I've seen plenty of mini PCs called “workstations” because they have a fast processor and a premium aluminum case.
This one is different.
FEVM has packed up to 192GB of unified memory into a chassis measuring roughly one liter.
That is a ridiculous amount of memory for something this small, but there is a reason for it: local AI.
The FAEX2 uses AMD's Ryzen AI Max+ PRO 495, whose unified-memory architecture allows a large portion of the system memory to be allocated to the integrated Radeon 8065S graphics.
For AI developers, that changes the question from “How fast is this mini PC?” to something more useful:
How large a model can you actually keep in memory?
The FEVM FAEX2 combines AMD's Ryzen AI Max+ PRO 495 with up to 192GB of unified memory in a compact 1-liter chassis.
What Is the FEVM FAEX2?
The FEVM FAEX2 is an ultra-compact workstation built around AMD's latest Ryzen AI Max+ PRO 495 processor.
The chip combines 16 Zen 5 CPU cores, 32 threads, Radeon 8065S integrated graphics with 40 graphics cores and an XDNA 2 NPU rated for up to 55 TOPS.
AMD lists the processor's maximum system-memory support at 192GB of LPDDR5X.
FEVM uses that capability to create a machine aimed squarely at users who want unusually large local workloads in a very small physical footprint.
Why 192GB Matters More for AI Than for Normal PCs
For a normal desktop, going from 64GB to 192GB can sound excessive.
For local AI, it can be transformative.
A large language model needs memory for its weights, runtime buffers, context and other working data.
Once a model cannot fit comfortably inside the accelerator's available memory, the system may need to move information between CPU memory and the GPU or use other forms of offloading.
Up to 160GB Can Be Assigned as VRAM
This is the specification that makes the FAEX2 particularly unusual.
FEVM says the 192GB configuration can allocate up to 160GB as VRAM.
That does not turn the Radeon 8065S into a discrete 160GB graphics card.
It means the GPU can access a very large portion of the unified memory pool.
That distinction matters because unified-memory bandwidth and architecture are fundamentally different from dedicated HBM or GDDR memory systems.
Overlooked Tip: Memory Capacity Is Not the Same as Memory Speed
A huge memory pool can let a model fit while still delivering much lower bandwidth than a high-end discrete AI accelerator. When comparing systems, check capacity, bandwidth, precision support and actual software performance together.
AMD Built the 495 for This Kind of Machine
The Ryzen AI Max+ PRO 495 is part of AMD's Gorgon Halo family of professional processors.
AMD lists 16 CPU cores, 32 threads, up to 5.2GHz boost, 40 graphics compute units and support for up to 192GB of LPDDR5X-8533 memory.
The processor also has configurable power support up to 120W.
That combination gives manufacturers enough compute and memory headroom to create systems that sit somewhere between a powerful mini PC and a compact AI workstation.
The FAEX2 Is Tiny, But It Is Not Barebones
The chassis measures roughly one liter, yet FEVM has squeezed in connectivity that would be unusual even on a larger workstation.
FEVM FAEX2 Hardware Highlights
- Three M.2 2280 slots: Support for PCIe Gen4 SSDs, giving the machine substantial local storage flexibility.
- OCuLink: Provides a pathway for external GPU expansion, although it shares resources with one M.2 slot.
- 10GbE + 2.5GbE: High-speed networking for local servers, storage and workstation workflows.
- USB4: Fast external storage and peripheral connectivity.
- HDMI 2.1 + DisplayPort 2.0: Suitable for high-resolution external displays.
- SD UHS-II: Useful for photographers, video creators and media workflows.
- Wi-Fi 7: Modern wireless networking alongside wired Ethernet.
The Price Is Where Things Get Serious
FEVM's Chinese pre-order pricing makes the FAEX2 a specialist machine.
The 128GB model is listed at CNY 21,599 without an SSD, while the 192GB configuration rises to CNY 43,599.
That works out to roughly $3,200 and $6,500 respectively using the exchange-rate conversions reported by Notebookcheck and other coverage at launch.
The huge jump between memory tiers is important.
You are not simply paying a little more for another 64GB.
The 192GB model is aimed at buyers whose workload genuinely depends on the additional memory capacity.
“I’d like to see everybody have an AI PC.”
AMD's current roadmap shows exactly why that vision is moving beyond ordinary office PCs: local models increasingly need more memory, compute and software support at the edge.
Who Actually Needs 192GB?
This is not the mini PC for someone who wants faster Chrome tabs.
The target audience is much narrower.
FAEX2 Makes the Most Sense for
- Local AI developers: Running larger language models without immediately moving everything to the cloud.
- Researchers: Working with memory-heavy datasets and local inference.
- Creative professionals: Combining AI with video, image or 3D workloads.
- Homelab builders: Running local AI services alongside storage and networking workloads.
- Enterprise prototypes: Testing privacy-sensitive workloads locally before larger deployment.
For ordinary office work, the extra memory is difficult to justify.
How It Compares With Other Ryzen AI Max+ PRO 495 Mini PCs
| System | Memory | AI Positioning | Key Difference |
|---|---|---|---|
| FEVM FAEX2 | Up to 192GB | Local AI workstation | Approx. 1L chassis with 10GbE and OCuLink. |
| MINISFORUM MS-S1 MAX-P495 | 192GB | AI workstation | Dual 10GbE, 80Gbps USB4 V2 and larger workstation chassis. |
| GMKtec EVO-X5 Pro | 192GB | Local AI and workstation | Competitive pricing and 160GB VRAM allocation. |
| CHUWI UniBox AI495 Pro | 192GB | Local AI mini workstation | 2.9L chassis, dual 10GbE and OCuLink support. |
The common denominator is AMD's unified-memory architecture.
