Looks like CCX latency may not be as much of an issue as originally thought, at least for games. Memory bandwidth is the real bottleneck here.
https://www.youtube.com/watch?v=Rhj6CvBnwNk
intelfags on suicide watch
>>59603184
this doesn't even take into consideration that the infinity fabric latency is linearly related to memory latency. The infinity fabric is always clocked at exactly half of the memory frequency, so once memory speeds can actually hit high marks, the inter core latency will also decrease drastically making the CCX argument fairly null.
The additional cache is actually advantageous for productivity reasons. Cache increases the effective memory bandwidth, I don't know if the effect is enough to offset the loss in inter-CCX communications though.
I've seen cross ccx latency drop from 100ns to 60ns with faster memory, it's definitely both.
Maybe AMD will make a SKU that runs the data fabric at full memory speed or in future running its own protocol that's not DRAM
It is an issue if youre using 2133Mhz RAM
>Looks like CCX latency may not be as much of an issue as originally thought, at least for games.
No fucking shit, these things are always blown out of proportion.
>>59603184
The latency thing is largely a misinterpretation by idiots.
It's actually mostly bandwidth bottlenecking between the CCXs, especially when threads get migrated and L2s need to re-warm from either the opposite L3 pool or the DRAM controllers.
Latency does scale inversely with the bandwidth, but it's a secondary symptom and not the real problem.
Surprise, surprise.
>>59603184
>>59603608
>>59605045
Someone hasn't seen the phoronix tests
http://www.phoronix.com/scan.php?page=article&item=amd-ryzen-cores&num=3
Spec the differences between 4+0 / 2+2 and 2+0 / 1+1
This is worst case scenario, they're using 2133mhz dual rank memory.