An OCuLink port and a second M.2 slot headline the spec upgrades on GMK's upcoming EVO-X3, a compact desktop built around AMD's Strix Halo silicon. The taller, slimmer chassis fits dual NVMe drives and exposes a PCIe 4.0 x4 link over OCuLink at up to 64 Gbps, opening a path to external GPUs or other PCIe accessories without opening the case.
Two processor tiers are planned. The model shipping first uses the same Ryzen AI Max+ 395 found in the EVO-X2, a 16-core, 32-thread Zen 5 part with 40 Radeon 8060S compute units and support for up to 128GB of 256-bit LPDDR5x-8000. A later PRO 495 variant pushes the memory ceiling to 192GB at LPDDR5x-8533 speeds and bumps NPU throughput. Both chips share the same core configuration, but the higher tier targets workloads that benefit from the extra bandwidth and capacity.
Unified memory is the real draw for anyone running local inference on Linux. Strix Halo lets the integrated GPU address the full system memory pool, so a 192GB configuration can hold model weights that would otherwise demand a multi-GPU server. AMD has been upstreaming ROCm and amdgpu support for the gfx1151 graphics block over the past year, and the platform now runs cleanly on current stable kernels. The ROCm compute stack for inference requires additional kernel configuration to expose the full unified memory pool rather than the small BIOS VRAM carveout the kernel defaults to, and initializing the HIP stack on gfx1151 requires setting HSA_OVERRIDE_GFX_VERSION=11.5.1. AMD's Strix Halo system optimization guide in the ROCm documentation covers those steps, with Ubuntu 24.04 LTS users needing HWE kernel 6.17 or newer for full RDNA 3.5 support.
The EVO-X2 has already attracted an active Linux community around the same silicon. The pablo-ross/strix-halo-gmktec-evo-x2 repository documents a complete Ubuntu 24.04 setup for llama.cpp inference with ROCm 7 and rocWMMA, and throughput figures tracked in Ollama's issue tracker show around 40 tokens per second on quantized 30B models via the ROCm backend. AMD's TheROCk nightly builds deliver significantly better performance than the stock ROCm 7.2 release for anyone willing to track prerelease packages, closing some of the gap against the NVIDIA DGX Spark systems aimed at the same desktop AI inference niche.
GMK first teased the EVO-X3 in May, and the Ryzen AI Max+ 395 configuration is expected to ship soon. Pricing has not been announced. The outgoing EVO-X2 started at $1,500 (€1,380) with the same chip, and higher memory tiers carried significant premiums on top of that base.
