Berth: A Lightweight GPU Resource Manager for CUDA Workloads
Berth v0.4.1 has been released as a Go-based GPU resource manager that coordinates access, prioritization, and live-swapping for CUDA workloads like llama.cpp and ComfyUI. It uses leases, a request queue, and an OpenAI-compatible API to prevent VRAM conflicts and streamline local AI operations. This tool solves a major pain point for local AI practitioners who frequently experience VRAM contention when running multiple heavy applications like language models and image generators simultaneously. By automating resource handoffs and process parking, it makes local multi-app AI workflows significantly more practical. Written as a single Go binary, Berth supports Linux NVIDIA out of the box with experimental AMD Linux and Windows support. It features automatic game management, GPU telemetry monitoring, model discovery across multiple directories, and agent integration for tools like Claude Code.
## BACKGROUND
CUDA is a parallel computing platform and programming model developed by NVIDIA that enables developers to use graphics processing units (GPUs) for general-purpose processing. Tools like llama.cpp and ComfyUI heavily rely on CUDA to accelerate the execution of large language models and diffusion-based image generation workflows locally.