Remember that AMD AGESA update (version 126.96.36.199) that Jeff told y'all about back in May? It's finally finding its way to end users by way of motherboard firmware updates. Some gerbils have already gotten theirs, but if you have a Ryzen machine with some hot memory that hasn't been updated—and a few hours to spend tweaking the machine—you should probably go ahead and do that. AMD recently updated its gaming blog with some testing results from the new firmware, and at least with the configuration AMD's tester was using, there's some serious performance potential laying on the table versus stock settings.
Before we begin, it's worth noting that these are manufacturer-provided numbers, and they should all be taken with the appropriate grains of salt. That said, AMD had veteran overclocker Sami Mäkinen fiddle around with a Ryzen 7 1700 CPU mounted on an Asus Crosshair VI motherboard along with some fancy RAM. Unfortunately, the exact models of memory modules aren't specified in the post, but the testing was performed with both single- and dual-rank modules. The dual-rank modules maxed out at a data rate of 3200 MT/s, while Mäkinen achieved 3520 MT/s with the company's chosen single-rank modules.
The AGESA 188.8.131.52 update allows motherboard vendors to offer toggles in their setup utilities for a couple of new options: Bank Group Swap (BGS) and Gear Down Mode (GDM). The first option has an effect on how physical addresses are assigned to applications, while the latter option may improve memory compatibility by reducing the effective data rate on the command and address buses. AMD found that Bank Group Swap actually improves the performance of synthetic applications when it is enabled, while disabling it offers a minimal-but-measurable improvement in gaming performance.
The effects of the Gear Down Mode option appear to be a little more complicated. While it is enabled, setting the "command rate" option for memory timings is pointless; it will always effectively be "2T" (although the company notes it's not quite the same thing—1T versus 2T command rates may still be worth exploring on their own). Disabling GDM will allow memory overclockers to manually set the DRAM command rate to "1T" and realize the full benefits of doing so. Tweakers have known for a long time that a 1T command rate can have huge effects on memory performance, and the Ryzen platform is no exception. However, if your memory can't run a 1T command rate at its spec'd clock rate, or if you're overclocking and can't hit a high clock with a 1T command rate setting, you could just be making your system less stable by disabling GDM.
AMD tested the effects of most of these settings, as well as a few memory myths. The questions of whether auto tunings are "good enough," whether builders should prefer single-rank or dual-rank memory, and whether tweakers should prefer a high frequency or tight memory timings timings (at the expense of data rate) are also addressed in the post. Naturally, the answer to every one of these questions is "it depends." In general, for the best gaming performance, AMD says you want a good balance between a high transfer rate and tight memory timings, and that's going to take some hand-tuning.
The best performance AMD's overclocker found was using single-rank DIMMs at 3466 MT/s with CAS latency 14, hand-tuned sub-timings, GDM and BGS off, and a 1T command rate. Overall, comparing the worst case (DDR4-3200 CL14 with auto settings) to the best case above, Mr. Mäkinen garnered a 10% improvement in Rise of the Tomb Raider framerates, a 16% improvement in Hitman performance, and a 9% improvement in 3DMark Sky Diver scores. That's perhaps a small gain for some hours' work, but it's a nice upgrade for exactly zero dollars. Check out AMD's blog post for the full details on the performance gains potentially available in AGESA 184.108.40.206.