The newly launched Lenovo Thinkstation PGX is designed for local AI tasks, featuring a robust Nvidia GB10 Grace Blackwell Superchip and 128GB of unified memory. Priced at $3539, this mini PC is tailored for developers who need to prototype or work on smaller AI models without a focus on speed. It stands out with significant VRAM capacity, making it suitable for quantization and fine-tuning processes.
Although the PGX is not as fast as individual high-end GPUs like the Nvidia H100 or RTX 5090, it serves a niche purpose within the Nvidia ecosystem, particularly with its compatibility with CUDA. The system includes ports such as 4 USB-C, HDMI, Ethernet, and 2 QSFP, enhancing its connectivity options for various applications.
Built on a reference design shared with other models like the Dell Pro Max AI and Asus Ascent GX10, the PGX operates on Nvidia’s DGX OS, which is based on Ubuntu and comes pre-installed with essential tools like CUDA 13 and TensorRT. This design allows for a consistent hardware experience across different platforms while varying in branding and support.