ComfyUI on Nvidia DGX Spark (blog.comfy.org)

NVIDIA’s new DGX Spark — a compact, NUC/Mac‑mini sized system built around the GB10 Grace Blackwell Superchip — can run ComfyUI out of the box on Linux and the author’s tests show all tried models work. The device ships with 128 GB of LPDDR5x coherent unified system memory (usable as both CPU RAM and GPU VRAM), NVIDIA’s AI software stack (CUDA, etc.), and dual ConnectX‑7 SmartNICs. Installation follows standard ComfyUI/NVIDIA Linux instructions, and the team promises formal benchmarks in a future post. Why it matters: the DGX Spark’s unified memory architecture lets you run models and workloads that exceed typical high‑end desktop GPU memory, so it’s useful for larger inference or model‑editing tasks that would otherwise be impossible on single‑GPU desktops. It won’t outpace a flagship desktop GPU (e.g., an RTX 5090) in raw throughput, but it offers a compact, turnkey platform for memory‑hungry ML workloads and potentially for multi‑GPU setups via its SmartNICs. For ComfyUI users this means easier handling of larger models and a portable, NVIDIA‑optimized environment for development and experimentation.