The original headline is: "Running Multiple ComfyUI Instances in Parallel on a Single GPU — What Actually Breaks First"

Short summary

An empirical investigation into running multiple ComfyUI instances in parallel on a single GPU, revealing that VRAM capacity is rarely the bottleneck — the GPU compute bus is. Tests on an RTX 3080 Laptop with SD 1.5 show that independent processes on separate ports work, but SQLite file-level write locks require the --database-url flag to avoid conflicts. Findings transfer to production setups using FLUX.2 Klein 4B on RTX 5090.

• •VRAM capacity is almost never the bottleneck; the GPU compute bus is
• •SQLite file-level write locks cause conflicts between instances — use --database-url to fix
• •Independent processes on separate ports work but performance degrades from compute bus contention, not memory

Generated with AI, which can make mistakes.