> Also peak performance is allready passed even with multicore because of RAM latency.

By that logic CPUs from today should be slower than those from 2005, since the real (uncached) memory latency has increased (significantly) for most systems.

Though I suppose this means you could actually engineer workloads that run slower on a modern 5 GHz CPU than on a 2005-era Athlon 64.

IBM already have a chip with a gigabyte of L3 cache, so maybe that combats the latency in another way

Reducing average memory read/write latency (as well as increasing bandwidth) is the point of big caches - but, cet. par. bigger caches inherently have higher lookup latency, deeper cache hierarchies increase the latency further and cache sizes are also limited by ISA concerns (4K pages on x86 limit the size of VIPT caches). So a system with a deep cache hierarchy and large caches will perform better on average, but actually going out there and getting bits from main memory will take longer. Not least because the actual, physical latency of DRAM itself improves only very, very slowly, this makes old systems stand up quite well in this particular metric against modern systems.

> I'm curious what the longevity at 100% will look like for these.

Indeed. At this feature size I'd already start to be worried how long the chip will work.

Feature size seems to be 15nm, from the article. 2nm is marketing speech for transistor density.