How Many ComfyUI Instances Can One GPU Handle? A Developer Found Out

A developer ran controlled experiments to determine how many parallel ComfyUI instances a single GPU can support before performance degrades, filling a gap in publicly available empirical data. Tests were conducted on an NVIDIA RTX 3080 Laptop with 16GB VRAM running ComfyUI 0.24.0, using a standard Stable Diffusion 1.5 txt2img workflow at 512x512 resolution. ComfyUI natively processes only one job at a time, forcing queued requests to wait sequentially regardless of available VRAM, a long-standing limitation the community has sought to work around. The experiment involved spinning up multiple independent ComfyUI processes on separate ports, each with its own VRAM allocation and execution queue. The key finding was counterintuitive: VRAM capacity was rarely the limiting factor — the GPU compute bus became the bottleneck first.