What is the ideal GPU horsepower required to make a specific application run at a premium, and how do you manage dynamic GPU workloads for AI and Machine Learning?
The short answer is, it varies. For some applications, it's a linear scaling capability. For example, four GPUs is four times better than one GPU.
For other applications, it plateaus. Perhaps by throwing more than two GPUs at the problem doesn't buy you any additional performance.
The way that a lot of customers handle it today is they have fixed servers with two, four or eight GPUs inside of it, and they hope that the workload that comes in best matches one of the free servers. This is where Composable Infrastructure shines. Watch this video to learn more.