Presbytier
[H]ard|Gawd
- Joined
- Jun 21, 2016
- Messages
- 1,058
From Ryzen: Strictly Technical thread @ Anand:
Clearly shows that there is a performance penalty when threads are on different CCX. Second video shows that Windows scheduler just assigns the threads "randomly" and why we are seeing such a varied results from game benchmarks.
I'm 100% sure this doesn't happen on intel 6-8 core processors because they don't have this CCX interconnect bw limitation as they are one monolithic die. It's no wonder we are seeing quite big gains with faster ram on ryzen because this "Infinity Fabric" (HyperTransport with a fancy name) which communicates between CCX is tied to ram speed.
The real question is that can it be fixed so that windows scheduler tries to keep threads which have dependency on each other on the same CCX. Maybe just make it behave like NUMA
That defeats the point of an 8 core CPU if I need to keep process locked to 4 logical core to ensure I don't get communication issues across the board. If this is true the 6 core Ryzen is in a world of hurt as that would be 3 cores per CCX and games generally scale well to 4 cores. I'm a bit doubtful that is the issue though because it doesn't seem to effect other applications outside of games that will spread across alll 8 cores.