--- title: 128GB unified memory is a real trick, but a used RTX... description: AMD's Ryzen AI Max+ 395-based mini PCs (like the Framework Desktop) offer up to 96GB GPU-addressable unified memory, enabling local inference of 100B+... canonical: https://daily.dev/posts/128gb-unified-memory-is-a-real-trick-but-a-used-rtx-3090-still-beats-amd-s-4k-mini-pc-for-most-peo-vteao8w1z twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: 128GB unified memory is a real trick, but a used RTX 3090 still beats AMD's $4K mini PC for most people | daily.dev og:description: AMD's Ryzen AI Max+ 395-based mini PCs (like the Framework Desktop) offer up to 96GB GPU-addressable unified memory, enabling local inference of 100B+... og:url: https://daily.dev/posts/128gb-unified-memory-is-a-real-trick-but-a-used-rtx-3090-still-beats-amd-s-4k-mini-pc-for-most-peo-vteao8w1z og:image: https://api.daily.dev/og/posts/vTEaO8W1z.png og:image:alt: 128GB unified memory is a real trick, but a used RTX 3090 still beats AMD's $4K mini PC for most people og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # 128GB unified memory is a real trick, but a used RTX 3090 still beats AMD's $4K mini PC for most people **[Trends](https://daily.dev/sources/trends)** · 2 min read · 1 upvotes · 0 comments ## Summary AMD's Ryzen AI Max+ 395-based mini PCs (like the Framework Desktop) offer up to 96GB GPU-addressable unified memory, enabling local inference of 100B+ parameter models that no consumer GPU can fit. However, a used RTX 3090 at ~$300 outperforms them on memory bandwidth (~936GB/s vs. the Halo's lower throughput), resulting in faster token generation for models under 24GB. The ROCm software ecosystem still lags behind CUDA for tools like ComfyUI, LM Studio, and Unsloth. At $3,999, the Halo's value proposition is narrow: it's the only affordable option for running 120B+ parameter models locally. For anything under 30B parameters, a used RTX 3090 is the better buy by a wide margin. ## Content AMD's Ryzen AI Halo mini PC (and Framework's similar Desktop) can do something no consumer GPU can: load 100B+ parameter models entirely in local memory. That's the pitch. The reality is more complicated. The hardware is genuinely impressive on paper. The Ryzen AI Max+ 395 chip with 128GB of LPDDR5x-8000 unified memory lets you allocate up to 96GB to the GPU, which means Llama 70B, GPT OSS 120B, and large Mixture-of-Experts models like Qwen 3 235B actually fit. No discrete consumer card can touch that. A used RTX 3090 tops out at 24GB VRAM. But here's where the story gets interesting: for everything *except* raw model size, the $300 used RTX 3090 wins. Its memory bandwidth hits around 936GB/s. The Halo's unified memory architecture, despite the massive capacity, can't match that throughput. The result is noticeably slower token generation on models that *do* fit in 24GB, and painfully slow prefill times on long prompts. Then there's the software problem. CUDA is still the default assumption for almost every local AI tool. ROCm has improved, but it's not there yet. The Linux version of the Halo ships pre-configured with ROCm and PyTorch, which is a real effort from AMD, but "pre-configured" and "works as well as CUDA" are different things. ComfyUI, LM Studio, Unsloth for fine-tuning — they all run, but you're still in second-class-citizen territory compared to Nvidia's ecosystem. The hands-on demos show 30-50+ tokens/second on mid-sized MoE models, which is usable. Multi-model agent setups and video generation with LTX work. Fine-tuning with Unsloth runs. It's not broken. It's just narrow. At $3,999 for the top Framework Desktop config (or similar pricing for the Halo workstation), you're paying a significant premium for one specific capability: fitting models that simply don't fit anywhere else at this price. If you genuinely need to run 120B+ parameter models locally and can't afford a multi-GPU server, this is your only real option. If you're running Llama 3 8B, Mistral, or anything under 30B? Buy the used 3090 and spend the remaining $3,700 on something else. ## Similar posts on daily.dev - [AMD's most exciting AI machine this year isn't a GPU — it's a $3,999 mini PC](https://daily.dev/posts/amd-s-most-exciting-ai-machine-this-year-isn-t-a-gpu-it-s-a-3-999-mini-pc-2w1ukbsmu) · XDA Developers · 0 upvotes · 0 comments - [AMD just dropped a compact AI workstation that makes discrete GPUs look outdated for running LLMs](https://daily.dev/posts/amd-just-dropped-a-compact-ai-workstation-that-makes-discrete-gpus-look-outdated-for-running-llms-gnypyfwfp) · XDA Developers · 0 upvotes · 0 comments - [Hands-On with the AMD Ryzen AI Halo](https://daily.dev/posts/hands-on-with-the-amd-ryzen-ai-halo-g0llzrz16) · Hacker News · 1 upvotes · 0 comments --- Tags: [#nvidia](https://daily.dev/tags/nvidia), [#amd](https://daily.dev/tags/amd), [#local-ai](https://daily.dev/tags/local-ai), [#rocm](https://daily.dev/tags/rocm) [View this post on daily.dev](https://daily.dev/posts/128gb-unified-memory-is-a-real-trick-but-a-used-rtx-3090-still-beats-amd-s-4k-mini-pc-for-most-peo-vteao8w1z) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"128GB unified memory is a real trick, but a used RTX 3090 still beats AMD's $4K mini PC for most people","url":"https://daily.dev/posts/128gb-unified-memory-is-a-real-trick-but-a-used-rtx-3090-still-beats-amd-s-4k-mini-pc-for-most-peo-vteao8w1z","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/128gb-unified-memory-is-a-real-trick-but-a-used-rtx-3090-still-beats-amd-s-4k-mini-pc-for-most-peo-vteao8w1z"},"datePublished":"2026-07-23T12:14:22.429Z","dateModified":"2026-07-29T12:18:00.928Z","description":"AMD's Ryzen AI Max+ 395-based mini PCs (like the Framework Desktop) offer up to 96GB GPU-addressable unified memory, enabling local inference of 100B+...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ebd22a4149662a2e08a91232a63d4263?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ebd22a4149662a2e08a91232a63d4263?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/128gb-unified-memory-is-a-real-trick-but-a-used-rtx-3090-still-beats-amd-s-4k-mini-pc-for-most-peo-vteao8w1z","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"nvidia,amd,local-ai,rocm","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"128GB unified memory is a real trick, but a used RTX 3090 still beats AMD's $4K mini PC for most people"}]} ```