NVIDIA artificially restricts peer-to-peer (P2P) GPU communication to their enterprise cards. Turns out this is a software limitation, not a hardware one. I patched my drivers to remove it, hacked vLLM to take advantage of it, and got a 15-50% throughput improvement running Qwen 3.5 35b on dual RTX 3090s.
NVApi is a small application I’ve written for monitoring and presenting utilisation metrics from Nvidia GPUs. This can be used to monitor GPU memory, temperature, power usage, and utilisation of GPUs in a system and can easily be integrated into tools such as HomeAssistant or Prometheus. The package uses the Nvidia Management Library (NVML) and provides a simple API for monitoring Nvidia GPUs along with a basic GUI client. NVApi on Github ...