GMKtec unveils an AMD Strix Halo mini-PC with 128 GB of RAM and 96 GB allocable to the GPU

GMKtec launches a more affordable version of its mini-PC Strix Halo. The particularity: 128 GB of unified memory with up to 96 GB allocable to the integrated Radeon 8060S GPU - a boon for LLM inference at home.

The context

Notebookcheck reports that GMKtec (a Chinese brand specializing in mini-PCs) is marketing a new variant - the GMKtec EVO-X2 - of its mini-PC built around AMD Ryzen AI Max+ 395 "Strix Halo". The remarkable point: 128 GB of LPDDR5X-8000 in a unified configuration, of which up to 96 GB can be allocated to the integrated Radeon 8060S GPU.

At a time when 70B language models run locally as long as you have VRAM, this configuration changes the game for AI hobbyists - without going through an NVIDIA card costing four digits.

What you need to know about Strix Halo

AMD Strix Halo (marketing name: Ryzen AI Max/Max+), officially announced at CES 2025, combines:

• Up to 16 Zen 5 cores (32 threads)
• Integrated Radeon 8060S GPU based on RDNA 3.5, up to 40 compute units
• XDNA 2 NPU at ~50 TOPS
• 256-bit memory bus LPDDR5X - this is the key: bandwidth ~256 GB/s, twice that of a classic laptop CPU in LPDDR5X 128-bit

The memory is unified (shared between CPU and GPU), in the manner of an Apple M-series. This is what allows dynamically allocating a large portion to the GPU - unthinkable on a classic PC with a dedicated GPU with 8-16 GB of VRAM.

Design / Handling

The EVO-X2 comes in a compact SFF chassis (approximately 14×14 cm, to be confirmed on the final product sheet), typical of the GMKtec range. Active cooling with a radial fan. External power supply.

On the I/O, we expect the recent GMKtec standard: USB4/Thunderbolt-compatible, HDMI 2.1, DisplayPort, several USB-A/C, 2.5G LAN. To be confirmed precisely on the product sheet.

Performance (expected)

Strix Halo positions its integrated GPU in the same league as an RTX 4060 mobile at ~4070 mobile, according to AMD tests and the first independent benchmarks (Notebookcheck, Level1Techs). For LLM inference:

• 30B model in Q4_K_M: between GPU offload and CPU, you can aim for 10-15 tokens/s locally
• 70B Q4 model: possible thanks to the 96 GB GPU, but around 3-6 tokens/s depending on the inference layer (llama.cpp Vulkan/ROCm)
• Image models (SDXL, FLUX): playable, slower than a dedicated RTX but usable

The real comparator remains Apple M4 Max with 128 GB of unified RAM (~4,500 €). The EVO-X2 should fall below this bar - the Notebookcheck article mentions a "cheaper" variant without giving a firm price, to be monitored.

Autonomy

Not applicable: it's a desktop. Expected consumption: configurable TDP 55-120 W depending on the mode, i.e., continuous consumption around 100-150 W during inference.

Provisional verdict

For those who want to run local LLM, light fine-tuning, or image generation without buying an RTX 4090 (24 GB VRAM) or an A6000, this configuration is the best accessible alternative in 2026 in the x86 world. The direct competitor on the Apple side is the Mac Studio M4 Max, more expensive but with a more mature software ecosystem (MLX, Metal) for AI.

Our reservations: effective availability (GMKtec delivers in waves), international warranty, and the state of ROCm support on Radeon 8060S at launch - historically AMD has been slow to support its iGPUs under ROCm, which limits the AI software stack.

Summary table:

To be closely monitored upon actual release.