Ryzen Beyond Recommendations
Organization: Why Do I Want Four Ranks?
A rank comprises the memory that connects to one 64-bit side of a DIMM. Most of these ICs are eight-bit, so that eight ICs in parallel compose a single 64-bit interface. Each module has two 64-bit interfaces, one in each side, which is why it’s called a Dual Inline Memory Module. A dual-rank DIMM has two interfaces filled with memory, while a single-rank DIMM has a filled side and an empty side.
Time and time again, we’ve seen proof that two ranks of memory per channel can significantly outperform one rank per channel. That’s probably due to interleaving, where for example one rank can start an operation while the other is finishing—think of it as latency overlap. It doesn’t matter whether those two ranks come from two single-rank DIMMs or one dual-rank DIMM, the performance improvement is the same.
A dual-channel motherboard with two ranks of memory per channel will have four ranks of memory. The problem for current buyers who want their 16GB to be spread across four ranks is that the 4Gb integrated circuits (ICs) that made up those 16GB quad-rank kits are no longer available. Most of today’s DRAM ICs have eight gigabits capacity, so that an eight-chip rank has 8GB capacity and four of those have 32GB capacity. Alternatives include buying very old memory with 4Gb ICs or very cheap memory with 8Gb ICs that have a 16-bit interface (so that four ICs make up a rank). Neither of those alternatives sounds very appealing.
We’ve already used a four-rank set of lower-density memory to show that our benchmarks don’t benefit from more than 16GB capacity when the number of ranks is the same, but how much more performance does four ranks really have than two?
The numbers appear to be all over the place when so many benchmarks are averaged in our overall performance chart for this change, but there appears to be a trend of four-rank configurations outperforming two. Things get interesting when we limit our focus to benchmarks that we know have memory bottlenecks:
Would you have believed that you could increase your frame rates by up to 10% in Ashes simply by using four ranks of memory? Or that you could reduce file compression time by up to 13%? Isn’t that performance gain similar to upgrading to the next level of graphics card or CPU? Given that 8GB per-rank capacity fills the enthusiast market, isn’t this a good enough reason for us to recommend 32GB to gamers, even to those who will never need more than half of its total capacity?
The DDR4-3600 Limit: Should I Try DDR4-3733 Anyway?
For the most part, the best configurations for DDR4 (and even DDR3) performance have been at multiples of 266.6x MHz, but DDR4-3600 doesn’t fit that scheme. Since we’re unwilling to take our DDR4-3600 all the way down to DDR4-3466, one might wonder why we don’t simply jump up to DDR4-3733? For Ryzen 3000, the answer lies within the assumed stability limitation of its I/O die. AMD switches its onboard Unified Memory Controller frequency (UCLK) to 1/2 of the DRAM data rate whenever DDR4-3600 is exceeded, and further protects controllers on the I/O die from instability by decoupling its Infinity Fabric clock when exceeding that frequency.
Now that we’re using an X570 test platform for memory reviews, we’ve seen plenty of proof that DDR4-3733 can lower performance compared to DDR4-3600 when those controllers are allowed to employ AMD’s default limitations. But hey, 1866 MHz (the clock rate for DDR4-3733) is only 66MHz higher than AMD’s approved 1800 MHz (the clock rate for DDR4-3600), and 66 MHz shouldn’t be that much of a reach, right?
Overclocking doesn’t get much simpler than this: We left motherboard firmware at the default DDR4-3600 timings and simply bumped its frequency up to DDR4-3733, followed by manually configuring the FCLK frequency to 1867 MHz and the UCLK frequency ratio at 1:1. MSI firmware’s CPU NB/SoC voltage was already set to 1.10V for our other tests: Would the system remain stable enough to complete our benchmarks without further voltage increases?
Since this is an overclocking section, we based our charts (100% rating) on the standard for which we were told Ryzen 3000 is optimized: 2x8GB DDR4-3200 C16. We can see that nearly every alternative performs better to some degree, but we really need to dive into the image gallery to see pronounced differences. Or, as usual, we can skip to the highlights.
There’s what we want to see. If you can run AMD’s controllers stably at 1:1 ratios, DDR4-3733 will give you up to 3% better performance than DDR4-3600. Four ranks of memory will give you up to 15% more performance than two, even when those four ranks have the disadvantage of 2T timings. The one place where it payed to slow our four DIMMS from DDR4-3600 to DDR4-3200 was in Ashes, where the lower data rate got the advantage of using a 1T command rate.
Ryzen Way Beyond The Basics
The easy conclusion is that every enhanced setting that you’d expect to boost the performance of AMD’s Ryzen 3000 series CPUs, does, that it’s not optimized for any specific data rate but scales proportionally with all stable data rates. So here’s a short list of our findings—that is to say, confirmations:
- Going from two ranks to four ranks of memory is by far the best way to improve performance in memory-bottlenecked applications. At current memory density, that usually means having at least 32GB (even if you don’t think you’ll use that much capacity).
- Switching from 2T to 1T command rate provides a similar performance boost to shaving two cycles from CAS Latency.
- The only time DDR4-3200 C14 and DDR4-3600 C16 have similar performance is when the slower kit has the better (1T) command rate.
- Overclocking to DDR4-3733 provides a nice performance boost, but only if you’re able to stably do it with synchronized UCLK and FCLK.
- The best practical configuration is four ranks of DDR4-3600 at the lowest stable latencies. Reminder: Four ranks can be achieved from either four single-rank or two dual-rank DIMMs.
Keeping in mind that not every overclocking adventure produces the expected results, a 32GB DDR4-3600 C16 kit such as the one we used today is a great starting point and an acceptable finish line.
Photo Credits: Tom's Hardware
MORE: Best Memory
MORE: All Memory Content