It's not really about FLOPS. These sorts of many core, low performance per core systems do well in handling lots of I/O where there isn't all that much process

It's not really about FLOPS. These sorts of many core, low performance per core systems do well in handling lots of I/O where there isn't all that much processing involved per node, and/or micro-latency isn't much of an issue, but lots can operate in parallel. For actual FLOPS/kW where the FLOPS matter, you still can't beat GPUs, or for non GPU-able code high-wattage high-performance CPUs that can chew through data like nobody's business. These low power systems' FLOPS take more kW in aggregate to perform the same work in similar ballpark times. But when things are I/O bound, the low power cores should be able to handle messages & transactions with much less wattage overhead than what the high single-thread performance CPUs have to spool up.