:strip_exif()/i/2005279580.jpeg?fit=%2C&ssl=1)
Is an Achilles heel of Ryzen that has been identified for a while, the latency between the cores can enhance, particularly if it’s a must to go from CCD to CCD.
additionally see https://www.anandtech.com…-5800x-and-5700x-tested/5
Within the Epyc story, suppose you need very low latency to community you’ll be able to set the community card as most well-liked IO, see level 3.1.12 in subsequent AMD doc
https://www.amd.com/syste…003-series-processors.pdf
However you even have to inform Home windows to run its NIC in the identical neighborhood as the place the NIC’s PCIe is, see 3.3.2 within the following AMD doc
https://www.amd.com/syste…dows-server-56782_1.0.pdf
If I translate that to gaming with an AMD with a number of CCDs, worst case state of affairs the sport runs on cores in a CCD the place the PCIe lanes don’t come from the GPU and the place the PCIe lanes don’t come from the storage.
That you just then get a efficiency hit is logical given the latency that may happen. A lot of cores and plenty of CCD, all very good however it provides latency and it begins to tackle proportions that you just discover them.
Level of criticism in direction of AMD, it shouldn’t be in 2022 {that a} recreation developer or a sysadmin has to determine which cores our load runs on and which cores our PCIe is on. These sorts of optimizations ought to occur mechanically.
And to instantly reply my criticism, a big a part of it’s within the OS so Microsoft, there’s nonetheless work to be completed!
*I’ve no expertise with Epyc in Linux, if one other sysadmin has, be at liberty so as to add.
