I see. So for a large dataset you would still hit a bottleneck, compared to a main mem store which also uses compression. But nonetheless it's a good idea to do certain operations on hot data on the GPU.

Which operations benefit the largest in your experience? I would guess scans with simple predicates and expressions. What about aggregations?