AMD just put serious AI horsepower in a box you can hold in one hand. Here is what the Ryzen AI Max+ 395 actually delivers — and where the hype meets reality.

The Viral Moment Everyone Is Talking About
On stage, AMD CEO Lisa Su held up a small mini PC and ran a 235-billion-parameter model live — no data center, no cloud bill, no rented GPU cluster. The clip spread fast across tech feeds because it pushed back on a long-held assumption: that serious AI inference requires either expensive cloud subscriptions or power-hungry discrete GPUs.
The hardware behind that demo is the AMD Ryzen AI Max+ 395, codenamed Strix Halo, sitting inside compact systems like the GMKtec EVO-X2. It is the first x86 chip where the CPU and a capable integrated GPU share one large unified memory pool — up to 128 GB of fast LPDDR5X. On Linux, roughly 110 GB of that can be handed to the GPU. For comparison, that is more than three times the memory of an RTX 5090 (32 GB) and over four times an RTX 4090 (24 GB). For large language models that stall when they do not fit in discrete VRAM, that is a meaningful shift.
Why Unified Memory Changes the Game
Traditional discrete GPUs hit a bottleneck: model weights and activations constantly shuttle across the PCIe bus between system RAM and limited VRAM. Unified memory removes most of that traffic. The Radeon 8060S integrated graphics (40 RDNA 3.5 compute units) and 16 Zen 5 CPU cores both reach the same high-speed memory directly. The result is that models which simply would not load on a 24 GB or 32 GB card can run comfortably on this little box.
CPU
16 Zen 5 cores, 32 threads, up to 5.1 GHz boost clock.
Integrated GPU
Radeon 8060S with 40 RDNA 3.5 compute units sharing the unified pool.
NPU
50+ TOPS via XDNA 2 — purpose-built for AI acceleration workloads.
Memory
Up to 128 GB LPDDR5X unified; roughly 96–110 GB usable as VRAM on Linux.
Power Draw
55–120 W configurable TDP; whole system typically around 120 W under full AI load.
Form Factor
Compact mini PC — the GMKtec EVO-X2 fits in one hand and sits on a desk.
Performance: The 3× Claims, in Context
AMD and early users report the system beating an RTX 5080 by more than three times on certain DeepSeek R1 inference workloads. That advantage shows up most clearly when the model exceeds the discrete card’s VRAM — which is exactly where unified memory earns its keep. It is fair, but it is not the whole picture.
Where It Shines
Retrieval-augmented generation over private documents, batch inference, chat-style workloads, fine-tuning smaller models, and fully offline private use.
Good With Trade-offs
Long-context agentic coding, where prefill and tokenization can lag high-end Nvidia setups due to memory bandwidth and the maturity of the ROCm software stack.
Not a Replacement For
High-concurrency, data-center-class inference or cutting-edge research that depends on the newest CUDA-only kernels.
Power Efficiency
About 120 W for the whole machine versus several hundred watts for a discrete GPU and CPU rig doing comparable memory-heavy work.
ROCm, AMD’s software stack, has improved dramatically but still trails CUDA by roughly 12 to 18 months in ecosystem breadth. Ollama, LM Studio, vLLM, and llama.cpp all work well out of the box for most users.
The Economics: When Does This Pay for Itself?
This is the part our clients ask about first. Many power users are currently stacking subscriptions — a $200-per-month premium coding plan, a $200-per-month pro chat plan, plus $20 to $50 for assorted tools. That adds up to $5,000 or more per year.
per year in stacked AI subscriptions for heavy power users
A $1,500 to $2,000 EVO-X2-class machine with 128 GB, plus roughly $9 to $12 a month in electricity, can replace a large portion of that spend once it pays for itself. After the payback period you get unlimited local inference, no rate limits at 3 a.m., and complete data privacy.
When the data never leaves your office, an entire category of compliance and confidentiality worry goes away — which is why lawyers, medical practices, government contractors, and developers with sensitive IP are paying close attention.