The difference is how each manufacturer uses the platform: chassis volume, cooling, networking, storage and external expansion.
Amazon: Alternatives for Local AI Builders
MINISFORUM MS-S1 MAX
A Ryzen AI Max-based workstation line with a strong focus on local AI, unified memory and workstation connectivity. Check the latest configuration available for your market.
Search MINISFORUM MS-S1 MAXGMKtec EVO-X5 Pro
Another Ryzen AI Max+ PRO 495-based system aimed at buyers who want a large unified memory pool for local AI and workstation workloads.
Search GMKtec EVO-X5 ProCHUWI UniBox AI495 Pro
The UniBox AI495 Pro combines the same Ryzen AI Max+ PRO 495 platform with 192GB unified memory, dual 10GbE and an OCuLink expansion path in a larger 2.9-liter chassis.
Search CHUWI UniBox AI495 ProWatch AMD's Ryzen AI Halo Platform in Action
AMD's official presentation explains the thinking behind Ryzen AI Halo and why unified-memory local AI systems are being designed for large-model inference and development.
Pros and Cons of the FEVM FAEX2
What Stands Out
- 192GB of unified LPDDR5X memory.
- Up to 160GB assignable as graphics memory.
- Very small 1-liter chassis.
- 10GbE plus 2.5GbE networking.
- Three M.2 storage positions.
- OCuLink provides external-GPU flexibility.
What to Consider
- 192GB configuration is extremely expensive.
- Memory is soldered and not a normal upgrade path.
- Unified memory is not equivalent to discrete high-bandwidth VRAM.
- International availability is still uncertain.
- High-end local-AI workloads can still outgrow an integrated GPU's compute capacity.
The Overlooked Problem: 192GB Does Not Mean 192GB Is Free for AI
This is where many headlines become misleading.
The FAEX2's 192GB is a shared system-memory pool.
The operating system, CPU workloads, GPU workloads and runtime all compete for that resource.
Even if 160GB is assigned to graphics, the remaining system memory is doing real work.
That means the practical model capacity will depend on the operating system, runtime, quantization and application behavior.
The Smart Way to Judge It
Do not ask only “Can it run a 300B model?” Ask at what quantization, with how much context, at what tokens-per-second, using which runtime and with how much memory available for the rest of the system. Those details determine whether the machine is genuinely useful for your workload.
The Bottom Line on FEVM FAEX2
The FEVM FAEX2 is interesting for one very specific reason.
It shows how far the definition of a “mini PC” is moving.
A roughly one-liter computer now exists with 192GB of unified memory, 160GB of potential GPU allocation, 10GbE networking and external-GPU connectivity built around an AMD processor designed for local AI.
That does not make it a replacement for a full data-center accelerator.
It does make it a radically different kind of edge workstation.
The 192GB configuration is expensive enough that ordinary buyers should look elsewhere.
For developers and researchers who are repeatedly hitting memory limits on conventional PCs, however, the architecture is much more compelling.
The important trend is not FEVM alone. AMD's Gorgon Halo platform is turning large unified-memory pools into a new middle ground between consumer PCs and dedicated AI servers.
And as local AI models become larger and more agentic, that middle ground could become increasingly useful.
Local AI Hardware: RAM vs. GPU VRAM
A 192GB mini PC proves how much memory you can pack into a tiny chassis, but is massive system RAM actually better than a dedicated graphics card? Read our complete guide to explore the critical differences between RAM and GPU VRAM, and discover which hardware upgrade will actually accelerate your local AI models.
Read the RAM vs. VRAM Guide →Sources checked for this article:
Notebookcheck — FEVM FAEX2 Specifications and Pricing
ITHome — FEVM FAEX2 Pre-Order Pricing
AMD — Ryzen AI Max+ PRO 495 Specifications
Frequently Asked Questions
What is the FEVM FAEX2?
The FEVM FAEX2 is a roughly 1-liter mini workstation built around AMD's Ryzen AI Max+ PRO 495, with up to 192GB of LPDDR5X unified memory, 10GbE networking, OCuLink and multiple NVMe storage options.
How much RAM does the FEVM FAEX2 have?
The FAEX2 is offered with up to 192GB of LPDDR5X-8533 unified memory. FEVM says up to 160GB of that pool can be allocated as graphics memory.
Can the FEVM FAEX2 run 300B AI models?
FEVM markets the FAEX2 for local AI models up to 300 billion parameters when using 4-bit quantization. Actual usability depends on quantization, context size, runtime, memory allocation and inference performance.
How much does the 192GB FEVM FAEX2 cost?
The 192GB configuration is listed in China at CNY 43,599 without an SSD, roughly equivalent to about $6,500 based on launch-time conversions. International pricing and availability had not been announced in the latest reports reviewed for this article.
Is 192GB of unified memory better than a discrete GPU?
It solves a different problem. A large unified-memory pool can let substantially larger models fit locally, while dedicated GPUs generally offer much higher memory bandwidth and more specialized compute. The better architecture depends on the workload.
No comments:
Post a Comment