On clusters of non-uniform machines, running some benchmarks could be interesting to inform workload placement. I suspect the more useful info will be aggregated CPU, file, and persistent storage performance, but depending on use case, you might want to go deeper.
Canonical's MAAS (disclaimer: I worked on it) could select physical nodes by GPU presence for some workloads.
On clusters of non-uniform machines, running some benchmarks could be interesting to inform workload placement. I suspect the more useful info will be aggregated CPU, file, and persistent storage performance, but depending on use case, you might want to go deeper.
Canonical's MAAS (disclaimer: I worked on it) could select physical nodes by GPU presence for some workloads.