We have TPC-H benchmarks for single node with a small scale factor in the contributors guide. We only benchmark against Spark though and not against Spark RAPIDS.

datafusion.apache.org/comet/contri...

I hate to say it, but "it depends". I'd recommend running your own benchmarks for your specific workloads. Performance will also vary greatly by environment (number of CPUs vs GPUs, different GPU types, and so on).

Is this using Arrow and/or DataFusion? If so, our Discord is probably a good place to ask.

datafusion.apache.org/contributor-...