A guide about VMware vSphere Bitfusion performance best practices has been published. Bitfusion provides a way to share GPUs among multiple VMs—a GPU card installed in one ESXi host becomes available to all VMs in a cluster (even those on separate hosts) that have the Bitfusion client installed. In addition, Bitfusion dynamically allocates a partial GPU, a single GPU, or multiple GPUs to a client VM and releases the GPU resources when they are no longer being used. The model of using GPUs as a shared resource lets you get the most out of your GPU card investment.
In the guide, we provide network and other tunings to get the best performance from Bitfusion. The recommendations in this guide are based on performance tuning we investigated and applied in our testbed.
The network section describes how DirectPath I/O (passthrough) achieves the best performance for all workloads we tested and especially for high performance computing, artificial intelligence, and machine learning applications. VMs running less compute-intensive applications can use VMXNET3 or PVRDMA, either of which gives you the management and virtualization features of vSphere.
You can read the guide here: VMware vSphere Bitfusion Performance Best Practices.
Other useful links: