"date":"2025-08-23T01:29:16","date_gmt":"2025-08-23T01:29:16","guid":{"rendered":"https:\/\/www.europesays.com\/ie\/17359\/"},"modified":"2025-08-23T01:29:16","modified_gmt":"2025-08-23T01:29:16","slug":"building-a16zs-personal-ai-workstation-with-four-nvidia-rtx-6000-pro-blackwell-max-q-gpus","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/ie\/17359\/","title":{"rendered":"Building a16z\u2019s Personal AI Workstation with four NVIDIA RTX 6000 Pro Blackwell Max-Q GPUs"},"content":{"rendered":"<p><a href=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-12.png\"><img fetchpriority=\"high\" decoding=\"async\" class=\"alignleft wp-image-378422 size-full\" src=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-12.png\" alt=\"a16z Personal AI Workstation with 4x NVIDIA RTX 6000 Pro Blackwell Max-Q GPUs\" width=\"2000\" height=\"1608\"  \/><\/a><\/p>\n<p>In the era of foundation models, multimodal AI, LLMs, and ever-larger datasets, access to raw compute is still one of the biggest bottlenecks for researchers, founders, developers, and engineers. While the cloud offers scalability, building a <strong>personal AI Workstation<\/strong> delivers complete control over your environment, latency reduction, custom configurations and setups, and the privacy of running all workloads locally.<\/p>\n<p>This post covers our version of a<strong> four-GPU workstation<\/strong> powered by the new NVIDIA <strong>RTX 6000 Pro Blackwell Max-Q GPUs<\/strong>. This build pushes the limits of desktop AI computing with <strong>384GB<\/strong> of VRAM (<strong>96GB<\/strong> each GPU), all in a shell that can fit under your desk.<\/p>\n<p><a href=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-13.png\"><img loading=\"lazy\" decoding=\"async\" class=\"alignleft wp-image-378421 size-full\" src=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-13.png\" alt=\"a16z Personal AI Workstation with 4x NVIDIA RTX 6000 Pro Blackwell Max-Q GPUs\" width=\"2000\" height=\"1175\"  \/><\/a><\/p>\n<p>Why Build This Workstation?<\/p>\n<p>Training, fine-tuning, and running inference on modern AI models require massive VRAM bandwidth, high CPU throughput, and ultra-fast storage. Running these workloads in the cloud can introduce latency, setup overhead, slower data transfer speeds, and privacy tradeoffs.<\/p>\n<p>By building a workstation around <strong>enterprise-grade GPUs with full PCIe 5.0 x16 connectivity<\/strong>, we get:<\/p>\n<ul>\n<li><strong>Maximum GPU-to-CPU bandwidth:<\/strong>\u00a0No bottlenecks from PCIe switches or shared lanes.<\/li>\n<li><strong>Enterprise-class VRAM:<\/strong> Each RTX 6000 Pro Blackwell Max-Q provides 96GB of VRAM, enabling dense training runs and large model inference without quantization. Each card consumes only 300W of power at peak (Max-Q version).<\/li>\n<li><strong>8TB of NVMe 5.0 storage:<\/strong> 4x 2TB of NVMe PCIe 5.0 x4 modules.<\/li>\n<li><strong>256 GB<\/strong> of total ECC DDR5 RAM.<\/li>\n<li><strong>Surprising efficiency:<\/strong> Despite its scale, the workstation pulls <strong>1650W at peak<\/strong>, low enough to run on a standard 15-amp \/ 120V household circuit.<\/li>\n<li><strong>Next-gen data GDS streaming:<\/strong> While we are still in the process of testing this support, this setup should be compatible with the <strong>NVIDIA GPUDirect Storage (GDS)<\/strong>, which allows datasets or models to stream directly from PCIe 5.0 NVMe SSDs into GPU VRAM, bypassing CPU memory, to reduce latency and maximize throughput.<\/li>\n<\/ul>\n<p><strong>We are planning to test and make a limited number of these custom a16z Founders Edition AI Workstations.