GMKtec has officially unveiled the EVO-X3, a mini-PC designed specifically for software developers testing and running local AI inference. Building on the popularity of AMD's Ryzen AI Max+ chips in the compact workstation market, this new iteration swaps the previous plastic chassis for a sleek, standing metal monolith. The system is headlined by the AMD Ryzen AI Max+ 395 processor, which pairs 16 cores and 32 threads with a Radeon 8060S iGPU featuring 40 compute units.

Memory and Connectivity Highlights

The standout feature for AI workloads is the 128GB of soldered LPDDR5X-8000 memory, providing the high bandwidth necessary for large model inference. Unlike most competitors, GMKtec has integrated an OCuLink port, offering PCIe Gen4 x4 connectivity for external GPUs. This port delivers significantly more bandwidth than Thunderbolt 4, though it requires the system to be powered down before connecting or disconnecting devices, a limitation GMKtec explicitly warns about with stickers near the port.

Design and Thermal Management

The EVO-X3 adopts a vertical, monolithic design measuring 353 x 186 x 41 mm, which GMKtec claims aids in thermal management. The chassis is fully metal, a notable upgrade from the EVO-X2, and features extensive ventilation on the sides and top. Despite the thin profile, the system relies on a 230W external power supply and includes a required base for stability. The rear I/O includes two HDMI 2.1 ports, a single USB4 port, and a 2.5GbE Ethernet jack, though the lack of 10GbE networking is a missed opportunity for a device marketed as an AI workstation.

Key Takeaways

  • The EVO-X3 features 128GB of LPDDR5X-8000 memory, ideal for local LLMs.
  • OCuLink support allows for high-bandwidth eGPU connections, surpassing Thunderbolt 4.
  • The system uses a vertical metal chassis with extensive ventilation for thermal control.
  • Networking is limited to 2.5GbE, which may bottleneck large dataset transfers.

The Bottom Line

For developers needing high memory bandwidth in a compact form factor, the EVO-X3 is a compelling, if slightly overpriced, option. The inclusion of OCuLink is a genuine win for expandability, but the lack of 10GbE networking feels like an oversight in an otherwise well-thought-out AI workstation. If you are building a local inference pipeline, this box checks the most important boxes: memory capacity and bandwidth. Just be prepared to manage the thermal constraints of a 41mm-thick metal slab and the hassle of hot-plugging restrictions with OCuLink.