Run large AI models locally with high memory and fast connectivity while reducing latency cloud use and keeping full control of data at the edge.
ASRock Industrial has launched the AI BOX-A395, a compact system designed to run AI workloads locally without depending on the cloud. It is built on AMD Ryzen AI Max+ 395 processors that combine CPU, GPU, and NPU, delivering up to 50 TOPS of AI performance. With support for up to 128GB LPDDR5x-8000 unified memory, the system can handle large AI models and data-heavy workloads directly on-device, reducing latency and keeping data local.
The system is positioned for enterprises, developers, and system integrators working on edge AI deployments. It supports workloads such as AI model development, local inference, engineering design, and media processing. It also enables applications like large language models, generative AI, and vision AI to run at the edge, lowering cloud costs and improving response time.
The processor includes up to 16 cores and 32 threads, along with an integrated Radeon 8060S GPU and XDNA 2 NPU. It supports both Windows and Linux, allowing flexibility across development and deployment environments. The unified memory architecture helps manage memory-intensive workloads without the need for external accelerators.
For connectivity, the system includes two USB4 ports, one USB 3.2 Gen2 Type-C, two USB 3.2 Gen2, and two USB 2.0 ports. It also provides dual Ethernet with one 10GbE and one 2.5GbE LAN port for high-speed data transfer. Display support includes HDMI 2.1 and DisplayPort 2.1 interfaces, enabling up to four displays with resolutions up to 8K.
Storage and expansion options include two M.2 Key M slots with PCIe Gen4 x4 support and RAID 0/1, along with an M.2 Key E slot for Wi-Fi 7 and Bluetooth 5.4. The system uses a thermal design with heat pipes, a copper base, and airflow optimization to support sustained workloads.
The unit is housed in a 200 × 100 × 232 mm aluminum chassis and includes features such as TPM 2.0 and redundant BIOS to support long-term enterprise deployment.
Click here for the original announcement.
