UMA vs. NUMA
As mentioned on the last page, you switch between memory access modes on Threadripper processors. 'Distributes' mode uses 'Uniform Memory Access' is enabled by default so this is what we've been testing with so far. However, you can switch to 'local' mode for 'Non-Uniform Memory Access' and this allows each of the Zeppelin dies to prioritize which cores access certain parts of the system memory. This basically prioritizes the nearest cores to improve overall latency for gaming applications that tend to place a premium on fast memory access. So let's see how this impacts gaming performance as well as a few productivity workloads...
Switching to NUMA has improved the Battlefield 1 performance and now Threadripper is delivering similar frame rates to the Ryzen 7 1800X. In fact, the 1950X creeps ahead ever so slightly and now is that much slower than the Core i7-7820X and Core i9-7900X.
F1 2016 also sees decent gains when switching to NUMA though this time isn't able to match the Ryzen 7 1800X and this means it's also well down on the Intel CPUs.
Performance in Civilization VI was already great and no extra performance was had when switching to NUMA.
Looking at application performance, we see a slight performance decline with POVRay, but nothing serious.
Blender also saw a slight decline in performance with the 1950X being about 4% slower using NUMA.
Last up we have the Corona test, in which both NUMA and UMA offered the same performance. There's not much to see here, so let's move on to some consumption figures.