GMKtec's EVO-X3 is now shipping with AMD's Ryzen AI Max+ 395, a 16-core, 32-thread processor paired with 40-core Radeon 8060S integrated graphics and 128GB of quad-channel LPDDR5X-8000 memory delivering 256GB/s of bandwidth. That unified memory pool is the real story here. On Linux with ROCm, the GPU can address shared system memory through AMD's GTT path without the fixed allocation ceiling that Windows imposes, and benchmarks show 20 to 30 percent better inference performance under Linux compared to Windows on the same hardware. GMKtec claims the 128GB configuration can run models up to 235 billion parameters entirely on-device.
The EVO-X3 takes a vertical triple-fan chassis design, a departure from the EVO-X2 that aims to improve thermals and reduce noise under sustained workloads. The metal enclosure measures 353 x 186 x 41 mm (13.9 x 7.3 x 1.6 inches). Inside, two M.2 2280 slots accept up to 16TB of PCIe 4.0 x4 storage, and connectivity includes USB4 at 40Gbps with video output, an OCuLink port for external PCIe 4.0 x4 accessories, HDMI 2.1 with 8K output, 2.5G Ethernet via Realtek RTL8125BG, and WiFi 7 with Bluetooth 5.4 through an AMD/MediaTek RZ717 card.
Phoronix testing of the Ryzen AI Max+ PRO 395 on recent distributions like Ubuntu 25.04 and Fedora Workstation 42 confirms strong Linux support out of the box, though ROCm rather than the Adrenalin UI handles GPU memory tuning through kernel boot parameters like amdgpu.gttsize. AMD's dedicated RDNA 3.5 system optimization guide covers those configuration steps for Strix Halo hardware specifically. ROCm 7.2, released in March 2026, auto-detects the Strix Halo GPU (gfx1151) and inference tools including Ollama, llama.cpp, and LM Studio work on that release, though AMD's compute compatibility matrix marks Ryzen APU support as Preview rather than production for general compute workloads. The Mesa RADV Vulkan driver offers a complementary path that sidesteps those caveats: community testing on Strix Halo hardware shows the RADV backend outperforming ROCm HIP on gfx1151 for inference while also being easier to configure, with llama.cpp via Vulkan/RADV reaching around 101 tokens per second on 30B-parameter models. vLLM requires ROCm nightly builds as a workaround since gfx1151 is not in upstream vLLM's supported GPU list. That flexibility makes Strix Halo machines particularly attractive for local LLM inference and self-hosted AI workloads where maximizing available VRAM matters more than plug-and-play convenience.
The Ryzen AI Max+ 395 silicon has attracted substantial community documentation that applies directly to the EVO-X3. The pablo-ross/strix-halo-gmktec-evo-x2 repository provides a complete Ubuntu 24.04 setup workflow covering llama.cpp compilation with ROCm 7 and rocWMMA along with benchmark results for the same hardware. The hogeheer499-commits/strix-halo-guide covers Ollama, llama.cpp Vulkan/RADV, and ROCm paths in depth, including runs of 120B GGUF models on-device. llm-tracker.info aggregates community inference benchmarks across backends for the platform. On the vendor side, GMKtec's partnership with AMD's ROCm Lab means the EVO-X3 ships with HelloROCm and a pre-configured ROCm development environment, providing a defined starting point for developers setting up the stack.
The EVO-X3 starts at $3,600 (€3,310) for the 128GB/2TB configuration, with a 128GB/4TB model at $3,850 (€3,540). Both are available now through GMKtec's website.
