NVIDIA just released NVIDIA DGX Spark. For those who don’t know this is a small form factor computer similar in size to a NUC or a Mac mini. What is special with this one is that both the CPU and GPU are in a single unit made by Nvidia and it has 128GB of unified memory that can be both ram and vram.
NVIDIA DGX Spark supports running ComfyUI extremely well. All the models I have tried work. It’s running Linux so to install ComfyUI on it you can just follow the regular NVIDIA instructions in our readme.
Performance wise it doesn’t beat a full desktop with a 5090 but it can run models and workloads that are too large for even high-end desktop systems.
Key features and technology callouts for DGX Spark platforms:
NVIDIA GB10 Grace Blackwell Superchip
128 GB LPDDR5x Coherent Unified System Memory
NVIDIA ConnectX-7 SmartNIC
(we have two of them so we might try it in the future to see how well it works with the ComfyUI multi GPU stuff)NVIDIA AI Software Stack
(comes with all NVIDIA CUDA, etc.. stuff already installed)
We will have benchmarks in a future blog post.
The unified memory architecture here is a game-changer for ComfyUI workflows. Unlike traditional setups where GPU and CPU memory are seperate, the GB10's 128GB unified approach eliminates data transfer bottlenecks when working with large models and high-resolution images. This means you can load massive multi-modal models alongside your image generation pipeline without worrying about memory fragmentation. The fact that it's being deployed to students, researchers, and developers shows Nvidia is serious about democratizing access to cutting-edge hardware for the next generation of AI creators.
The integration of ComfyUI with NVIDIA DGX Spark is a smart move. Having a pre-configured Slurm cluster with ComfyUI makes it much easir for teams to scale their workflows without the typical infrastructure headaches. The GB200 GPUs and liquid cooling system should provide excellent performance for complex image generation tasks while mantaining efficiency.