GMKtec EVO-X3 workstation debuts with 128GB unified memory for AI

The launch of this hardware signals a shift in compact computing where memory capacity, rather than raw clock speed, becomes the primary bottleneck for generative AI production. By providing 128GB of unified memory, GMKtec enables creators to run massive models locally that typically require enterprise-grade data center GPUs. Within the streaming ecosystem, this lowers the barrier for localized video generation and real-time AI metadata processing at the edge. As production workflows increasingly move toward local AI inference to reduce cloud costs and latency, watch for how the 140-watt thermal design handles sustained rendering loads in such a small chassis.

AMD's Ryzen AI Max+ 395 processor, codenamed Strix Halo, has become the anchor silicon for a growing class of compact AI workstations targeting local inference workloads. Framework announced its Desktop platform in early 2025 using the same Ryzen AI Max+ 395 chip, pairing it with up to 128GB of unified memory in a mini-ITX form factor aimed at developers and content creators who need GPU-class memory bandwidth without a discrete card. The Radeon 8060S integrated GPU inside Strix Halo delivers 40 compute units with access to the full unified memory pool, a configuration that AMD has positioned as a direct alternative to entry-level discrete GPUs for AI and creative tasks. AMD formally detailed the Ryzen AI Max+ 395 specifications at CES 2025, highlighting 50 TOPS of NPU performance and support for up to 96GB of GPU-accessible memory, though OEM partners like GMKtec have since pushed the unified memory ceiling to 128GB.

The compact AI workstation segment is attracting multiple hardware vendors competing on memory capacity, thermal design, and expansion flexibility. Minisforum launched its MS-A1 workstation in mid-2025 with the same Ryzen AI Max+ 395 processor and 128GB unified memory, positioning it as a local LLM inference box capable of running 70-billion-parameter models without cloud dependency. ASUS also entered the segment with its ProArt PX13, a convertible laptop using the Ryzen AI Max+ 395, targeting mobile creators who need on-device AI acceleration for video editing and image generation. The competitive pressure among these vendors centers on sustained performance under thermal constraints, OCuLink expansion for external GPU scaling, and pricing relative to cloud inference costs.

Independent benchmarks of Strix Halo systems have focused on memory bandwidth as the critical variable for LLM inference throughput. Testing by Notebookcheck on Ryzen AI Max+ 395 systems showed that the 256-bit LPDDR5X-8000 memory bus delivers approximately 256 GB/s of bandwidth, which is roughly half the bandwidth of an NVIDIA RTX 4090 but at a fraction of the power draw and cost. For streaming production use cases such as real-time transcription, scene detection, and AI-assisted metadata tagging, that bandwidth is sufficient for models up to approximately 30 billion parameters at interactive token rates. Igor's Lab measured the EVO-X3's sustained power consumption under mixed CPU-GPU loads at approximately 120 watts, confirming that the 140-watt TDP envelope leaves headroom for thermal throttling avoidance during extended rendering sessions, a relevant consideration for studios evaluating these systems as in distributed production pipelines.