CCX / Cache / Memory Performance Scaling

One of the most debated questions on quad-core Ryzen processors was AMD's decision to configure the "Summit Ridge" silicon such that each of its two quad-core complexes (CCXs) has two of its cores enabled, and in some of the lower models with even just 4 MB L3 cache per CCX. Some of the earlier comparisons between such chips, and artificially disabling an entire CCX on an 8-core chip (to simulate a quad-core, single CCX chip with 8 MB L3 cache), showed a mixed bag of results.

We picked WinRAR to investigate this performance scaling as it's one of the tests that saw the biggest differences in our Ryzen Memory Performance Analysis article. For these tests, all settings were identical - the only difference is the memory frequency.



Many people expected that Raven Ridge wouldn't be as dependent on memory speeds as Summit Ridge due to the single CCX configuration.

Unfortunately, the opposite is the case. As we change memory speeds, the gap between Summit Ridge and Raven Ridge gets bigger, with Raven Ridge showing the larger performance loss. AMD's explanation is that due to the smaller cache size on Raven Ridge, the memory speeds matter more, which makes sense. Four cores and eight SMT units compete for access to L3, and whatever can't be stored stays on the memory.

The memory controller is also serving the iGPU and not just the CPU, so we ran a second round of tests using the Vega 11 integrated graphics and Witcher 3 (lowest settings, 1080p).


Performance of the integrated graphics, which relies on system memory entirely, scales linearly with memory clocks, which is as expected. It's important to point out that the difference in performance between 2133 MHz and 3200 MHz is huge (30%), and faster memory can make the difference between "playable" and "slide-show."