I'm very curious how they work the cache hierarchies of these hybrid clusters. You really want to be tuning the latency and throughput of the cache to the consumers attached to it and that wouldn't be straightforward in this case.