The side channel for most of the spectre variants is the latency of misses on the cache lines. L1 and L2 cache lines are local to a physical core. As far as I k

The side channel for most of the spectre variants is the latency of misses on the cache lines. L1 and L2 cache lines are local to a physical core. As far as I know, nobody has made any of the spectre variants work by measuring the latency of L3 cache misses, which are local to NUMA nodes if I understand correctly, but I'd love to hear otherwise. The most recent round of spectre variants measured the latency of the line fill buffers and other parts which are local to a physical core's memory subsystem.