<\/strong><\/p>\n<p><a href=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-10.png\"><img loading=\"lazy\" decoding=\"async\" class=\"alignleft wp-image-378424 size-full\" src=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-10.png\" alt=\"Detail of a16z Personal AI Workstation with 4x NVIDIA RTX 6000 Pro Blackwell Max-Q GPUs\" width=\"2000\" height=\"1500\"  \/><\/a><\/p>\n<p>Core Specifications<\/p>\n<p>Let\u2019s break down the hardware:<\/p>\n<ul>\n<li><strong>GPUs:<\/strong>\n<ul>\n<li>4 \u00d7 NVIDIA RTX 6000 Pro Blackwell Max-Q<\/li>\n<li>96GB VRAM per GPU (<strong>384GB total VRAM<\/strong>)<\/li>\n<li>Each card on a dedicated PCIe 5.0 x16 lane<\/li>\n<li>300W per GPU<\/li>\n<\/ul>\n<\/li>\n<li><strong>CPU:<\/strong>\n<ul>\n<li>AMD Ryzen Threadripper PRO 7975WX (liquid cooled with Silverstone XE360-TR5)<\/li>\n<li>32 cores \/ 64 threads<\/li>\n<li>Base clock: 4.0 GHz, Boost up to 5.3 GHz<\/li>\n<li>8-channel DDR5 memory controller<\/li>\n<\/ul>\n<\/li>\n<li><strong>Memory:<\/strong>\n<ul>\n<li><strong>256GB<\/strong> ECC DDR5 RAM<\/li>\n<li>Running across 8 channels (32GB each)<\/li>\n<li>Expandable up to 2TB<\/li>\n<\/ul>\n<\/li>\n<li><strong>Storage:<\/strong>\n<ul>\n<li><strong>8TB<\/strong> total: 4x 2TB PCIe 5.0 NVMe SSDs x4 lanes each (up to 14,900 MB\/s \u2013 theoretical read speed for each NVMe module)<\/li>\n<li>Configurable in RAID 0 for <strong>~59.6GB\/s aggregate theoretical read throughput<\/strong> (we are in the process of testing real throughput).<\/li>\n<\/ul>\n<\/li>\n<li><strong>Power Supply:<\/strong>\n<ul>\n<li>Thermaltake Toughpower GF3 1650W 80 PLUS Gold<\/li>\n<li>System-wide max draw: 1650W, operable on a <strong>standard, dedicated 15A 120V outlet<\/strong><\/li>\n<\/ul>\n<\/li>\n<li><strong>Motherboard:<\/strong>\n<ul>\n<li>GIGABYTE MH53-G40 (AMD WRX90 Chipset)<\/li>\n<\/ul>\n<\/li>\n<li><strong>Case:<\/strong>\n<ul>\n<li>Off the shelf Extended ATX case with some custom modifications.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p><a href=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-11.png\"><img loading=\"lazy\" decoding=\"async\" class=\"alignleft wp-image-378423 size-full\" src=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-11.png\" alt=\"a16z Personal AI Workstation with 4x NVIDIA RTX 6000 Pro Blackwell Max-Q GPUs\" width=\"2000\" height=\"1500\"  \/><\/a><\/p>\n<p>Design Highlights<br \/>\nFull PCIe 5.0 Bandwidth<\/p>\n<p>Each GPU is connected via its own <strong>dedicated PCIe 5.0 x16<\/strong>, ensuring maximum data transfer rates between CPU and GPU. Unlike multi-GPU setups that rely on bifurcated lanes, multiplexers, or external bridges, this build guarantees <strong>no compromise on lane allocation or defaulting in lower PCIe versions<\/strong>.<\/p>\n<p>Storage for High-Speed Datasets<\/p>\n<p>The four PCIe 5.0 NVMe SSDs provide read speeds of up to <strong>~14.9 GB\/s each (theoretical)<\/strong>, scaling to <strong>~59 GB\/s theoretical in RAID 0<\/strong>. While we are still in the process of testing full <strong>NVIDIA GPUDirect Storage (GDS)<\/strong> compatibility, it could allow GPUs to fetch data directly from NVMe drives, enabling direct-memory access (DMA).<\/p>\n<p>Power Efficiency &amp; Practicality<\/p>\n<p>The overall system consumes <strong>1650W peak<\/strong> and fits comfortably into a home or office environment without requiring dedicated circuits or 220V wiring. With built-in wheels, it is designed for effortless transport between locations.<\/p>\n<p>Baseboard Management Controller (BMC)<\/p>\n<p>Integrated AST2600, a Baseboard Management Controller (BMC) that serves as a dedicated processor for remote out-of-band server management, operating independently of the host CPU and OS to handle critical monitoring and control tasks.