Local AI Hardware vs. Cloud: A Quick Comparison
| Option | Usable Memory | Approx. Power Draw | Best Use Case |
|---|---|---|---|
| Ryzen AI Max+ 395 (EVO-X2) | Up to ~110 GB unified | ~120 W whole system | Private, offline, large-model local inference |
| RTX 4090 | 24 GB VRAM | 450 W GPU alone | Fast inference on models that fit in VRAM |
| RTX 5090 | 32 GB VRAM | 575 W GPU alone | Top single-card speed, VRAM-limited |
| Cloud subscriptions | Effectively unlimited | N/A (hosted) | Peak frontier capability, no maintenance |
The honest takeaway is not “Nvidia is finished” or “all subscriptions are over.” It is that high-memory unified hardware just became practical and affordable for private, local AI — which shifts the math for a real slice of workloads and points toward a hybrid future: local for privacy and cost, cloud for peak capability.
Who Should Buy One Right Now?
- Developers and indie builders tired of token meters and context limits
- Teams that need private retrieval over internal documents
- Privacy-conscious professionals in legal, healthcare, and finance
- Anyone wanting to fine-tune or run large open-source models locally
- Teams needing maximum speed on frontier-quality agentic workflows
- Users who want zero maintenance and constant model updates
- High-concurrency, public-facing applications where cloud still wins
Getting Started With Local AI on This Hardware

- Use Linux for the best memory allocation, giving the GPU 110 GB or more of usable VRAM.
- Install Ollama or LM Studio as your local inference runtime.
- Pull strong open models such as Qwen3 235B, DeepSeek variants, or Llama 3.3 70B.
- Point your tools or agents at localhost — no API key, no cloud dependency.
- Set the GPU memory split appropriately in the BIOS or ROCm settings for maximum usable VRAM.
Many users are already running 70B-class models comfortably and watching the subscription math tip in their favor.
The Honest Take
This is not magic. The 235B model in the demo is a Mixture-of-Experts design, which activates far fewer parameters per token than a dense 235B model would. Prefill speeds can trail an optimized Nvidia setup on long contexts, and the software ecosystem, while strong and improving fast, is not at CUDA parity yet. But the direction of travel is clear, and the early movers who fold local high-memory inference into their stack will hold a real edge in speed, cost, and data control over the next 18 to 24 months.
How WD Morgan Thinks About This
We are a veteran-owned digital marketing agency, and we watch this kind of hardware shift closely because it changes how our clients can work. If you run a business where privacy, cost control, or offline capability matters — a law firm, a medical practice, an MSP, or any company handling sensitive client data — a machine like the EVO-X2 is worth testing now. Our growth is the mission, and that includes helping the businesses we serve understand where technology genuinely moves the needle and where it is just noise. If you want a straight answer about whether local AI fits your operation, we are happy to talk it through.
Frequently Asked Questions
How much does the GMKtec EVO-X2 cost?
Configurations with 128 GB of unified memory generally land in the $1,500 to $2,000 range, depending on storage and the seller.
Is ROCm good enough for running local LLMs?
For most production inference, yes. Ollama, LM Studio, vLLM, and llama.cpp run well out of the box. ROCm still trails CUDA by roughly 12 to 18 months for bleeding-edge frameworks, so the newest CUDA-only research kernels may not be available.
Will this replace my cloud AI subscription?
For a meaningful share of workloads — private document retrieval, chat, batch inference, and offline use — it can. For frontier-quality agentic coding and high-concurrency public applications, a hybrid approach that keeps cloud access for peak capability usually makes more sense.
Is my data private when I run models locally?
Yes. When inference runs on hardware in your own office, the data never leaves your network, which is a major reason regulated industries are interested.
Local hardware is one end of the AI infrastructure spectrum; the other end is genuinely exotic, covered in our piece on space data centers and the SpaceX-Nvidia orbital AI bet. And if you’re deciding what size of model actually belongs on hardware like this, see our guide to small language models and domain-specific LLMs.
Your growth is the mission.
WD Morgan Solutions is a veteran-owned digital marketing agency helping Florida businesses navigate technology decisions that actually move the needle. Whether you need guidance on SEO, web design, or figuring out where tools like local AI fit your operation, we cut through the noise. Get in touch and let’s talk through what makes sense for your business.

























