Exploring AMD’s New Hardware: A Deep Dive into the Minisforum 395 Platform
A few weeks ago, I embarked on an extensive exploration into the evolving landscape of local AI deployment and shared my insights on why local AI is finally becoming practical and accessible. If you’re interested in the broader context of my research, feel free to read the full article here: Running an Opus-Level Coding Agent. In that piece, AMD was mentioned briefly as an emerging player, but I promised to investigate its offerings more thoroughly—so, here we are.
Recently, I took the plunge and purchased a preconfigured AI-focused hardware box, anticipating that it might deliver the best balance of price and performance currently available. This device is the Minisforum 128GB model, equipped with AMD’s latest 395 platform. While the newer 495 platform is not yet widely accessible, this setup offers compelling value. In the UK, it costs approximately £3,500, and after accounting for regional premiums, it can be acquired in the US for around $3,500 or even less.
Why AMD’s Platform Looks Promising
One of the most attractive aspects of the AMD platform is its affordability. Despite the lower price point, it provides impressive specifications, particularly in memory bandwidth—clocking in at around 256GB/s. This makes it suitable for intensive AI workloads, where memory throughput is critical.
Technical Setup and Performance
To maximize resources, I replaced the default Windows OS with a clean installation of Ubuntu Server. Notably, I found myself leveraging Vulkan for some of the performance-critical tasks—a surprising development given the current state of AMD’s ROCm ecosystem, which remains somewhat underdeveloped.
Currently, I’m experimenting with the Qwen 3.8 flash model, with a multi-tasking process (MTP) depth of approximately 4. This setup allows for a context window of at least 128K tokens, enabling substantial input processing capabilities. As a quick test, I tasked the system with creating a simple meteor-themed game in a single shot, achieving a sustained processing rate of 40 to 50 tokens per second. This level of performance is very encouraging and appears quite practical for real-world applications.
Ongoing Development and Future Plans
I’m engaged in a more in-depth process of building and iterating entire AI applications multiple times, refining hardware configurations and software setups along the way. I plan to share more updates and insights through my newsletter at https://ailocal.substack.com.
For those interested in the collective effort to optimize local AI deployments, I maintain a centralized resource compile at https://github.com/boxabirds/awesome-local-ai, where I gather the best configurations, tools, and workflows tailored to various hardware setups.
Final Thoughts
The AMD 395 platform, particularly in the form of compact, cost-effective boxes like the Minisforum, holds significant promise for developers and AI enthusiasts seeking accessible yet powerful local AI solutions. While certain software ecosystems still have room for improvement, the current performance and affordability make this an exciting time to explore AMD-based AI hardware.
Stay tuned for more updates as I continue testing and optimizing this platform for a range of AI projects.