<\/p>\n<p><a href=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-9.png\"><img loading=\"lazy\" decoding=\"async\" class=\"alignleft wp-image-378418 size-full\" src=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-9.png\" alt=\"CPU connection diagram of a16z Personal AI Workstation with 4x NVIDIA RTX 6000 Pro Blackwell Max-Q GPUs\" width=\"2000\" height=\"1713\"  \/><\/a><\/p>\n<p>Use Cases<\/p>\n<ul>\n<li><strong>Training and fine-tuning LLMs<\/strong> with up to <strong>tens of billions of parameters<\/strong> in full precision.<\/li>\n<li><strong>Running dense multimodal inference<\/strong> across image, text, audio, and video models simultaneously.<\/li>\n<li><strong>Experimenting with model parallelism<\/strong> (tensor, pipeline, or expert-based sharding) across four GPUs.<\/li>\n<li><strong>Streaming high-throughput datasets<\/strong> directly from SSD RAID into GPU memory for reinforcement learning or diffusion-based workloads.<\/li>\n<\/ul>\n<p>With libraries like vLLM, DeepSpeed, SGLang, etc., this machine serves as a foundation for training and serving custom LLMs, RL training pipelines, multimodal models, autonomous agents, etc., without cloud dependency and with a custom setup and environment.<\/p>\n<p>This RTX 6000 Pro Blackwell workstation represents a <strong>sweet spot between datacenter power and desktop accessibility<\/strong>; all while staying within the footprint and power draw of a high-end AI Workstation under your desk.<\/p>\n<p>Whether you\u2019re a researcher exploring new architectures, a startup prototyping private LLM deployments, or simply an enthusiast, this build demonstrates an efficient, AI Workstation under your desk.<\/p>\n<p>Some temperature tests:<\/p>\n<p><a href=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-14.png\"><img loading=\"lazy\" decoding=\"async\" class=\"alignleft wp-image-378420 size-full\" src=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-14.png\" alt=\"Full Utilization statistics of a16z Personal AI Workstation with 4x NVIDIA RTX 6000 Pro Blackwell Max-Q GPUs\" width=\"2000\" height=\"1038\"  \/><\/a><\/p>\n<p><a href=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-15.png\"><img loading=\"lazy\" decoding=\"async\" class=\"alignleft wp-image-378419 size-full\" src=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/08\/250821-Building-a-Next-Gen-AI-Workstation-15.png\" alt=\"Idle statistics for a16z Personal AI Workstation with 4x NVIDIA RTX 6000 Pro Blackwell Max-Q GPUs\" width=\"2000\" height=\"1038\"  \/><\/a><\/p>\n<p>\u00a0<\/p>\n<p>\u00a0<\/p>\n","protected":false},"excerpt":{"rendered":"In the era of foundation models, multimodal AI, LLMs, and ever-larger datasets, access to raw compute is still&hellip;\n","protected":false},"author":2,"featured_media":17360,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[261],"tags":[291,289,290,18,19,15548,17,82],"class_list":{"0":"post-17359","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-artificial-intelligence","8":"tag-ai","9":"tag-artificial-intelligence","10":"tag-artificialintelligence","11":"tag-eire","12":"tag-ie","13":"tag-infra","14":"tag-ireland","15":"tag-technology"},"share_on_mastodon":{"url":"","error":""},"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/posts\/17359","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/comments?post=17359"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/posts\/17359\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/media\/17360"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/media?parent=17359"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/categories?post=17359"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/tags?post=17359"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}