Hacker News Reader

Traditionally AMD does NUMA differently from Intel. Would like to see a comparison focusing on NUMA.

3950X and the single socket versions of new EPYC chips aren't NUMA. Instead of each CPU die having its own memory controller there is a central I/O die with an IF link to each CPU die.

Ryżem branded cpus have single NUMA zone, threadripper have 2, EPYC have 2 per socket

This was true up until now, but the upcoming 3900X and 3950X have 2 separate dies ("chiplets"). I assume this means it will have the same architecture as the current TR 2920X and 2950X.

Edit: opencl post above says apparently not, it's a different memory architecture.

The 2 smaller chiplets are for core complex, the memory controller is on the IO die(the other bigger one) that these 2 CCX share. So there is still one single memory controller, and hence, not NUMA.

All memory access is through the single I/O die, so neither the new Ryzen nor the new Epyc are technically NUMA.

However, because all cores in a single chiplet share a single L3 cache, doing NUMA-like optimizations will still yield significant benefits.

AFAICT.

The bigger EPYCs have four NUMA zones per socket (at least the 7551P does).

Reply on news.ycombinator.com