Setting Up a 16x ASUS GX10 (DGX Spark) Local AI Cluster
A user has shared a hardware configuration for a local 16-GPU cluster using ASUS Ascent GX10 systems (DGX Spark) and a MikroTik CRS804-4DDQ switch. This setup is designed to run massive, frontier-level open-source AI models locally, such as DeepSeek v4 Pro and Kimi K3. This DIY setup demonstrates the feasibility of running frontier-level models with over 2 trillion parameters locally, bypassing the need for cloud-based API providers. It highlights the growing trend of high-performance local AI infrastructure for enterprise-grade or research workloads. The cluster links 16 ASUS GX10 units, which feature NVIDIA Blackwell GPUs with 128GB of unified memory, using a MikroTik CRS804-4DDQ 400G switch with four 400G-to-100G breakout cables. The user plans to split the cluster to run two separate models on 8 GPUs each, or combine them to run 2T+ parameter models.
## BACKGROUND
The ASUS Ascent GX10 is a compact AI supercomputer powered by NVIDIA Blackwell GPUs and equipped with ConnectX-7 networking for multi-system stacking. The MikroTik CRS804-4DDQ is a high-speed 400G switch designed to handle the high-bandwidth demands of AI clusters and storage fabrics. Stacking these systems allows users to run massive models that exceed the memory capacity of a single machine.